In our deal flow

Software and services kept meeting in the middle. We saw software companies move toward doing the work for customers, while services companies tried to productize work they had historically sold as labor. They arrived with different advantages: one side had the product, while the other already had customers and people doing the job. The distinction became more interesting when we looked at delivery. Some engagements appeared capable of producing reusable software and knowledge. Others still depended heavily on doing the work again for each customer.

Several companies wanted to stay inside the customer’s existing tools. Their products showed up in enterprise systems, text, WhatsApp, and other companies’ agents. The appeal was straightforward: customers could get something useful without moving the work into another application. That left us asking where the company’s lasting position would develop. The interface might belong to someone else, but the product could still become important through the information it held, the exceptions it handled, or the workflow it supported.

Agents started appearing in go-to-market plans. One company wanted to become the tool a large platform’s hiring agent calls. Another was building around agent payment protocols. Another listed a frontier model’s app ecosystem as a potential channel. These were plans at different stages, with limited evidence of repeatable acquisition in the companies we reviewed. Still, it stood out that founders were already making product decisions around how an agent might discover, select, and use them.

Elsewhere in the market

Trayd added accounting to its construction platform.

On September 24, Trayd launched accounting alongside the payroll and workforce tools its contractor customers already use. The product brings job costing, billing, accounts payable, and financial statements into the same system, with AI-enabled workflows for coding contracts, purchase orders, and invoices. The announcement names early contractor users and says integrations will remain available for customers keeping their existing accounting systems. It gives us a concrete expansion to follow from a product already handling an important part of the back office.

Trayd’s announcement →

BlackLine and Clio bought their way into adjacent work.

BlackLine acquired NetNow on September 21, adding customer onboarding and credit-risk work that happens before an invoice is created. On September 30, Clio acquired Learned Hand, which builds AI tools for judges and court staff, marking Clio’s first direct expansion into the judiciary. The deals cross different distances: BlackLine moves earlier in a financial workflow, while Clio enters another institution within the legal system, with its own users and buying process.

BlackLine’s announcement → · Clio’s announcement →

The model companies moved further into professional workflows.

Anthropic launched Claude for Financial Advisors on September 14, connecting custodial, portfolio, planning, CRM, and meeting information to tasks such as preparation and follow-up. OpenAI announced Astra for Law on September 17 for law firms and legal-technology companies, including application providers such as Harvey and Legora. Both announcements make the relationships between model providers and specialist software companies more interesting: the model company can supply intelligence, distribute a partner’s capabilities, and offer more of the workflow itself.

Anthropic’s announcement → · OpenAI’s announcement →

Cognizant brought agents into existing claims systems.

On September 28, Cognizant announced agentic workflow processing inside TriZetto’s health-plan administration systems. It is designed to handle eligible routine pended claims using a plan’s operating procedures, while routing denials, exceptions, and complex cases to people. The company announced general availability but also said early-adopter pilots were underway and performance results would follow. That makes it a useful case to revisit when there is evidence of how much work gets completed and how much oversight remains.

Cognizant’s announcement →

Insurers started deciding how to handle agents as customers.

Insurify blocked Meta’s Muse from its marketplace on September 23, citing missing coverage context and costs associated with automated quote requests. It continued supporting agent access through its own APIs and integrations. Five days later, Liberate launched AI Intercept to identify agents calling insurance businesses and route them to insurance-specific AI. Together, the announcements put some practical detail around “agents as a channel”: the receiving business has costs, responsibilities, and a say in how the interaction works.

Insurify’s announcement → · Liberate’s announcement →

Vanguard’s Altruist agreement carried into September’s conversations.

Announced August 26, just outside this month’s window, Vanguard’s agreement to acquire Altruist remained worth following. Vanguard first invested in the company in 2020 and expects Altruist to retain its leadership, brand, and operating model after closing. The announced rationale brings together advisor technology, custody, and established advisor relationships. That combination caught our attention: the software is part of a much broader position in how financial advice gets delivered.

Vanguard’s announcement →

If something here connects with what you’re building or seeing in your own market, we’d like to compare notes.

A note from Daring: We occasionally invite people in our community to write about the parts of building companies they understand unusually well.

For our first guest essay, Yelena Ambartsumian, founder of AMBART LAW and a Daring Ventures Community Partner, looks at a question applied AI founders are going to encounter more often: when software acts, who ultimately has to stand behind what it does?

Last year, a civil engineering client of ours received an unusual request. One of its customers had contracted with an AI vendor to produce a survey of the interior of a large manufacturing facility. The survey was finished, but it was not usable. That is because no licensed engineer had stamped it, and, without a stamp, no permitting office in Connecticut would accept it.

So the customer made an offer. Would our client “review” the survey and stamp it, for a modest fee? Easy money, right?

Well, the customer was greedy, and they wanted to use their own paper, instead of my client’s standard terms and conditions. So that’s how the question came to me.

Legally, our client could stamp an AI-generated survey. In fact, Connecticut had answered this question decades ago. Under § 20-300-10(b) of the Board of Examiners’ regulations, a licensee may not seal work she did not prepare or supervise unless she prepares, signs, seals, and retains for six years a “thorough written evaluation of the professional services represented by the documents,” down to the design calculations and the applicable codes.

In other words, the regulation permitted the customer’s request, but it first required the engineer to do the work the customer was hoping to skip.

More worrying than the modesty of the fee were the liability and indemnity terms in the agreement that accompanied it: put your license on the result, absorb the downside if the AI got it wrong, indemnify us for third-party claims, and do it for a fraction of what the survey itself would have cost.

I see a similar request on legal platforms nearly every week. A business has an AI-drafted complaint, though of course it never says so. Rather, the request reads: we are looking for a litigator to assist with a filing. You do not have to do anything! The issues have already been analyzed, and a thorough complaint has been prepared. We only need a lawyer admitted in the right court to sign and file it, for a flat fee of $1,000 or so.

But Rule 11(b) of the Federal Rules of Civil Procedure attaches to the signature, and it certifies that the signer made a reasonable inquiry into the facts and the law. The courts that have sanctioned lawyers for filing hallucinated citations since Mata v. Avianca were not asking who wrote the brief. They were asking who signed it.

Anything a lawyer submits to a court, she has to stand behind. (Oh, and we can be sanctioned too.)

Founders are told to ship. Professionals are required to sign.

The software can produce the survey or the complaint, but it cannot be licensed, disciplined, or sued. The liability, therefore, attaches to the nearest license.

And that gets us to the next example, the radiologist.

The radiologist

Consider the scenario Maddi and I discussed when we first spoke about this piece. An AI system reads a chest CT alongside a radiologist and misses a nodule. The radiologist signs the report. The patient is harmed.

Who is accountable?

Today, the answer is the radiologist, who is responsible for the final interpretation of the CT scan, and the law is reasonably settled on this. The standard of care belongs to the physician.

The FDA clears computer-aided detection software as an aid to the reading physician, and the vendor’s terms of service repeat the point: the tool “does not replace or substitute for clinical judgment,” a phrase that appears in nearly every AI vendor agreement I have reviewed.

The radiologist’s signature on the report is her stamp.

But the answer is less stable than it appears, and it is eroding in two directions at once.

First, the standard of care can invert. Once a detection tool is widely adopted, the exposure may shift toward the radiologist who did not use it.

Price, Gerke, and Cohen anticipated this in JAMA in 2019: a physician is safest when she follows the prevailing standard, and the prevailing standard may eventually include the tool.

Sometimes, as the technology becomes readily available, you are faulted for not having it.

In a 1932 Second Circuit case about negligence, The T.J. Hooper, the Court held that two tugboats sailing without radio receivers, which would have alerted them to the evening report that predicted worsening weather, were unseaworthy.

Interestingly, at the time, most operators were not purchasing radio receivers for their crews, but the Court explained:

“An adequate receiving set suitable for a coastwise tug can now be got at small cost and is reasonably reliable if kept up; obviously it is a source of great protection to their tows.”

Radios became a standard within a few years.

Second, the radiologist cannot inspect what she is being asked to vouch for.

In a 2024 paper in Future Healthcare Journal, Lawton and colleagues warned that clinicians risk becoming “liability sinks” for artificial intelligence, absorbing blame for errors in systems they neither designed nor can audit.

The engineer asked to stamp an AI survey could decline. The radiologist cannot; the hospital chose the tool, and it comes with the job.

From “in the loop” to “over the loop”

In the engineering, legal, and radiology scenarios, there is a human reviewing the output, or, as we say in AI governance terms, “human in the loop.”

Without that, the accountability question would not be answerable at all.

Agentic AI, however, removes that step.

In a piece for Women in AI earlier this year, I argued that governance for autonomous agents must move from “human in the loop,” a person approving each decision, to “human over the loop,” where people set boundaries, define escalation triggers, and monitor behavioral telemetry rather than reviewing every action.

In a human-in-the-loop system, accountability attaches at the point of decision: the moment the professional signs.

In a human-over-the-loop system, it attaches at the point of design: the moment someone chose the boundaries, set the thresholds, and decided what the agent could do without asking.

In practice, that is usually the deployer, as an institution. The hospital’s governance committee, not the individual radiologist.

And it is the vendor to the extent the vendor set those parameters and did not allow the customer to change them.

I see this in the contracts we negotiate for our AI-native clients. Their enterprise customers increasingly ask the vendor to stand behind the consequences of the agent’s design and default configuration: how it was tuned, what it escalates, what it does when it is unsure, rather than behind the software’s conformance to its documentation.

That is a different warranty, and it tracks the shift we are discussing here: the customer has worked out that the decision was made upstream, at the design stage, and wants the party who made it to carry the accountability.

Where the law is today

For the founders reading this, here is where each of the three parties stands.

The professional still carries the standard of care, and will for some time. Courts move slowly, and malpractice doctrine is built around a human decision-maker.

The deployer carries negligent selection, configuration, and supervision. This is where “over the loop” matters legally.

A deployer that runs an autonomous agent without logging its decisions is choosing to have no evidence when something goes wrong.

Under Article 26 of the EU AI Act, deployers of high-risk systems must assign human oversight to people with the competence and authority to exercise it, and must keep the logs those systems generate.

That is a legal description of human over the loop.

The vendor is the interesting case, because the answer depends on the jurisdiction.

In the United States, vendor liability for AI software is almost entirely a matter of contract. The 40-page vendor-risk package your first enterprise customer sends, the security questionnaire, the DPA, the “AI addendum,” is the deployer trying to push accountability back upstream, to you.

In the European Union, the revised Product Liability Directive (Directive (EU) 2024/2853) now defines software as a product. Beginning in December 2026, a vendor whose AI causes personal injury or property damage to a consumer will face strict liability that cannot be disclaimed by contract.

Founders selling into both markets are therefore negotiating two different accountability regimes with the same product.

The founder’s position

Most of Daring’s founders are the vendor in this story.

Limitation-of-liability caps, the allocation of responsibility for outputs, the definition of “misuse,” and the question of who controls the agent’s boundaries after deployment: these terms decide who answers when the agent gets it wrong.

Negotiate them accordingly.

And human over the loop means accountability earlier, not less of it.

The boundaries you code, the escalation triggers you define, and the logs you keep are the evidence a court or a regulator will ask for.

If your governance lives in a policy document rather than in the architecture, you have kept the liability and discarded the defense.

The engineer who declined to stamp the AI survey and the lawyer who declines to file the AI complaint are making the same judgment about the risk to their licenses.

Agentic AI has not eliminated that question. But it has moved it earlier into the process, to the people who decide what the system is allowed to do before it does anything at all.

For founders building applied AI, that is you.

About the author:
Yelena Ambartsumian is the founder of AMBART LAW and a Daring Ventures Community Partner. She advises companies on commercial and technology matters, including AI-related contracting and governance.

The views expressed are the author’s own. This article is for informational purposes only and does not constitute legal advice.

Not because they had every answer on day one. Because once they understood the vision, things started moving fast. Work that had been sitting with the founders started getting picked up and pushed forward without needing to be constantly directed.

That sparked a bigger conversation about hiring for agency.

One of the hardest things to figure out in an interview is whether someone will actually take ownership once they’re in the seat. You can ask about it. You can talk about past experiences. But sometimes, you don’t really know until they’re onboarded and you see how they operate.

The takeaway: As a founder, your job isn’t to keep doing every job better. It’s to find people who can take jobs off your plate and run with them.

And that’s exactly what we dug into in this week’s Small Wins, along with what happens when AI starts taking ownership of the work too.

Watch the full conversation →

tl;dr

• handshake: the two-party agreement on what makes work usable

• AI made “looks right” free, so usable-by-the-next-person is the product

• entry point like the ritual: ritual gets a product used, handshake gets its work accepted

ALT H FD S O X G E ↵ ALT H F C M TAB ALT B 255 ↵

So what is that gibberish?

It is the Konami code for people with no personal life.

More specifically, it’s the keyboard shortcut sequence to select all the “hardcodes” or numbers that have been manually typed into cells in Microsoft Excel and color them blue.

It’s just a color, but it means that someone you have never met can open your model and understand how to approach it. You’ll disagree about almost everything else, but not about how to read it.

Excel is basically a sandbox that lets you do whatever you want. So in terms of structure, it contributes essentially nothing to this.

Let’s break down the shortcut:
ALT H FD S opens Go To Special
O X G E selects the constants and unchecks everything except numbers
ALT H F C M TAB ALT-B 255 turns the font a specific blue

Excel knows which cells were typed and which were calculated but doesn’t need to know why anyone cares, and it will happily let you color data invisibly white if you want.

The users supply the opinions, the software is the blank slate.

Now that software is doing the work, no one is standing between the software and the next person, holding that understanding on its behalf. Either the product knows what that person needs in order to use the work, or the work comes back.

A lot of people call tools with this knowledge baked in highly opinionated, but I think that obscures the true source of it’s value as a wedge in software.

Instead, I’ve been calling it a handshake.

Real quick, some awesome news—Forward This is presented by Ambart Law, our newest community partner. Learn more here:

Two entry points.

Few weeks ago, I wrote about the concept of a ritual as an entry point for B2B software startups. A ritual is a moment the institution already performs, such as a debrief, an inspection, or a shift handoff. It works because products that enter one the right way don’t ask users to change behavior.

The handshake is another one of these entry points because it ships meeting an existing standard: the recipient of the work it produces can work from the output on day one.

Work that looks right and work you can use.

In real life, a handshake is an agreement in context. Doesnt really matter if it’s a fist bump or firm corporate handshake. Ultimately, it boils down to:

In professional workflows, some handshakes are explicit while others aren’t. Contracts specify what to do and professional standard defines the bar it must clear. We learn them through certifications, training, or from our boss after messing something up.

Everyone has the same model.

A foundation model is a lot like Excel. It can do nearly anything and does not know what any of it means. The model supplies the capability, and the founder supplies the understanding of what the work has to be when it reaches the other side.

That is where a generic capability becomes a specific product.

Bloom, one of our portfolio companies, works with visual researchers, who gather and arrange references so other creatives can work from them. Anyone can point AI at collecting images.

But “here are some images” and “here is the direction we are working toward” are different deliverables, however similar they look from outside the profession. The researcher is judging throughout: the light in one image, the composition in another, the pacing of a sequence.

The next creative can only work from the result if those judgments arrive with it. Bloom leaves the judging to the researcher and concerns itself with what has to arrive.

Hostie, another portfolio company, works with restaurants. Consider what the phone is for. A party of eight calls. One person cannot manage stairs, and everyone needs to leave in time for a show.

A usable answer is a commitment: the restaurant knows what it has agreed to provide, and the guest knows what has been arranged.

Getting there takes seating, timing, policies, sometimes a follow-up question, and an understanding that sometimes a person from the restaurant stepping in is a feature, not a failure.

The two companies share almost nothing on the surface but, in both, the product is defined by what the person on the other side needs in order to rely on the work.

What follows.

The handshake as an entry point for software startups suggests an attractive growth pattern that starts from a very simple perspective:

The other side does not have to change.

This isn’t a new insight, but the economic viability of this type of solution exists in tiny wedges now.

Work that crosses a handshake is accepted or it comes back, and the person deciding held that job long before the product existed. A founder does not have to teach the market what good output looks like. The counterparty will tell them, usually the same day.

Accepted work earns a position. The counterparty adopts nothing, but they see the work over and over, in a form they already trust. That is how a narrow product moves toward the rest of the exchange and the transaction behind it, and it explains expansion better than three concentric circles in a deck.

And it can be tested in one conversation, before there is any traction to examine. Ask the founder what gets work sent back. People who have done the job answer immediately, in detail, and with some lingering resentment.

Recommended for you:

Knowing the handshake is not enough.

The convention and the handshake are different things. Some conventions are only adaptations to bad tools, and automating them faithfully preserves the workaround.

Nobody owns a handshake. Knowing the handshake is an edge in building the product and not a moat around it. Durability has to come from what accepted work earns.

Some exchanges check who did the work as well as what it says. A licensed engineer’s stamp, an auditor’s attestation, and a physician’s signature are all handshakes, and none of them can be satisfied by output that merely reads correctly.

The diligence questions are specific:
What gets work sent back?
What does the other side need in order to rely on it?
Does the counterparty have to change anything?
Is there real economic activity past the counterparty?
Does accepted work put the product in front of it?

A convincing answer should survive an awkward exchange as well as a clean demo.

The small things tell you where to look.

Yes, the spreadsheet shortcut looks absurd written out.

Day -to-day, no one actually experiences it as a decision. I actually didn’t really remember what the shortcut was, but it was just muscle memory because making the model ready for the next person is such a familiar moment.

That familiarity makes these details easy to overlook. People inside an industry barely think to explain them. People outside it see formatting, preferences, bureaucracy, a weird flex, or ceremony.

A founder who knows the work knows what the next person needs in order to use it. In software, that knowledge is most of the product.

Because building is hard, and most wins look kind of ridiculous next to the thing you’re actually trying to pull off.

You’ve already got LinkedIn, the company Slack, and your investors for the big stuff.

This is the place for “cracked three likes on my Substack note.”

We kicked off Small Wins last week with Taylor Rose, the founder and CEO of Kinstead Health.

Taylor’s small win: finally hiring an engineer after months of searching, three rewrites of the job description, and eventually realizing they were looking for a pretty different person than they thought.

The person they found had 15 years of startup experience, deep ties to healthcare, and two nurse practitioners in the family.

From there we got into hiring, healthcare, where AI actually helps, and all the company-building decisions that refuse to fit neatly into a process.

Now tell us one thing that went right this week for you.

Get more from Joseph Alalou in the Substack app

Available for iOS and Android

Get the app

Coffee shops, event spaces, open-plan offices with tastefully understated Knoll furniture that belies how much a large pension fund this startup paid for it.

Not ours.

Let me preface this by saying we are very appreciative of Reitler for letting us use their space. (Thanks, Theo!)

With that out of the way, it’s time for me to say this out loud.

Founders & Fractionals is held in an objectively ridiculous room.

It is a straight-up boardroom. Thirty-plus chairs around one long, narrow table. Great for hostile takeovers and the like. Not so great for networking. Definitely doesn’t scream early-stage startup event.

People walk in and there is usually a beat of confusion.

Am I in the right room?

Am I walking into someone’s deposition?

Am I walking into my own??

Eventually people find their seats, chat with whoever ended up next to them, and the panel gets going. Pretty normal stuff from there.

The magic happens when we open up for Q&A. People just react differently at that moment. For some people, there’s still tension. They’re looking over at Maddi or I before they jump in. It’s a bit of: are we doing hand-raising , are we taking turns, like what’s the format here?

But for others you can see a spark, a twinkle, a something that says showtime!

Yeah, so we’ve tried this several times and here’s where it broke…

Real disagreement. Unapologetic viewpoints. People get sharper and speak up. They stand up to hand-wavy suggestions, question fluffy rationale, but none of it is combative or mean. It’s just a level of comfort with saying stuff.

Recommended for you:

I think the phenomenon is amplified by how much start-up culture can be almost comically different. There is a lot of improv and a lot of “yes, and.” That is probably necessary when everybody is trying to make something out of nothing, but it can also create this weird social pressure where every contribution has to be additive and every idea gets treated like it might secretly be amazing.

Sometimes you end up with a room full of people politely building on an idea nobody actually believes.

Meanwhile, in corporate America, people learn pretty quickly that “yes, and” turns into a meeting.

Then another.

And then another but to bring additional stakeholders up to speed.

Which means dashboards.

And decks. And trackers. And trackers tracking trackers.

You get the point. Yes, and is a rookie mistake that no one makes twice. That’s why we learn corporate’s incredible defensive vocabulary.

Take it offline.

I’ll ping you separately.

I’ll circle back.

I’m convinced let’s put a pin in it has resulted in more corporate productivity than Microsoft Teams ever will. It’s an artform. It’s also incredibly irritating and I think it’s probably half of why many of us left corporate to do our own thing.

A pathological aversion to new is so incompatible with startups and the people who operate in and around them. But it is extremely funny to watch those instincts reappear the second you put people back around a boardroom table.

You would think everyone would get up and move somewhere more conducive to meeting people.

They don’t. They stay for another hour.

Networking.

Around a table where the person you want to meet might be four chairs away and therefore approximately 70 feet from you.

Somehow it works.

tl;dr

• It’s harder than ever to use cheap proxies to identify value.

• Both sides now arrive over-produced: founders with demos and audiences, investors with decks of logos and intros.

• Density is the question underneath: how much future company is packed into what exists today?

Pre-seed investing is a strange business because the evidence is supposed to be incomplete.

By the time a company is obviously important, the interesting part of the job is over. Everyone can see it. The price reflects it. The judgment has already been made by someone else, earlier, with less to go on.

The job is really getting there first.

Getting there first is more than hearing about the company first or taking the meeting first. We want to be first to recognize how much company is already hiding inside something that still looks small.

We’ve been calling it density which is our shorthand for a single question:

How much future company is packed into what exists today?

Venture-shaped objects may be smaller than they appear.

AI has compressed the cost of appearing further along than you are.

A founder can now produce, very quickly: a polished product, a credible website, a working demo, early revenue, customer logos, an outbound motion, content, a plausible-looking funnel. This is all real. It’s also much less informative about the company’s future than it used to be.

Over the last few decades, building the product itself filtered for some sort of fundable capability. The artifact was the evidence that some combination of team, technology, and temperament for execution yielded a working thing.

Today, the distance between “this looks like a company” and “this can become a consequential company” is much larger than it has ever been.

The same thing has happened to audience.

Ten years ago, having thousands of people paying attention to you was unusual. It suggested distribution, credibility, or some kind of earned pull.

Now everyone has an audience.

Founders arrive with newsletters, podcasts, LinkedIn followings, communities, Discords, waitlists, and thousands of people willing to like, repost, comment, subscribe, or show up. Audience building is hard. But again, it just doesn’t mean the same thing it used to.

Audiences are increasingly becoming less reliable customer bases as dozens if not hundreds of voices vie for their dollars every day.

Audience can make a young company look unusually large from the outside. Announcements travel farther. Launches look bigger. Waitlists fill faster. Every milestone arrives with applause.

It couldn’t be less of a factor in B2B and enterprise software.

A parade of people cheering for you is not the same thing as someone changing how they work, putting the product through procurement, giving you budget, and coming back the next month.

But density is not how many people are watching or how slick the demo is.

It is what the people who actually need the product are doing.

So the underwriting problem has changed.

The investor has to look beneath visible progress and ask what produced it.

Recommended for you:

Weight, but how?

If density is the amount of future company compressed into the company that exists today, how do we see it?

It starts with our view that a dense company may still be tiny. It may have very little revenue. Its product may cover one painfully narrow workflow. Its team may be two people (often one plus the target CXO who has been helping and will join post raise.)

What matters is what sits underneath those facts.

Density shows up as some combination of:

Product conviction. Unusually specific beliefs about what should exist and why, the kind that usually come from having lived inside the problem.

Commercial pull. Evidence that the problem matters enough for customers to behave differently than they otherwise would.

Relationships to execute. Access, credibility, distribution, domain trust, or a network that would take an outsider years to reconstruct.

A narrow wedge with enormous consequence. Something small enough to win now, but structurally connected to a much larger company.

Density is not just another way to say “great founder.” Exceptional founders create it, but it is bigger than them.

It is the set of conditions they have already produced around the company: the relationship between the founders, the market, the product, the customers, and the opportunity.

A brilliant founder pointed at the wrong wedge is not dense. Neither is the right wedge in the hands of someone who doesn’t know what they don’t know about getting the first meeting.

The founder matters because these conditions are not accidental and, once created, it becomes a gravity well of talent and capability.

Density creates gravity that scales.

Small differences at the beginning do not stay small.

Two companies can look almost identical at pre-seed. We’ve seen several companies that are practically indistinguishable at first glance.

But then you notice the little details and they accumulate. One founder understands the customer slightly better. One wedge is slightly more central to the workflow. One team has slightly stronger relationships. One product gets pulled into an adjacent use case a little sooner.

Each advantage makes the next advantage cheaper to acquire. Which is why motion is not growth, and speed is not velocity. The question is no longer how quickly the company is moving.

It is whether each unit of progress increases the company’s ability to make the next unit of progress.

Looking like a startup and moving like a startup have never been easier. Growing like a startup is still really hard.

What it looks like in practice

Density is easiest to see in wedges.

We like narrow initial products, which sounds like a preference for small ambition.

It is the opposite.

A narrow wedge looks unimpressive if you evaluate it by surface area. But narrowness creates density when the wedge sits inside an unavoidable workflow, an institutional ritual, an expensive failure point, a regulated process, a recurring approval, a system of record, or a deeply embedded operational moment.

The important question is not:

“How big is this product today?”

It is:

What happens if this company earns the right to own this moment?

A tiny wedge in the right place can contain an enormous company. A broad product in the wrong place usually contains a feature.

The test of an investment thesis is not whether it can explain great companies after everyone agrees they are great.

It is whether it explains the companies you backed before the market did.

At pre-seed, the company is supposed to be malleable. The material has not been fired yet. Product, positioning, even the shape of the company may still change, but the thesis can’t. It has to be hardened before the first check. Otherwise it can always be reshaped around whatever eventually works, which is less conviction than narration. A finished company can still be a great investment. But buying the ceramic does not tell you whether your kiln works. It says you waited in line at the store.

Density is simply a name to something we were already underwriting.

PraxisPro

Investment date: March 2025.

What existed. AI coaching for pharma sales reps. A narrow product inside a heavily regulated industry.

What the market saw. Too niche. Sales tech. A point solution.

What we saw. Coaching was the collection mechanism for a compliant intelligence asset. Every interaction produced signal about reps, messages, products, and territories inside an environment where almost nobody else can legally sit.

The thing that made it look small was the thing that made it investable.

What happened next. A $6M oversubscribed seed led by AlleyCorp. The company today is live market intelligence for life sciences.

Hostie

Investment date: March 2025.

What existed. An AI phone agent for restaurants. Eleven customers and $50K ARR.

What the market saw. Another voice agent in a crowded category. A feature a reservation platform could bolt on.

What we saw. The phone was the last uninstrumented surface in the restaurant, sitting upstream of the reservation, the order, and the guest record. The technical problem was increasingly cheap. Distribution was not. Restaurant software spreads through relationships, and Randall had spent his life inside the industry.

What happened next. Obvious Ventures lead Series A

The trajectory: from answering the phone to owning guest demand, reservations, and orders.

Bloom

Investment date: May 2026

What existed. Rights-aware visual research. 208 monthly active users.

What the market saw. Another AI memory tool in a category with a visible ceiling. Rights looked like a compliance feature.

What we saw. The rights layer was structural. Professional references eventually have to be cleared, forcing the product into licensed sources and professional workflows that general-purpose tools have little reason to serve. And the users did not merely consume references. They shipped them to other creative professionals, making the wedge its own distribution.

The memo put the mispricing plainly: we were buying the latter at the price of the former.

The trajectory: from visual research to the intelligence and rights layer behind commercial creative.

Stealth-mode startup we cannot wait to share

Investment date: May 2026

What existed. A workflow product for biomedical IP commercialization, with 13,000 IP assets already ingested.

What the market saw. Software for technology transfer offices.

What we saw. The workflow sat directly in the path of transactions between research institutions and commercial buyers. Every deal could produce proprietary data about which IP matches which buyer, which structures close, and which assets have commercial potential. The workflow was the entry point. The accumulating data was the asset.

The memo called it the data moat at the price of a workflow tool.

Two of these companies are already further down the curve.

Two are still early.

The underwriting pattern is the same.

The elephant in the room

Yes, this cuts both ways. There is absolutely an implication here for venture firms too. At pre-seed, the alpha is increasingly less about who you know and more about what you know.

Networks still matter. Introductions matter. Access matters.

But access has become abundant too.

Every investor has a customer network, a recruiting network, downstream funds, operators, executives, founders, and a list of people they can introduce you to.

The existence of the network tells you less than it used to.

What is harder to manufacture is judgment.

That matters most at pre-seed because the best opportunities are often least legible at exactly the moment when the investment is most interesting.

What really matters

How the buyer actually behaves. Why a particular workflow matters. Which customer behavior contains information and which is noise. Why a seemingly narrow wedge can expand into something much larger. What the founders understand about the market that an outsider would need years to learn.

And founders are doing exactly the same evaluation in reverse.

The founders with the deepest knowledge of their markets are often the quickest to recognize which investors genuinely understand what they are building.

They can tell the difference between someone who can interrogate the business from first principles and someone who has learned the vocabulary.

They can tell who sees the wedge, who understands the buyer, who knows where the commercial friction actually sits, and who is mostly offering access to other people.

They are underwriting us.

Which means density cannot be something venture firms demand from founders without applying the same standard to ourselves.

Our firms need density too.

tl;dr

• AI made reasoning, orchestration, and translation abundant. Capability by itself is no longer evidence of a durable company.

• The key diligence question is where “correct” lives: in the model, in the customer’s existing systems, or in a record the company is building itself.

• The interesting companies institutional memory from durable records that a competitor cannot recreate from the same models and connectors.

I’ve been thinking a lot about grammar recently. Not because of the watermarking or AI checker hoopla, by the way.

It’s actually because I’m realizing grammar is a really useful lens to apply to applied AI startup pitches.

A lot of the language sounds concrete until you ask whether the noun names an actual thing or just work the software performs.

Take this very common premise:

“We are the reasoning layer across the enterprise.”

It is becoming a common description of AI companies, and an increasingly uninformative one. You can pull up an invoice, a claim, a contract, a payment, or a ticket. You can say who owns it, where it lives, which system is authoritative, and what breaks if it disappears. Reasoning has none of those properties. It describes work being performed rather than anything the work produces.

That distinction used to matter less. If a company could reason across several messy systems, normalize their data, resolve inconsistencies, and take action, there was usually substantial infrastructure underneath the claim: integrations, custom logic, implementation work, and people handling edge cases. The capability itself told you something about the company.

So what changed?

Real quick, some awesome news—Forward This is presented by Ambart Law, our newest community partner. Learn more here:

You guessed it. Those two letters. AI.

AI changed the relationship between the capability and what had to be built to deliver it. A company can now perform sophisticated work over systems it barely controls. The product can be useful, the demo can be excellent, and the customer can be happy without the company acquiring much ownership of the workflow underneath it.

So the question I care about has shifted. After the reasoning is done, what exists that did not exist before?

What sits underneath the reasoning

Take a product that handles deduction disputes. It pulls a dispute from a portal, checks it against the order, decides whether the exception is valid, routes the case, and posts an adjustment.

At the top is orchestration: moving work through systems.

Underneath is reasoning: deciding which record to trust, which exception matters, and what should happen next.

Both can be valuable. Both are also increasingly available to another company with comparable models and connectors.

The underlying context is more concrete. The order history, contract terms, payment behavior, prior disputes, and account record all exist somewhere. Usually they belong to somebody else. The ERP owns the invoice, the CRM owns the account, and the retailer portal owns the deduction. The AI company gets permission to read them.

That means a surprising amount of the product can be assembled from things the company neither created nor controls. The orchestration is built through integrations. The reasoning comes from a model. The context comes from read access.

The interesting question is whether anything accumulates beyond those inputs: a durable account of what correct looks like inside this particular institution.

That is different from a policy document. Policies describe what is supposed to happen. The useful record is built from what actually happened: which decisions worked, where practice diverged from the written rule, which divergences were mistakes, which became accepted practice, who made the call, and what happened afterward.

That knowledge cannot simply be fetched from the ERP or inherited from a frontier model. It has to accumulate case by case.

Where the standard comes from

When a product makes a judgment, there are a few places the standard behind that judgment can come from.

Sometimes it comes almost entirely from the model. The company combines frontier-model priors with a prompt and enough context to perform the task. The work may be good, but another team can access the same underlying capability. If the reasoning disappears, it can usually be regenerated. What matters most in that business is execution: distribution, implementation, product quality, and speed.

Sometimes the standard already exists in the customer’s records. The company reasons over policy manuals, contract terms, ERP objects, payer rules, and other material the institution already maintains.

An invoice-to-cash product can have excellent adjudication logic while every important object remains authoritative elsewhere. The invoice lives in the ERP, the remittance comes from the portal, and the ledger entry lands in the general ledger. Switch the product off and no record disappears. The work gets slower and worse, which can still support a good business, but the customer has not lost the underlying record of what happened.

This position also creates dependence on whoever owns those records. With physical assets, that dependence is obvious. Nobody assumes a company can become the software Teslas run without Tesla’s cooperation because the asset is visible and clearly owned. Software obscures the same relationship. A company can look deeply embedded in a workflow while its position still depends on access to objects controlled by somebody else. A dispute record has an owner just as surely as a car does.

Recommended for you:

The more interesting case is when the product creates the standard it uses.

Consider a company that captures debriefs after consequential events, synthesizes the participants’ accounts into a structured record, and compares that record with written policy. Each event creates a new piece of institutional memory. Over time, the records show where doctrine and practice diverge, which deviations were mistakes, which were improvements, and how the organization actually responds under real conditions.

Two hundred of those records are more than two hundred transcripts. They describe something the institution previously did not have in durable form. The knowledge lived in people’s memories, spread unevenly across teams, and disappeared with turnover.

If the product becomes the place where that history accumulates, the company owns something meaningfully different from the model reasoning over it. A competitor can connect to the same source systems and use the same models without recreating the record, because the events that produced it have already passed.

That is the case I want to find.

What survives if you unplug it

The easiest screen is still what happens when the company disappears.

If nothing specific disappears, the company is primarily selling a capability.

If the important records all survive in other systems, the company has built a workflow around someone else’s property. That can be valuable, sticky, and difficult to replace, but the source of authority still sits elsewhere.

If an accumulated record disappears, the situation is different. The company may be creating a new system the institution starts consulting because it contains knowledge that exists nowhere else.

That does not automatically make it a good company. A dispatch schedule is a perfectly concrete object with discrete instances and a clear custodian, and the business built around it can still be terrible. The record may be tied to an asset somebody else owns. An incumbent may control distribution. Customers may simply not care enough about it to switch.

A newly created record can also remain peripheral forever. It can accumulate without becoming authoritative. It can be useful without becoming critical. It can sit inside a workflow the customer can recreate elsewhere.

So this is only a screen. Once the object appears, ordinary underwriting begins. Who uses it? Who refers back to it? Does it improve as cases accumulate? Does the institution start trusting it? What becomes difficult to reconstruct if the company disappears?

What the work leaves behind

This distinction matters beyond reasoning products.

Middleware used to derive much of its value from translation. Moving information from one system into the format another system expected required parsers, mappings, connectors, and accumulated integration knowledge. Models are making more of that translation cheap.

The middleware businesses that remain durable tend to own something beyond translation: state, guarantees, audit history, responsibility, liability. Visa does not become irrelevant because a model learns to format a payment message. The important part of Visa is the network and the transaction state, rules, and obligations around it.

The same applies to service-as-software. Performing a service repeatedly becomes more interesting when each performance leaves behind something durable: an adjudication history, an exception record, a prepared file that captures how unusual cases were handled. After two hundred cases, the company may hold something a new entrant cannot reproduce merely by connecting to the same systems.

AI is making capabilities abundant. That does not make every company offering those capabilities identical. One may be selling execution over systems other people own. Another may become deeply embedded without ever controlling the authoritative record. A third may quietly create a record the institution starts depending on.

The pitch tells me what the company can do.

I want to know what is left after it does it.

tl;dr

• Transmission separates rituals from habits. A habit dies with the person; a ritual belongs to a role and outlives whoever holds it.

• Rituals reduce behavior change and increase adoption rates.

• Venture-scale density = a small entry point already holding the ingredients of a much larger company.

One of the first questions I ask a founder about a workflow tool is “will anyone actually stop and use/do/reference this?”

Documentation products die on this question all the time. Asking busy people to pause and narrate their work, for the benefit of some future reader, is asking for a habit that almost nobody sustains.

A few weeks ago I spent an hour with a founder building voice software around after-action reporting. The premise, stripped to its simplest form: after consequential work, the people involved should record what happened and what they learned, and software should turn that account into something the institution can actually use.

The product, in other words, was entering a moment of reflection the institution already performs, rather than attempting to create reflection from nothing. The behavioral question shrinks from “will people adopt a new habit” to “will people let a tool into a habit they already have.”

Intuitively, we’ve always loved these products that create data exhaust, but recently we’ve started identifying the very specific, narrow wedge they target as a ritual.

Recommended for you:

Job, ritual, workflow

Founders and investors already have good vocabulary for two adjacent ideas, and the ritual sits between them.

The job to be done answers why someone needs to act. In the after-action case, the job is to learn what happened and prevent recurrence. The unit of analysis is the outcome someone wants.

The workflow answers how the work moves from beginning to end. Here: capture the accounts, synthesize them, compare them against expectations, assign what follows. The unit of analysis is the sequence of tasks, handoffs, tools, and outputs.

The ritual answers a different question: when, with whom, and under what obligation does the action occur? In this case, the relevant team gathers immediately after the event to debrief. The unit of analysis is the recurring, role-bound institutional moment.

A ritual is more than something that happens repeatedly.

• It has a recognizable trigger: the consequential event just ended.

• It has understood participants: the people who were there, sometimes joined by whoever is responsible for what happens next.

• It has roles: someone leads, others contribute what they saw, someone carries authority over the follow-up.

• It has shared language: participants belong to the same institution and describe the work in its terms.

• It has obligations and consequences: it exists precisely when the work carries stakes.

• It has an expected output: it could be as simple as a shared understanding and a few commitments.

The ritual sits alongside the job to be done and the workflow and the three layers are complementary rather than competing. Specifically because each one answers a question the others cannot.

The job tells you the progress people want, but a job floats free of any particular time or place; people want to learn from failure continuously and abstractly.

The workflow tells you what the work looks like once it begins, but a workflow says nothing about why it ever begins.

The ritual supplies what both are missing: the where, the when, the who, and the obligation. It locates the abstract job in a concrete institutional moment and gives the workflow its starting gun.

Layers make a stack

The relationship among them is the useful part: the wedge enters through the ritual, improves the workflow, and helps complete the job.

Consider what a software product normally has to establish on its own before anyone uses it well:

• When the product should be used

• Who should participate

• Why they should care

• What information should enter it

• What artifact or decision should emerge

• What happens next

Most products manufacture these conditions from scratch. They send reminders to create a trigger. They chase seats to assemble participants. They run change management to generate urgency. They beg for data entry. They evangelize an output nobody asked for. Much of what we call go-to-market is really the expensive construction of context.

A ritual supplies that context before the product arrives. The trigger exists: the event just happened. The participants are assembled: the moment already draws them together. The urgency is real: the work was consequential, and the institution expects an account of it. The information is present, still fresh, in the heads of people standing in the same room. The expected output is understood, because the ritual has always produced one, however informally. And the moment already points at what happens next, because the debrief exists to shape it.

The wedge therefore has a reliable insertion point. It appears when attention is already concentrated and action is already expected. That is a rare and valuable position for a narrow product, and it is worth distinguishing from the usual situation of a tool waiting hopefully in a browser tab. A product outside a ritual competes for attention against everything else in the workday. A product inside one arrives at the single moment when the institution has already cleared space for exactly the work the product does.

None of this amounts to free distribution or guaranteed adoption. The ritual creates favorable conditions; the product still has to be useful, trusted, and well fitted to the moment it enters. What the ritual removes is the burden of inventing the moment itself. But the conditions are borrowed rather than owned, which raises a question about durability I return to below.

Behavior change, narrowed

A ritual narrows the behavior change without eliminating it. Some new behavior remains, and the product’s fit with the moment decides whether what remains is acceptable.

The after-action case makes the point through the question of capture, where there are three options.

The first is to ask workers to record insights as they occur. This requires a new habit with no context, no trigger, and no social support, and it fails for the reason most documentation initiatives fail.

The second is to record everything continuously. Ambient capture removes friction, but it also removes the boundary that tells people when they are speaking for the record. Continuous recording reads as surveillance, and the trust cost lands on exactly the people whose candor the product depends on.

The third is deliberate capture inside the understood debrief. Someone pushes to talk and everyone present knows why. The recording has a purpose, an audience, and a boundary matching one the ritual already maintained, so the behavior change shrinks to a gesture that makes sense to the people performing it.

That answer belongs to the ritual rather than to the problem, and elsewhere it inverts. Outside sales teams in pharmaceuticals hold what amount to windshield calls: reps phone one another between appointments to talk through what the customer pushed back on and how somebody else has answered it.

The moment takes that form for a regulatory reason, since the rules governing what a representative may say constrain the official channel tightly enough that the knowledge gets squeezed out of enablement and into the car. In the debrief, producing the record is the point.

In the car, the absence of a record is what permits the candor, and a tool that captured the conversation would end it. The design problem becomes capturing the output without capturing the exchange that produced it.

This is what it means to fit snugly into a wedge. The ritual has a timing, a cast, a tone, a set of norms, and a product that violates them does not inherit the moment’s advantages but contaminates them. Fitting the ritual is the price of the adoption conditions above.

Where institutional judgment becomes visible

The moment is valuable for a second reason: it concentrates information that exists nowhere else in the institution.

Walk one compact loop. Several participants contribute perspectives on the same event, and those perspectives differ because each person saw the work from a different position. Their accounts become one structured record of what happened, in what order, and decided by whom. Here, rituals can do more than push institutional knowledge downward. They expose where it should change, because the debrief is one of the few moments when practice and doctrine are held against each other by people with the standing to say which is right.

Notice what accumulates in a single recurring moment: information hours old rather than quarters old, tacit knowledge nobody has written down, conflicting accounts whose conflict is itself the signal, the rules invoked as the standard of comparison, the exceptions surfaced and examined, and the authority to decide what follows.

A product living inside that moment captures something the institution has never held in durable form. The transcript is raw material; the asset is institutional memory, a connected record of what the organization actually does, how that differs from what it says it does, and how the gap gets resolved event after event.

No system built from forms and fields contains it, because the information surfaces only in the ritual and otherwise decays with turnover.

Expansion follows the consequences

The expansion path is unusually legible, because the ritual sits inside a web of dependencies the product can follow: improve the ritual, own its artifact, connect occurrences across time, then influence the systems before and after it.

Improving the ritual earns the place. Owning the artifact follows, because once the debrief produces its structured record through the product, the product is the system of record for that moment. Connecting occurrences is where compounding begins: one debrief is an anecdote, two hundred are a dataset about how the institution operates, where it deviates, and what those deviations cost or save. It is also where the product stops depending on the moment that admitted it, since the institution comes to depend on the artifact more than the product depends on the occasion. The ritual is the entry, not the permanent residence.

From there, follow what feeds the moment and what the moment feeds. Policy, training, staffing, plans, and prior incidents sit upstream, and the debrief exists to test the work against them. Assignments, retraining, policy changes, escalations, and performance records sit downstream, and the debrief exists to trigger them. Whoever owns the connected record holds the evidence those systems run on, and the larger company emerges one adjacency at a time, each justified by what the ritual produced.

This is a different logic from horizontality. A capability that could apply anywhere gives you a list of markets, each won from a standing start. A ritual gives you a chain of consequences that tells you what to build next, and every adjacency is one the customer already connects to the moment, which makes the expansion legible to the buyer as well as the builder.

Nor is the pattern confined to after-action review. A shift handoff opens onto operating history, staffing, and quality. An inspection opens onto the asset record, maintenance, and compliance. A financial close opens onto controls, planning, and finance operations. Each has its own trigger, cast, and artifact, and each sits inside its own web of dependencies waiting to be followed.

Recurrence alone is insufficient

The concept degrades if every repeated task gets called a ritual, so it is worth stating what qualifies. Would the moment still occur without the product? Is there a recognizable trigger, and are the participants and their roles understood? Would a new participant be absorbed into it without being told it exists? Is participation reinforced by policy, regulation, professional norms, or operational necessity? Does consequential information surface there, must something binding emerge, and does skipping the moment carry an identifiable cost? And what holds the moment in place would that thing survive better tooling?

The cleanest line between a ritual and a habit is transmission. A habit belongs to a person and dies with them, while a ritual belongs to a role that outlives whoever occupies it, so a replacement is absorbed into the moment rather than trained into it. This is why the standing status meeting is the instructive failure rather than the obvious one. It recurs and it has a cast, but nothing triggers it except the calendar, nothing binding emerges from it, and attendance is a courtesy; it disappears the week the manager who convened it leaves. The habit dies with its holder and the standing meeting dies with its convener. A ritual survives both, because nobody in particular is holding it up.

Rituals also differ in what holds them there, and the difference decides whether a wedge is buying a fixture or a countdown. Some moments are held by consequence: physics, liability, regulation, an asset that runs through the night whether or not anyone is watching. Better software does not repeal the reason those exist. Others exist only because coordination is currently expensive, like the reconciliation meeting that convenes because two systems cannot speak to each other, and those are workarounds wearing the costume of a ritual. They pass every other test right up until the constraint that produced them is removed, which makes this the failure mode worth guarding against: it is the one that looks most like success.

Formality, finally, is evidence of institutional reinforcement rather than the thing itself. The debrief that started this essay was informal, and no regulation required it, yet operational necessity produced it reliably, with understood roles and real consequences. What matters is that the institution, in some fashion, insists on the moment.

Calendared repetition tells you a product might get opened weekly. A ritual tells you the product sits where the institution concentrates its people, its information, its authority, and its risk.

Why Daring notices rituals

Much of our time goes to narrow wedges in institutional settings, and the recurring difficulty is judging which narrow things can become large. A wedge locked into a genuine ritual is a signal we weight heavily, because so much of what we look for is already concentrated in the moment: an adoption occasion supplied by the institution, identifiable users and roles, existing urgency, proprietary information that surfaces nowhere else, stakes high enough to have produced the ritual in the first place, and expansion paths running in both directions. Together they amount to venture-scale density, a small entry point already holding the ingredients of a much larger company.

A ritual is optional. Plenty of significant companies began with wedges that created new behavior outright, and we would not hold its absence against a founder. The concept earns its place as a diagnostic, a way of seeing why a narrow product is adoptable now and expansive later, when the same product described capability-first would look like a feature.

When a founder can point to the moment their product enters, name who is standing in it, and explain what record must come out of it, we pay closer attention.

Some think CVCs are just traditional VCs backed by large corporations. Others assume every corporate investor is looking to acquire them or become a customer. The reality is far more nuanced.

In many ways, “CVC” has become like “family office.” It’s a catch-all label that groups together investors with very different structures, incentives, and ways of operating. The label tells you who the investor is. It doesn’t tell you how they’ll invest.

That’s exactly where most of the confusion starts.

What is CVC?

At the highest level, corporate venture capital is an investment arm or dedicated fund set up by a corporation to invest in external startups. Unlike traditional venture capital firms, which invest on behalf of outside limited partners (LPs), CVCs invest capital from a corporate parent.

At first glance, CVCs can look a lot like traditional venture capital firms. They source deals, conduct diligence, invest in startups, join boards, and support portfolio companies. But beneath the surface, they often operate under a different set of incentives, which shapes everything from how they evaluate opportunities to how they define success.

While CVCs have been around for decades, they've become an increasingly important part of the venture ecosystem. Today, corporate investors participate in nearly one in four venture deals, and more than 3,000 corporations invested in startups in 2025 alone. From healthcare and financial services to manufacturing and enterprise software, corporations are using venture investing to stay closer to the technologies and business models shaping their industries.

Why Do Corporations Invest in Startups?

One of the biggest misconceptions about corporate venture capital is that it’s simply an acquisition pipeline.

In reality, most large companies think about innovation through multiple levers. Depending on the opportunity, they may choose to build a solution internally, buy a company, partner with a startup, or invest in one. Each option serves a different purpose and comes with different levels of commitment, risk, and control.

Unlike an acquisition, a venture investment allows a corporation to learn from and work alongside a startup without fully integrating it into the business. That flexibility is one of the reasons CVC has become an increasingly important part of many corporate innovation strategies.

Just as importantly, not every CVC is optimizing for the same outcome. While traditional venture capital firms are primarily measured by the financial returns they generate for their limited partners, corporate venture capital programs often fall on a spectrum of strategic and financial objectives.

Here are some of the most common reasons corporations invest in startups:

1. Staying Close to Innovation

For many corporations, venture investing is a way to see where an industry is heading before those trends become obvious.

By building relationships with startups working on emerging technologies or business models, companies gain early visibility into shifts that could reshape their markets. These insights can influence product roadmaps, internal strategy, or future investment decisions.

2. Hedging Against Disruption

Sometimes the greatest risk to a corporation isn’t missing the next big opportunity, it’s being disrupted by it.

Rather than making a large acquisition or overhauling its core business too early, a corporation can invest in startups that are challenging existing ways of doing things. This provides a lower-risk way to learn, monitor market adoption, and build relationships while waiting for stronger evidence that a new technology or business model is gaining traction.

3. Exploring New Markets and Business Lines

Corporate venture capital can also help companies explore opportunities beyond their existing products or customers.

Rather than building an entirely new business from scratch or acquiring an unfamiliar company outright, corporations can invest in startups operating in adjacent markets to better understand customer needs, competitive dynamics, and long-term growth opportunities.

These investments often provide strategic insight into markets the corporation may eventually choose to enter, whether through partnerships, acquisitions, or internal product development.

4. Creating Commercial Opportunities

An investment can also be the beginning of a broader business relationship.

Portfolio companies may become customers, suppliers, technology partners, or distribution partners. In many cases, the value of a successful commercial partnership can far exceed the financial return from the investment itself.

5. Creating Strategic Optionality

Investing gives corporations a front-row seat to a startup’s progress without requiring the commitment of an acquisition.

Rather than buying a company before the technology or market has fully matured, a corporation can build a relationship, monitor execution, and better understand how the business evolves. While some portfolio companies are eventually acquired, many are not. The investment simply preserves optionality as the market develops.

6. Generating Financial Returns

Not every CVC is primarily driven by strategy. Some are expected to generate competitive venture returns alongside, or even ahead of, strategic outcomes.

These organizations often resemble traditional venture firms, with dedicated investment professionals, disciplined portfolio construction, and return expectations comparable to institutional VC funds.

The key takeaway is that there isn't a single reason corporations invest in startups. Every CVC balances strategic and financial objectives differently, and those priorities shape everything from how the investment team is structured to how it behaves as an investor. Understanding those motivations is often just as important as understanding the sectors a CVC invests in.

Recommended for you:

Not All CVCs Are Structured the Same

One of the biggest misconceptions of corporate venture capital is that every CVC operates the same way. In reality, there are a few different ways corporations can structure their investment programs.

The first distinction is where the investment capital comes from.

Some CVCs invest directly from the parent company’s balance sheet. In this model, there isn’t a separate venture fund. Instead, each investment is funded with corporate capital as it’s approved. This has historically been the most common approach, particularly among large public companies.

Other CVCs operate with dedicated committed capital. The parent company allocates a specific amount of money, such as a $50 million venture program, that’s reserved exclusively for venture investing. While the capital still comes from the corporation, it’s ring-fenced for the investment team to deploy.

The second distinction is how the investments are held.

Some corporations invest directly, meaning the parent company owns the startup shares itself. Others create a separate legal investment vehicle, such as an LLC or limited partnership, to make and hold investments. This gives the venture team greater operational independence and dedicated governance. Although these entities are legally separate, they are often still consolidated into the parent company's financial statements, which is why the term "off-balance sheet" isn't always technically accurate.

Finally, there’s how long the capital is intended to last.

Many CVC programs are evergreen, meaning they don’t have a predetermined end date. The corporation can continue funding new investments over time and may recycle proceeds from successful exits back into the program.

Less commonly, a corporation will establish a closed-end fund that looks much more like a traditional VC fund, with a fixed pool of capital, a defined investment period, and a target lifespan of around ten years.

Three Questions That Define a CVC

Not all corporate venture capital programs are built the same. Understanding how a CVC is structured can tell you a lot about how it makes investment decisions, how flexible it is, and what kind of partner it may be after writing a check.

Every CVC can be understood by answering three questions:

These three questions won't tell you whether a CVC is the right investor for your company, but they will give you a clearer picture of how the investment team operates, how resilient the program may be through changes in corporate strategy, and what you can reasonably expect from them as a long-term partner.

Lastly, there are a few edge case structures worth noting.

Multi-corporate funds
Some venture funds are backed by multiple corporations at once. This spreads risk and reduces dependence on a single corporate sponsor, which can make the structure more stable in certain situations.

Externally managed CVCs
In this model, the corporation provides the capital, but an outside VC firm runs the investment process. It can give the program more institutional investing experience while still keeping a strategic lens from the corporate side.

Evergreen funds with outside LPs
Less common, but some corporate-sponsored vehicles blend capital from both the corporation and external limited partners. These structures sit somewhere between traditional VC funds and pure corporate venture programs, often to scale capital or broaden alignment.

The CVC Behavior Matrix

The biggest mistake I see founders make with CVCs isn’t taking corporate money. It’s assuming all corporate investors are playing the same game.

Some CVCs behave almost exactly like traditional VCs. Others are deeply embedded in the business and can help open commercial doors. Others are primarily there to learn, monitor markets, or explore technologies that could shape the future of their industry.

The problem is that “CVC” tells you who the investor is, not what they’re optimizing for.

That’s why founders often walk away disappointed. They assume every corporate investor can unlock customers or accelerate commercial partnerships, when in reality many were never investing for those outcomes in the first place.

That’s why I created the following framework. Instead of asking whether a CVC is “strategic,” it helps answer a much more useful question:

What is this CVC actually optimizing for?

The CVC Behavior Matrix

• X-axis: Financial Focus → Strategic Focus

• Y-axis: Business Insights → Commercialization / M&A Outcomes

This isn’t a framework for how a CVC is structured. It’s a framework for how it’s likely to behave after making an investment.

How to Use This Framework

No CVC fits perfectly into one quadrant. Priorities evolve over time, leadership changes, and individual investments may serve different purposes than the overall fund strategy.

The point isn’t to label every CVC. It’s to understand what success looks like for the investor sitting across the table.

If a CVC sits closer to the Financial end of the spectrum, you should generally expect it to behave more like a traditional venture investor, where financial returns drive decision-making.

If it leans more Strategic, it’s more likely to evaluate opportunities based on how they can create value for the broader business, whether through learning, partnerships, new business lines, or long-term strategic positioning.

The vertical axis tells you what that strategic value actually looks like.

Some CVCs invest primarily to generate business insights. They’re looking to understand emerging technologies, customer behavior, and where their industry is headed. Others are investing to create commercialization or M&A outcomes, where success is more closely tied to commercial partnerships, product integration, or, in some cases, acquisitions.

Four Questions to Ask Every CVC

Once you understand the framework, the next step is figuring out where a specific CVC falls.

These are four questions I always recommend asking before taking corporate capital:

The answers to these questions will often tell you far more than the CVC’s website or investment thesis.

Why This Matters

One of the biggest misconceptions founders have is that taking money from a corporate investor automatically opens the door to enterprise customers.

That’s rarely how it works.

Even the most strategically motivated CVCs usually can’t force a business unit to buy your product. Commercial introductions can absolutely happen, but they’re still subject to procurement processes, internal priorities, budgets, and executive buy-in. A warm introduction is valuable, but it isn’t the same as a signed commercial agreement.

The opposite misconception is that every corporate investor is looking to steal your technology or quietly acquire your company. While acquisitions certainly happen, they’re far from the default outcome.

Take a Kubernetes security company, for example. Its value comes from becoming the trusted security platform across cloud providers and enterprises of every size. That business is often worth far more as an independent platform than as a product owned by a single corporation. A CVC can still generate significant strategic value through market insight, partnerships, and financial upside without ever intending to acquire the company.

The takeaway is simple: don’t evaluate a CVC by its brand alone.

Instead, evaluate:

Those answers will give you a much more realistic picture of the partner you’re bringing onto your cap table than the word “strategic” ever will.

That was a real founder question.

My answer (as always): it depends.

Here’s some background.

The founder asked for a sales intro.

Not a platform strategy. Not a community. Not a portal. Not a webinar called “Scaling Revenue With Intention.”

One person. One problem.

Four hours later, someone named Colton is walking them through the startup-support version of a complimentary multi-point inspection at your friendly neighborhood Ford dealership.

Here is the platform lead. Here is the operating partner. Here is the vendor list. Here is the founder portal. Here is a Notion doc. Here are all the other Notion docs linked from this Notion doc. Here is a Slack community. Here is a person who once ran sales at a company that sounds similar but is actually not similar at all. Here’s a dude who basically invented Square’s GTM. Or maybe it was Foursquare. Either way, everyone calls him Peaches. He’s a gawd.

Here is a 30-minute intro call before the real intro call.

The founder asked for the wiper.

Now they are stuck at the dealership.

That is where founder support has gotten weird.

Founders want help. Of course they do. They want the right intro, the right recruiter, the right fractional CFO, the right GTM person, the right person who has seen their exact problem before and can say, “Oh yeah, this is the part where everyone screws it up.”

But every investor is value-add now. Every fund has a platform person. Every platform person has a guy. Behind every guy is the real guy.

Every studio wants to studio you. Every accelerator wants to accelerate you. Frameworks. Packages. People who help startups help themselves. Strategic advisors with advisory strategy for RevOps. DevOps? BevOps¹?

Every expert has a framework. Every operator has an advisory offering. Every vendor has a startup package. Every community has a Slack.

The founder asks for help and suddenly they are not getting help. They are navigating help. The problem is that support has become another thing founders have to manage. Founders hiring “junior operators” to screen inbound VC calls. Platform team getting handed off to Ryan, he’s an associate on the operations team here and would be happy to kick things off. Okay fine that hasn’t happened… yet.

Too much platform turns help into overhead. More calls. More choices. More follow-ups. More vaguely useful things to sort through. The best support should reduce load. A lot of support now adds load.

Founders do not need more people standing around the car.

They need the exact person who can fix the exact thing.

There is a certain kind of startup helper who claims to do everything. They can help with GTM, hiring, sales, partnerships, positioning, fundraising, ops, finance, pricing, culture, strategy, executive coaching, category design, community, enterprise motion, and probably your marriage if the retainer is high enough.

No.

Founders do not need superheroes with menus like the Cheesecake Factory. They need people with lanes. The person who is terrifyingly good at one thing. In-n-Out. Double-double, ketchup & mustard instead, and fries well.

The person who has done the same painful, specific job enough times to know where the bodies are buried. The person who can look at the mess and immediately see the three things that actually matter.

If someone says they can do everything, they are probably not the killer you need for the thing currently on fire.

This is also not really a pitch for consultants.

Consultants diagnose. They advise. They deliver. And then they dip.

They make decks. Sometimes that is useful. Sometimes it leaves the founder with a strategic recommendation and twelve more things to do.

That is not what I mean.

The distinction that matters is pretty simple: a consultant says, “Here is what I would do.” A fractional operator says, “Here is what I am doing.”

More importantly, here’s how fast they’re going to get out of your way. Like a top tier pit crew. Scary efficient.

Not a massive internal platform team. Not a vendor marketplace. Not an accelerator, a studio, a consultant directory, or a menu of random helpers who all swear they are great with founders.

Just a bench of specific people for specific problems. That’s what we’re calling our crew of operators we’re partnering with for our Fund I portfolio support.

It is less bullshit between the founder and the help.

Recommended for you

The best fractional people are not drive-by advisors. They can grow with the company. At seed, you may need someone to clean up basic finance, hiring, or sales chaos. At Series A, that same person might help build the operating cadence, prep for the next fundraise, or hire the first full-time leader. They can be useful before the company is ready for the full-time executive. They build context over time. They help the founder avoid hiring too early, hiring too senior, or hiring someone full-time for a problem that is not yet a full-time job.

Fractional is not less than full-time. It’s founder-sized and startup speed.

The old flex in venture was having a giant platform team. Look at all our resources. Look at all our operating partners. Look at all our programs. Look at our portal. Look at our community. Look at our content engine.

The new flex should be restraint.

Look how fast we can get you to the right person. Look how little time we waste. Look how few steps there are. Look how quickly the problem gets solved. Look how cleanly the expert leaves when the work is done.

More headcount is not more help. More platform is not more value. Sometimes it is just more dealership.

The future of founder support is not a bigger service department.

It is a sharper bench.

¹Hire a head of BevOps and every day is a liquidity event.

TL;DR

• SFOs, MFOs, FFOs, RIAs, OCIOs, and consultants all use the same “family office” language. They are not the same thing.

• What separates them is not the label but the number of approvals between your conversation and a wire.

• For Fund I, the best targets are principals and small SFOs where one person’s conviction can move capital.

There are a few terms we can’t stand at Daring Ventures. One of them is “family office.”¹

Like most language that spends too much time inside opaque markets, the term has been stretched until it means almost nothing.

What used to describe a very niche structure built to manage one family’s significant wealth now gets applied to everything from billion-dollar investment offices to wealth managers, consultants, software vendors, placement agents, random LinkedIn profiles with “family office” in the headline, and people who have more in common with a club promoter than an investment professional.

When I started fundraising outreach for Daring Ventures Fund I, I naively thought the following words were basically interchangeable: family office, single family office, multi-family office, HNWI, UHNWI, and private wealth.

After a few months of cold outreach on LinkedIn to “family office” people, I realized how wrong I was.

Most of the “family office” executives I met with were not LPs. Some were service providers trying to sell software. Some were advisors. Some were consultants. Some were legitimate people who sat near wealthy families but had no control over allocation decisions. Some were just wasting my time.

Within five minutes of a call, I usually knew the next twenty-five minutes were going to be a waste. I would leave every call more frustrated.

Then my co-GP, Joe, told me to take a step back and ask one simple question:

Who actually controls the capital allocation decision?

That is the filter.

Once I started asking that question, the quality of my LP conversations improved and conversion improved. I wasted less time with service providers, consultants, and people who used family office language but had no ability to write a check.

The family office universe is real. It can be a valuable source of capital for Fund I managers. But the label itself tells you almost nothing.

What matters is how close the person is to the actual decision to allocate capital.

Why “Family Office” Means Almost Nothing Now

The term dates back to 1882, when the Rockefellers founded one of the first full-service single family offices to manage wealth from the industrial revolution. It was specific. It meant managing investments, tax, philanthropy, and other affairs for one family.

Now the term gets used for almost everything adjacent to private wealth.

It can mean a billion-dollar single family office with a full internal investment team. It can mean a multi-family office serving dozens of families. It can mean an RIA with a family office offering. It can mean a fractional service provider. It can mean a consultant. It can mean a founder with a big exit investing off their own balance sheet. It can mean a solo operator with a Squarespace website and a newsletter about “the evolving family office landscape.”

That ambiguity is the problem as a fund manager.

Nobody is going to introduce themselves by saying, “I do not control capital.” Nobody is going to self-identify as a time-waster. And nobody using the term “family office” is going to tell you upfront whether they are the check writer, the gatekeeper, the advisor, the service provider, or the person trying to sell you something.

So you need a better framework.

The Real Framework: Distance From the Check

The useful way to evaluate this market is not by title. It is by distance from the capital allocation decision.

There are two questions I try to answer as quickly as possible.

First: what is this person’s relationship to the capital?

Second: what is my access point?

Those sound similar, but they are different.

A person can sit very close to principal capital and still have no authority over it. A junior investment professional at a massive family office may work inside an organization managing billions of dollars, but their “yes” may only mean they are willing to bring it up in a meeting.

A person can also be an agent legally but function almost like a principal for fundraising purposes. A CIO at a small single family office is not investing their own money, but if the family has delegated investment authority to them, their conviction can cause capital to move.

That is the distinction that matters.

Principal-like does not mean “owns the money.”

It means the person is close enough to the capital allocation decision that their judgment can turn into a check.

Agent-like does not mean bad.

It means the person’s relationship to the capital is mediated. They may be a fiduciary, advisor, consultant, gatekeeper, or service provider. Some agents are extremely high-quality. Some are not. The question is whether their approval actually moves capital.

Before spending time on any “family office” lead, I now run five quick tests:

1. Whose money is it?
Is this a principal investing their own capital, one family’s capital, multiple families’ capital, institutional capital, or no capital at all?

2. Can I trace the capital back to a real economic principal?
Who created the wealth? Was it a founder, a family business, an operating company, a public company exit, an inheritance event, or another identifiable capital base? The point is not that the family needs to be famous. Many real families are intentionally private. The point is that there should be a real source of wealth behind the allocation.

3. Who actually has discretion?
Is the person you are speaking with authorized to make allocation decisions, or are they collecting information for someone else?

4. How many people need to say yes before a check gets written?
A principal may need no one. A small SFO CIO may need one family member. A large SFO may need an investment committee. A consultant may need to convince an institutional client’s committee. A connector may need to introduce you to someone who introduces you to someone else.

5. What does a “yes” from the person I’m speaking with actually buy me?
Sometimes it buys a check. Sometimes it buys a committee memo. Sometimes it buys an intro to the actual decision-maker. Sometimes it buys placement on an approved list. Sometimes it buys nothing.

Everything below should be read through that lens.

The label matters less than the path from conversation to wire.

The Starting Point: HNWIs and UHNWIs

HNWIs and UHNWIs are not family offices by default.

A founder with a large exit investing their own money is a principal. They may be an excellent LP. They may be sophisticated. They may have a team around them. But if they are making allocations for their own account, that is not the same thing as a family office.

A family office implies some separation between the principal who owns or created the wealth and the professional infrastructure managing it.

That distinction matters because it changes the dynamic.

A principal answers to themselves. Their investment process may be highly sophisticated, or it may be completely idiosyncratic. They may care about your market, your story, your network, your personal background, or a specific strategic angle. They can move quickly because there is no client committee, no fiduciary overlay, and no obligation to build consensus across unrelated stakeholders.

That can be very good for Fund I.

But it is not the same thing as raising from a family office.

Capital relationship: principal capital
Typical access point: the person who made or controls the money
What a yes buys: potentially a check
Fund I relevance: high, if they are aligned with your strategy and check size

Single Family Office (SFO)

Principal capital, delegated decision-making

A single family office is one entity serving one family.

A senior investment professional or team of investment professionals has been hired to direct investment strategy, manage the family’s wealth, and coordinate some combination of investments, tax, philanthropy, estate planning, reporting, and other family affairs.

Every SFO looks different because every family is different. Some outsource legal and accounting. Others keep everything internal. Some are mainly investment offices. Others look like full operating companies.

The common thread is maximum control, maximum customization, and maximum privacy.

The important thing to understand is that scale changes the decision path.

A small SFO may be as simple as one former allocator running the CIO function out of a home office, managing a handful of external manager relationships. In that case, the CIO is technically an agent, but may function almost like a principal. If the family has delegated discretion, you may only need to convince one person before a check can get written.

A large SFO is completely different.

Same label, different decision path.

The largest and most sophisticated SFOs are full operating companies. They may run an internal RIA, manage properties across multiple estates, oversee aviation and yacht management, operate a philanthropy arm, and maintain all the supporting verticals like legal, tax, HR, accounting, and reporting.

At that end of the spectrum, you are dealing with an org chart that looks more like a holding company. Senior leadership may come from the highest ranks of Wall Street. Gregg Lemkau leaving Goldman Sachs after 28 years as co-head of investment banking to run Michael Dell’s family office is a good example of the level of talent these platforms can attract.

The cost of running an SFO is significant, so families typically need a net worth of at least several hundred million dollars for it to make sense. That also makes SFOs the hardest for fund managers to find. They are exempt from SEC registration under the Family Office Rule, which means they do not show up on public databases like Investment Adviser Public Disclosure (IAPD). You usually find them through warm introductions and relationship networks. Occasionally you can find them through LinkedIn, but it is harder because many have limited public information and intentionally low visibility.

When you do find the right SFO, the check can be one of the most conviction-driven in your raise.

Investment decisions often come down to alignment with the family’s business, a personal connection to the fund manager’s story, a strategic interest, if they can access similar strategies elsewhere, or pure opportunism.

Less process, more judgment.

But you need to know who you are actually talking to. A CIO at a small SFO may be terminal. A junior investment professional at a large SFO may only be one node in a much longer internal process.

Decision-maker: CIO, CFO, or investment committee

Capital-control question: Does this person’s yes get me to a check, or just to the next internal conversation?

How to research: LinkedIn, firm website if they have one, warm introductions, PitchBook, Fintrx, Preqin, With Intelligence

Examples: Hillspire, MSD Capital, Willett Advisors

Recommended for you:

Multi-Family Office (MFO)

Multiple principals, shared agent structure

A multi-family office serves multiple affluent families through one shared platform. These families are typically large enough to want more than a standard wealth advisor, but not always large enough or interested enough to maintain a dedicated single family office.

A real MFO offers a full stack of services: investment management, tax, estate planning, risk management, philanthropy, reporting, and sometimes lifestyle or concierge services. The shared model lets multiple families pool the cost of infrastructure, talent, and access.

This is where the principal-agent relationship becomes more layered.

The capital still belongs to families, but now a professional firm is managing money for multiple unrelated principals. That usually means more process, more documentation, more fiduciary discipline, and more formal manager selection than a small SFO.

The right MFO can still be a real Fund I target, especially if it is boutique, alternatives-oriented, and led by a CIO or founder-partner with actual discretion.

But as MFOs scale, they become more institutional.

More families means more stakeholders. More stakeholders means more process. More process means your contact’s yes may not be enough.

That is why the label alone is not useful. A boutique MFO and a scaled commercial MFO may both call themselves “multi-family offices,” but the fundraising experience will be completely different.

A boutique MFO may behave more like an SFO: concentrated decision-making, strong relationships, more judgment, and more openness to conviction-driven allocations.

A scaled MFO may behave more like an institution: formal diligence, investment committees, approved lists, allocation models, and a slower path to commitment.

Many MFOs are Registered Investment Advisors, which means they show up on IAPD. That is useful because you can actually research them before reaching out.

How to evaluate one using IAPD:

Step 1: Pull the firm’s Part 2 Brochure. If “venture capital” or “private equity” is not mentioned, move on. It is probably not a fit.

Step 2: If venture or private equity is mentioned, go to the latest Form ADV, Section 5.K.(1), Separately Managed Accounts. Look at what percentage is allocated to “Securities Issued by Pooled Investment Vehicles (other than Registered Investment Companies or Business Development Companies).” If the allocation to alternatives is below 30%, it is probably not worth prioritizing. Above 30%, add them to your list.

MFOs run more formal processes than SFOs. Expect more diligence and a stronger focus on financial returns. They can be approachable at Fund I, but only if they have a real private markets allocation and you can get in front of the person with authority.

Decision-maker: CIO, senior partner, Head of Investments, or investment committee

Capital-control question: Is this a boutique office where one investment lead has discretion, or a scaled platform where every commitment goes through committee?

How to research: IAPD, LinkedIn, firm website, PitchBook, Fintrx, Preqin, With Intelligence

Examples: Bessemer, ICONIQ, Pitcairn, Whittier Trust

Fractional Family Office (FFO)

Partial agent structure, usually lighter capital control

A fractional family office serves multiple families but provides only part of the family office function, not the full suite.

This is not just a smaller MFO. It is a thinner service model.

That distinction matters.

An MFO is supposed to provide full family office infrastructure across investments, tax, estate planning, philanthropy, reporting, and related services. An FFO typically provides a narrower set of services for families that want more personalized attention than a standard wealth advisor but do not need or cannot justify a full SFO or MFO.

The reason to include FFOs here is that they often use similar language. They may say “family office.” They may serve wealthy families. They may sit close to capital.

But service depth changes what they can actually do.

Private markets infrastructure is often one of the first things that gets cut when the service model is lighter. A fractional family office may be coordinating advisors, organizing reporting, or helping manage a family’s financial life without actually running a serious external manager selection program.

That does not make FFOs fake. It just means they are usually not your primary target pool at Fund I unless they have a clear alternatives capability and an identifiable decision-maker.

Decision-maker: founder, CIO, lead advisor, but may ultimately roll up to a family principal

Capital-control question: Do they actually allocate to external private funds, or are they mainly coordinating services around a family’s wealth?

Fund I relevance: usually low to medium, unless there is a real alternatives program

Registered Investment Advisors (RIAs)

Fiduciary agent structure, varies wildly by alternatives capability

RIA is a regulatory category, not a useful fundraising category by itself.

It covers an enormous range of entities that register with the SEC or state securities regulators and have fiduciary responsibility to act in their clients’ best interests. Some RIAs have sophisticated private markets programs. Others are standard wealth managers with no alternatives infrastructure. Some market themselves as family offices because it sounds better, not because they are realistic LP targets.

The problem is that “RIA” alone tells you almost nothing.

You need to know which type you are dealing with.

Alternatives-Enabled RIA

An alternatives-enabled RIA is a mid-to-large firm, often with $500M to $5B+ in AUM, that has built a private markets capability either internally or through a platform like iCapital or CAIS.

These can be real targets. The decision-maker is usually the CIO, Director of Alternatives, or Head of Private Markets. They run a formal diligence process and may have a dedicated alternatives allocation.

Examples: Cerity Partners, Mariner Wealth Advisors

Capital-control question: Does the firm have a centralized alternatives function with authority, or are individual advisors making one-off client decisions?

RIA With an Internal Fund of Funds

Some RIAs create their own pooled private markets vehicle that allocates to external managers. This can make them relevant because the firm has already created a structure for accessing private funds.

Investment decisions usually go through a formal alternatives committee.

These can be worth pursuing if you can get in front of the right person, especially if the platform has an emerging manager allocation or a mandate that fits your fund.

Examples: Mercer Advisors, Hightower Advisors

Capital-control question: Who controls the internal vehicle, and does it allocate to funds at your stage?

Aggregator RIA

Aggregator RIAs are large networks that provide infrastructure to hundreds or thousands of advisor practices.

This category is harder to navigate because decision-making may be decentralized. Individual advisors may make their own allocation decisions, or the central investment team may maintain an approved list that advisors can use with clients.

For a Fund I manager, aggregators are usually better as later-stage targets once you have more signal and track record.

Examples: Dynasty Financial Partners, LPL Financial

Capital-control question: Are you selling to a central investment team, an approved list, or hundreds of individual advisors?

Wealth Management RIA

This is the least relevant category for Fund I.

These firms may manage $50M to $500M for high-net-worth clients, but they typically have no real alternatives infrastructure. They are usually focused on public equities, bonds, ETFs, financial planning, and maybe some liquid alternatives.

They may use family office language in marketing, but that does not make them LP targets.

Examples: Creative Planning, CAPTRUST

Capital-control question: Is there any actual private markets capability, or is this just a wealth management firm using family office branding?

RIA + VC/PE Hybrid

Some VC or PE managers are registered as RIAs and also manage capital for external clients alongside their own funds. Sometimes this can become relevant for Fund I if they have an emerging manager program, a fund-of-funds strategy, or a specific allocation reserved for Funds I to III.

But you need to be careful with this category. The fact that a VC or PE manager is an RIA does not automatically make them an LP target. You need to understand whether they allocate to external managers and who controls that allocation.

Examples of the broader category: Sequoia Capital, General Catalyst, Lightspeed Venture Partners

Capital-control question: Do they actually invest in external managers, or are they simply registered as an adviser because of their own fund management business?

For any RIA, run the same IAPD process as an MFO. Same two steps. Same 30% threshold as your signal for whether alternatives are actually embedded in the strategy.

Decision-maker: CIO, Director of Alternatives, Head of Private Markets, or investment committee

Best signal of a real target: dedicated alternatives capability, not individual advisors

Outsourced Chief Investment Officer (OCIO)

Pure agent with discretion

OCIOs are easy to confuse with the CIOs you are trying to reach at an MFO or RIA. They are not the same thing.

An OCIO is a third-party firm providing investment management services to organizations that do not have internal investing capabilities. Their clients are usually large institutional investors: endowments, foundations, pension funds, and insurance companies. Not wealthy families or HNWIs.

Conceptually, OCIOs are important because they show why “agent” does not always mean “no discretion.”

An OCIO is a pure agent. They are not investing their own money. They are managing someone else’s institutional capital. But if the client has delegated investment authority, the OCIO may actually control the allocation decision within an agreed investment policy.

That makes them agent-like in relationship to the capital, but principal-like in decision authority.

For Fund I, this distinction is usually more theoretical than practical. OCIO capital is institutional, process-heavy, and typically not available to first-time VC funds. I learned this after a conversation with someone at Commonfund. Given their clients are institutional investors, the earliest they invest in VC firms is usually Fund III+ with a minimum fund size around $100M.

For Fund I, skip OCIOs entirely. Revisit at Fund III when you have a track record and fund size that can meet their threshold.

Decision-maker: asset class lead, CIO, or investment committee inside the OCIO

Capital-control question: Has the client delegated discretion, or does the recommendation still need to go back to the underlying institution?

Fund I relevance: low now, potentially relevant later

Investment Consultant

Pure agent without discretion

Investment consultants sit right next to OCIOs and often get confused with them, so it is worth separating them clearly.

An investment consultant advises institutions on how to manage their portfolios. Their clients are pensions, endowments, foundations, insurers, and sometimes very large family offices. They help with investment policy statements, asset allocation, manager searches, due diligence, and performance monitoring.

But in the pure consultant model, they do not control the check.

They recommend. The client’s investment committee decides.

That makes them valuable, but in a different way. A consultant’s yes can get you into research coverage, onto a recommendation list, or in front of multiple institutional clients over time. But it usually does not produce a direct commitment.

It creates influence, not immediate capital.

This is the clean contrast with OCIOs. Both are agent structures. Both sit far from principal capital. Both work with institutional clients. But an OCIO may have discretion, while a consultant typically does not.

For Fund I, investment consultants are mostly relationship-building targets, not near-term LP targets. If you can get on their radar early, that may matter by Fund II or Fund III when you have a track record. But you should not confuse consultant interest with capital commitment.

Decision-maker: asset class lead, research lead, or consultant covering venture/private markets

Capital-control question: Does this person control capital, recommend managers to people who control capital, or simply know people in the ecosystem?

Examples: Cambridge Associates, Mercer, Aon, Callan, NEPC, Meketa, Wilshire, Albourne, Aksia

When the Access Point Is Not Capital

Once you start mapping people by capital control, the “watch out” category becomes much clearer.

Some people are legitimate but not LPs. Some are access points but not decision-makers. Some are selling services. Some are wasting time. Some are actively problematic.

The mistake is treating all of them like capital.

Family Office Service Providers

Attorneys, fund administrators, placement agents, consultants, wealth tech platforms, reporting tools, and other vendors fall into this bucket.

They use family office language on their websites and LinkedIn profiles because their clients are family offices. The language is designed to attract clients, not to tell you who they are.

These companies may be legitimate. Some are very good businesses. They are just not LPs.

The key distinction is that they sell to capital. They do not allocate capital.

Common tells:

• They describe who they serve, not what they do with capital

• Their title includes “advisor,” “consultant,” “strategist,” “solutions,” “business development,” or “partnerships”

• They have multiple short-tenure roles across different firms

• They post constantly about family office trends, education, or how the family office landscape is evolving

• The employer has large headcount

• The company page has a contact form, demo request, or lead capture element

• There is no identifiable economic principal behind the capital because there is no capital

Examples: Addepar, Arch, Juniper Square

Again, these are legitimate companies. They are just not LPs.

Capital-control question: Do they allocate capital, influence capital, or sell products to people with capital?

Fake Family Office Individuals

This is where it gets more serious.

There are people who pose as family office executives to take advantage of new fund managers. Fraud and time-wasting are real hazards when you are fundraising, especially when you are a first-time fund manager trying to build momentum.

The best way to protect yourself is to stay anchored to the same framework.

• Whose money is it?

• Can I trace the capital back to a real economic principal?

• Who has discretion?

• How many people need to approve before a check gets written?

• What does this person’s yes actually buy me?

If those questions never get answered, be careful.

The Phantom Anchor LP

They express serious early interest, imply they could anchor your raise, and request your full data room. Then they string you along for months.

Their goal may be competitive intelligence. They may be shopping your deal to other GPs for a finder’s fee. They may be trying to look important. Either way, the process never moves.

Every meeting is one-on-one. They cannot produce a reference from a prior fund commitment. They avoid specifics on allocation process. They never introduce the actual decision-maker. The conversation never advances to legal docs.

Red flag: high enthusiasm, no verifiable allocation history, no process.

The Fee-Before-Capital Placement Agent

They present as a capital introduction specialist with access to a curated HNW or family office network. Then they ask for an upfront retainer, platform fee, or access fee before raising anything.

That is a major red flag.

Legitimate placement agents generally work on success fees tied to capital raised. Any placement agent asking for money before they raise money should be treated with extreme skepticism.

Red flag: you pay before capital moves.

The Conference Circuit LP

This person is a regular at emerging manager summits, LP/GP events, and private wealth conferences. They collect decks aggressively, have a polished pitch about their “family office,” and seem to know everyone.

But they never actually wire capital.

Often this person is after competitive intelligence, relationship leverage, consulting opportunities, or deal flow they can repackage elsewhere.

The most useful question you can ask other fund managers is simple:

“Have you ever seen them actually close?”

Red flag: universally visible, constantly networking, no confirmed commitments.

The Offshore “Sovereign” Fund

This usually starts as an inbound inquiry from a foreign entity describing itself as a sovereign wealth fund, royal family office, dynasty fund, or government-adjacent investment platform.

They express interest in a large commitment quickly with minimal diligence.

That should make you more skeptical, not less.

The goal may be advance-fee fraud, sanctions evasion, money laundering, or something else you do not want anywhere near your fund. This is not just a wasted relationship. Accepting capital from the wrong source can create catastrophic legal exposure.

The key red flag is that the speed and size of interest are inversely proportional to the amount of diligence.

No real $10M LP commits in two weeks over three emails.

Red flag: huge check, little diligence, unclear source of wealth, low-transparency jurisdiction.

The Credential Blender

This person presents with a confusing mix of high-status credentials, vague institutional affiliations, and family office language.

They may claim to be a CFA, former banker, advisor to families, partner at an investment platform, or representative of private capital. But the details do not verify. The website is thin. The one-pager is strange. The story changes. The capital source is vague. The process is unclear.

This type is dangerous because they can sound sophisticated in a first conversation. But if you cannot verify the credentials, the firm, the capital base, or the decision-making authority, do not treat them like an LP.

Red flag: impressive-sounding biography, unverifiable details, no clear capital source.

The One Question That Cuts Through All of It

Before you spend time researching investment strategy, checking IAPD, sending a deck, or booking the call, ask yourself one question:

How close is this person to the capital allocation decision?

The family office universe is genuinely useful for Fund I managers, but the signal-to-noise ratio is brutal until you stop taking titles at face value. “Family office” is not a category. It is a clue. The job is to figure out whether the person in front of you controls capital, influences capital, sells to capital, or is just using the language of capital.

¹ More on these later.

tl;dr: the cheaper the memo gets, the more expensive the name on it becomes.

fast framework: price the apology, not the artifact: who has to show up when it breaks?

open question: if clients can sue the firm but not the model, is the entire industry just liability arbitrage with good branding?

A few weeks ago, the FT reported that Kirkland & Ellis, the largest law firm by revenue, was committing $500 million to build internal AI tooling. Firm chair Jon Ballis explained the logic plainly: off-the-shelf tools like Harvey are “raising the floor for everyone,” so any real edge has to come from something a rival can't simply license.

Like with all AI news, cue the “______ is cooked” takes. In the hotseat this time: Harvey.

And like many reactionary takes, it smuggles in a category error worth naming because it’s one I’ve been guilty of myself.

At one point in my life, I was a complete outsider to professional services. I marveled at the prospect of being one of those uber-smart young professionals cranking out that brilliant work product.

Generally, I thought:

• BigLaw writes documents.

• MBB consulting firms make slides.

• BB+ investment banks build models and pitchbooks.

AI can write documents, make slides, and build models. Ipso facto, AI is coming for legal, consulting, and financial advisory.

AI will absolutely make the artifacts of professional services cheaper to produce. It will produce first drafts, diligence summaries, market maps, research memos, buyer lists, comp sets, board pages, redlines, operating-model slides, and first-pass recommendations.

But let’s take a step back. The same way Carvana isn’t really in the “selling cars” business, Kirkland isn’t in the “writing documents” biz. Carvana is a subprime auto loan origination machine with a cute vending machine gimmick. Kirkland is an authority broker & professional pointy-end-of-stick.

Clients do not hire Kirkland because they need a document. The document is just a delivery mechanism for “we mean business.”

The product is the message. The message is medium-agnostic.

Could be a fax. A candygram. A poem. If I was jockeying for position in a liquidation and I got a hand-quilled couplet on Kirkland & Ellis cowhide I might be more intimidated.

We signed the deal at midnight, left you out;
Subordinated now, you'll do without.

But the message isn’t the words either.

Think about the same cowhide with the same couplet. Except this time it’s signed Kirklandt & Ellison. Everyone’s laughing. “We’ve got BigLaw at home” energy.

The words are just another medium. The message is who’s saying it.

People don’t hire these firms because they need brilliant memos. They hire these firms to convert work product into reliance. Into something they can act on: sign, disclose, reorganize, defend to a board, take to market, close.

The deliverable isn’t the service. It’s the residue. It’s the container. The thing that survives the meeting and goes into the data room. The thing journalists can obtain and Congress can subpoena.

The product is the moment a client is willing to act, under someone else's accountability. Memos and slides and CIMs are worthless as work product without the trust that converts it to revenue.

The Reliance Layer

So if they aren’t selling documents, what do they make?

Let’s call thing being sold the Reliance Layer: the institutional machine that turns work product (whether paralegal, word processor, or AI-assisted) into trusted action.

It’s a combination of firm knowledge, brand equity, regulatory insulation, lobbyist influence, and a shortlist of names who will end up holding the bag if/when shit goes down.

Every elite firm has one; they just brand it differently. I know because I sold it. In banking, we called it “unparalleled industry coverage” and “24x7 transaction execution support.” I’m sure BigLaw and MBB have equally irritating terminology. Same same, but different, but also kinda the same.

To uncomplicate this, let’s complicate it first. I think about professional services the way airlines think about themselves. An airline looks like it sells flights. It actually manages perishable inventory. The seat flies whether or not someone is in it, the plane costs the same either way, and a seat that takes off empty is revenue that never existed. So the metric that runs the business isn’t flights flown or tickets sold. It’s PRASM: passenger revenue per available seat mile. Yield per unit of capacity that expires on a schedule.

Professional services has the same shape. A firm looks like it sells documents. It actually manages perishable capacity: a partner-hour, a principal review slot, a senior-banker call. Each one converts to revenue or expires worthless, and you can’t warehouse Tuesday. So the metric that matters isn’t memos produced.

It’s the firm’s equivalent of PRASM: revenue per available unit of trusted expert capacity. The Reliance Layer is the airline. The scarce input is not artifact production. AI just made that nearly free.

The inventory is the layer itself: the supervised judgment that converts artifacts into something a client will rely on.

So the metric measures the layer, and the test that follows asks who owns it. For any AI bet in or around knowledge work, ask three questions:

1. Artifact or conversion? Does the position own the production of work product (commoditizing fast) or the Reliance Layer that converts it into trusted action (where the rent lives)?

2. Who captures the PRASM uplift? AI raises throughput per expert. Does the layer’s uplift accrue to the incumbent firm, to a vendor, or get competed away to clients? Strong demand plus a binding capacity constraint means the firm keeps it; weak demand and no way to redeploy saved labor means it leaks to price.

3. Where does accountability sit? Whoever signs, insures, and bears the liability owns the durable position. A model can generate legal text, but it can’t be admitted to the bar, lose a license, or take the board’s call when someone asks whether this is safe.

The closer a business sits to low-accountability artifact production, the more AI commoditizes it. The closer it sits to high-stakes reliance, the more valuable the layer becomes. And the more defensible. The layer is not a defense against AI. It’s the intake valve.

Recommended for you:

In law, “AI passed the bar” is the same category error in miniature. Passing the test is not the job. The job is entering an accountability system: admission, privilege, sanctions, malpractice exposure. BigLaw AI isn’t document generation; it’s yield management under liability.

In consulting, a generic tool can say “centralize procurement” or “exit the segment.” The partner knows whether this CFO can force the regions to comply, or whether this CEO has the board support to do it. By this client, at this moment. Most consulting value lives in that clause, and a model doesn’t have it. The production side is already moving. McKinsey’s internal assistant runs north of 500,000 prompts a month (roughly one every five seconds, around the clock). But adoption of the tool is not the same as capture of the value.

In M&A, a CIM is not an auction and a fairness deck is not board confidence. A dead deal, a busted process, a missed market window. That’s dead inventory, opportunity cost with nicer fonts. The bank’s scarce asset isn’t analyst hours; it’s the senior capacity to turn a live situation into a closed fee before the window shuts. The economics are lumpy. A handful of high-stakes mandates drive the year. That is exactly why senior judgment, not throughput, is the binding constraint, and why faster artifact production alone doesn’t move the number that matters.

A generic market summary is exposed; a live board recommendation is not. A first-pass contract summary is exposed; a negotiated risk allocation in a contested deal is not. In every case the artifact gets cheaper and the conversion gets more valuable.

Who owns the Reliance Layer is the investment question.

There are three answers, and each is a position.

Incumbents. If demand is strong and trusted capacity is the bottleneck, AI lets a firm push more high-value work through the same senior people. PRASM goes up and the firm keeps the spread. Kirkland’s $500 million isn’t a Harvey-replacement story; it’s a Harvey-absorption story. The bottleneck was never generating drafts. It’s the internal trust infrastructure that turns drafts into filings: routing, review, supervision, privilege, billing. Build that, and the firm can safely consume more vendor AI than any rival and convert more of what it consumes. They’re not trying to become Harvey. They’re trying to become Harvey’s biggest customer. More Harvey in, more Kirkland out. The levers are realization, write-offs, leverage, and the move toward value-based pricing where the work supports it.

Vendors. The tooling layer is real and growing fast. In banking, Rogo raised a $160 million Series D in 2026 at a $2 billion valuation; its agent, Felix, screens deals, drafts CIM sections, runs buyer outreach, and works through data-room diligence for more than 35,000 professionals at over 250 institutions, Lazard, Moelis, and Rothschild among them. The vendor risk isn’t that incumbents clone it. It’s abstraction: if the firm owns context, permissions, privilege, QA, routing, and sign-off, the vendor becomes one callable service inside the firm’s machine. The vendor wants to become the workflow; the firm wants to own the workflow. If Kirkland’s build works, Harvey sells more, not less. It just sells it as an ingredient. That fight is live and unsettled. (A major bank’s own growth-equity arm sits on Rogo’s cap table.)

Where do vendors win outright? In greenfield or low-accountability work with no incumbent Reliance Layer to defend, and in becoming the system of record. A vendor that owns the data and the workflow can starve the incumbent’s trust infrastructure from below (the bull case for the tooling layer). The Rogo client roster tells you which one is furthest along.

Clients insourcing. The cheaper artifacts get, the more work moves in-house. Real, but concentrated in exactly the low-accountability tasks the framework already flags as exposed.

None of this is a clean build-versus-buy choice, and treating it as one is its own mistake. The durable pattern is internal workflow plus external infrastructure: firms own the relationship, context, judgment, and sign-off; vendors supply models, data, and rails.

My view is that Kirkland will build a proprietary platform which will make it easier to keep licensing third-party tools. Their play is to own the path from artifact to action, not the tooling to get there.

So the sharper question for any position is not “build or buy” but “who ends up owning the Reliance Layer once the dust settles?”

The underwriting heuristic: run it at 10x

The cleanest stress test is to assume the models get ten times better. If a position only produces artifacts, 10x kills it. If it owns the Reliance Layer, 10x may strengthen it: more raw output flooding into the same scarce conversion capacity raises the value of whoever can route, supervise, and stand behind it.

The firm with the widest layer gets to has the most capacity to weather the storm. So underwrite the conversion layer, not the generation layer. The winners aren’t the firms or tools that generate the most output; they’re the ones that convert the most output into trusted action per scarce unit of capacity.

Pressure testing the Reliance Layer

The Reliance Layer isn’t magic. There are four real pressures on the thesis. Each comes with a mitigation, and the mitigations are not evenly distributed.

Pricing leak

The pressure: if AI just makes artifacts cheaper and demand is soft, clients won’t pay old fees for faster work and the uplift leaks to price. The shift toward value- and outcome-based pricing is the tell for who’s actually capturing the gain.

The mitigation: a pricing leak needs a forcing function. Medicine got outcome-based payment because doctors don’t control the payer. Lawyers control everything. The fee rules are written by lawyers, enforced by lawyers, and appealed to judges who are lawyers, in a country where lawyers are still the largest bloc in Congress and half the Senate.

The apprenticeship problem

The pressure: firms train senior judgment through junior artifact work. Strip out too much of it without redesigning training and you optimize this year’s margin by starving the judgment you sell in ten years.

The mitigation: a precedent pointing the other way. The last time the junior artifact got automated was the spreadsheet, and analysts barely existed before the spreadsheet. The analyst class didn’t shrink. It was born.

Vendors moving up

The pressure: a tool that quietly accumulates context, accountability workflow, and the client relationship stops being a component and starts becoming the firm.

The mitigation: As of right now, non-lawyers can’t own law firms in the US outside of a handful of states with pilot programs.

Demand elasticity

The pressure: the layer’s whole PRASM story assumes a binding capacity constraint. No constraint, no story. Just deflation.

The mitigation: the legal book hedges itself. Boom times, bust times, legal work still exists. The window doesn’t close as much as it rotates. It is not a coincidence that the firm on top of both cycles is the one spending the $500 million.

The mitigations rank the pillars. Consulting has no licensing wall. Banking answers to regulators who aren’t bankers, and its vendors are already inside. Law regulates itself, prices itself, and hedges itself. Which is why the anchor of this essay is a law firm (and an area we are very bullish on investing in).

What to take home

The question to ask of any AI play in knowledge work isn’t “can it make the deliverable?” Of course it can. It’s: who reviews it, owns it, prices it, and stands behind it when the recommendation becomes a contract, a filing, a board vote, or a closed deal?

The “Harvey is cooked” takes had it exactly backwards. Tools that raise the floor only threaten firms that compete at the floor. Kirkland isn’t spending $500 million to use less Harvey. It’s spending $500 million on the internal trust infrastructure that lets it safely consume more Harvey than anyone else: routed, reviewed, privileged, signed. Not the name under the document. The name under ten times the documents.

TL;DR

• Bubbles do not just leave assets behind. They leave work.

• We think the forgotten story of the telco bubble was the software that made the wreckage usable.

• AI’s version is already forming in power, leases, GPU debt, inference bills, and procurement.

Over the last twelve months, the AI bubble question has become a regular part of LP conversations.

At first, it arrived as a perfunctory diligence question near the end of the meeting.

“And what if this is all a bubble?”

It had the energy of a question people mostly ask to prove they asked it. But that is not quite right either, because everyone actually wanted to hear the answer. It was more like checking under the bed for the boogeyman. You do not really believe in the boogeyman. You still check.

Yes. We checked. Twice, obviously.

Then there was the other version, which I liked much more: the LP who came in with an analogy already chambered.

“Isn’t this just fiber, except the fiber depreciates every three years?”

Those are my people. I think that way too. Give me a historical analogy, a broken capital structure, and one weird mismatch in asset duration, and I am happy. The question was never annoying. It was just early.

Now everyone asks it.

And once everyone asks the same question, the question itself stops being the interesting part. The interesting part is what the question assumes.

“What happens if the AI bubble pops?” makes the pop sound like the event. One moment of market-clearing violence. A chart going down. A few spectacular bankruptcies. Some awkward board meetings. The usual cleansing fire.

But bubbles do not just pop.

They leave things behind.

People reach for analogies because there is no clean way to talk about a buildout this large while it is still happening. Fiber. Shale. Railroads. Nuclear. Cloud. Every prior boom gets pulled off the shelf and held up to the light.

The analogies are useful because they turn a live, messy operating reality into a clean historical shape.

Fiber is the comfortable one. The story goes like this: capital overbuilt the network, investors got wrecked, bandwidth got cheap, and society inherited the infrastructure for the next internet.

Perhaps.

Shale is darker. Real output, bad returns, constant reinvestment, asset depletion disguised as growth.

Maybe that too.

Nuclear is less fun to talk about, because the punchline is a power bill. Cost overruns do not vanish. They migrate into rate cases, surcharges, and a thirty-year argument over who should have known better.

Maybe that too.

The problem with all of these analogies is not that they are wrong. It is that they are too clean. They make the cycle sound like a question of asset value. What is the salvage value of the fiber? What is the residual value of the GPU? Who owns the railroad after receivership? Who gets stuck with the plant?

Those are good questions. They are not the only questions.

Buildouts do not just leave assets behind.

They leave a lot of work.

Contracts. Permits. Leases. Tariffs. Billing disputes. Vendor fights. Security gaps.

It becomes consulting. Litigation. Regulation. A private-equity rollup. Which then becomes repeatable workflows, scalable markets, and real software opportunities.

Boom or bust, the rise of process is when software has something to sell because the work it inevitably creates has become somebody’s job.

One power bill, ten workflows

Start with the power bill.

A data center does not just use electricity. It changes who has to reserve capacity, who pays for grid upgrades, who gets priority in an interconnection queue, which costs can be passed through to ratepayers, which projects get delayed, and which regulator has to explain why the same household is paying more for a system it does not think it uses.

That is the part the bubble analogies tend to miss. The asset is easy to point at. The work is not.

In the regional grid market that covers much of the Mid-Atlantic and Midwest, capacity prices for the 2025-26 delivery year cleared at $269.92 per megawatt-day, up from $28.92 the prior year. That is an 833% increase. Data centers accounted for 40% of the December 2025 capacity costs. By 2028, one estimate has the average household in the region paying $70 more per month on its electricity bill.

It’s only one price signal, but it has a cascading effect.

The ratepayer gets a bill. The public utility commission gets a hearing. The state legislature gets a bill. The utility holdco gets a new five-year capital plan. The EPC (energy, procurement, construction) contractor gets a revised schedule. The equipment supplier gets an order book. The tax office gets an abatement fight. The insurer gets a new exposure. The auditor gets a new lease question. The procurement team gets a vendor problem. The associate gets a PDF named final_final_v7.

GE Vernova entered 2026 with 83 gigawatts of equipment under firm order. Siemens Energy is sold out years into the future. Transformer lead times are now measured in quarters, not months. EPC contracts that used to treat equipment as a passthrough now negotiate over turbine slots, transformer reservations, and switchgear queues as first-class economic rights.

Every one of those is a workflow.

The same thing is happening in financing.

A data center lease used to be a real estate document with power density. Now it can be the anchor for a special-purpose vehicle, the collateral behind an investment-grade bond, the proof point for a private-credit fund, the reason a utility files a new rate class, and the invisible bridge between a hyperscaler’s balance sheet and somebody else’s cost of capital.

Meta and Blue Owl’s Louisiana structure has a name now. People call it the Beignet template. Blue Owl owns most of the development entity and bears most of the cost; Meta owns the minority stake and signs the long-term lease back. The debt is rated like high-grade paper but yields like something that knows it is carrying more risk than the rating suggests. Thirty-plus hyperscale campuses are reportedly using or copying versions of the structure.

That is not just a financing innovation. It is a workflow factory.

How do you reconcile a 25-year physical asset against a shorter accounting treatment of the lease? Who carries the renewal risk? Whose insurance covers the residual? Where in the waterfall does the gas supplier sit relative to the preferred equity? What disclosure is required when the same anchor tenant is also a meaningful counterparty elsewhere in the capital stack?

None of these questions sound like software.

That is usually the first clue.

In West Texas, the same thing is happening around behind-the-meter power. The simple thesis is easy to understand: the Permian produces so much associated gas that it can trade at deeply negative prices, and that stranded gas can power colocated compute faster than a utility interconnection. The hard part is everything after the sentence.

Does the project have a named, creditworthy offtake? Does it have an air permit? Does it have turbines secured? Who owns the gas supply? Does the gas contract match the data-center contract? What happens in year fifteen if the gas plant still has a decade of useful life and the compute customer has moved on? Is the water right real? Does air cooling work in a West Texas summer without destroying the power economics? Does the developer have a project, an option, or a press release?

Again, workflows.

Some of them will be absorbed by lawyers. Some by utilities. Some by private-credit analysts. Some by consultants. Some by whichever developer can afford to hire the person who knows the answer. Some by an associate who becomes the only person in the room who understands why the tax abatement, GPU depreciation schedule, gas supply agreement, and tenant guarantee do not quite line up.

A small slice will become software companies.

The companies that get built will not announce themselves as “AI infrastructure software.” They will look, at first, like someone fixing a very specific workflow that suddenly became too expensive to keep fixing by hand.

The asset changes. The cleanup work rhymes.

There is one prior cycle close enough to look at carefully.

Recommended for you

The story people remember

Telecom is the cleanest prior cycle because the public memory of it is so settled.

Everyone knows the headline version. In the late 1990s, capital flooded into fiber. Carriers borrowed too much, built too much, and assumed demand would arrive on schedule. It did not. The bubble broke. Balance sheets collapsed. The physical network survived. Bandwidth got cheap. The next generation of internet companies inherited better infrastructure than the first one had.

That story is true as far as it goes.

It is just not the part that matters most here.

Cheap assets explain why better-capitalized operators could buy distressed networks for pennies on the dollar. They do not explain why the decade after the bust produced a quiet wave of companies most people outside telecom have never heard of. To see those companies, you have to look between the bankrupt asset and the eventual internet abundance.

That is where the work was.

In April 2002, PSINet sold its U.S. national backbone to a four-year-old company called Cogent for $10 million in cash, plus $3 million for the right to do due diligence.

PSINet had spent billions building that backbone. Cogent’s founder Dave Schaeffer had previously run a taxi fleet and a commercial real estate business, and by the late 1990s he had been close enough to dark-fiber economics to know what was about to come out of the Southern District of New York bankruptcy court. Between 1999 and 2002, Cogent did the same trade thirteen times. By 2014, Schaeffer was telling investors Cogent carried roughly a fifth of the world’s internet traffic.

That is the version of the telecom bust I remember from NPR on the way to soccer practice: cheap assets, better operators, second-owner economics.

It’s only half the story though. The neat half. Underneath all the “fiber” were carries that did not know what they owned, enterprises paying invoices for circuits no one used, and a network of operators who had spent enough time inside the pipes to know where the bodies bundles were buried.

The first problem was inventory.

After the bust, the surviving carriers had absorbed fragments of dozens of failed networks. Routes, switches, circuits, leases, customers, billing systems, and physical plant had been stitched together through bankruptcy sales and emergency integrations. Knowing what the carrier owned was no longer a database question. It was an archeological problem.

By 2003, SBC and Telecom Italia were phasing out a Bell System-era network inventory system called TIRKS and replacing it with software from a startup called Granite Systems. Telcordia bought Granite in 2004 for roughly $70 million. Two years later, Amdocs paid $375 million for Cramer Systems, a UK company solving a similar problem. NEC bought NetCracker for around $300 million. Oracle bought MetaSolv for $219 million.

Nobody outside telecom has heard of most of these companies. That is the point. They were not consumer internet stories. They were not Web 2.0. They were software companies built inside the operating mess the bust left behind.

The second problem was the bill.

A 2005 Gartner study found that up to 80% of enterprise telecom invoices contained some kind of mistake. Enterprises were paying for disconnected circuits, abandoned wireless lines, duplicate services, and contracts no one on staff remembered authorizing. The network had become cheap enough to sprawl and complicated enough that no one could audit it by hand.

A category called telecom expense management appeared to clean this up. Tangoe, Rivermine, ProfitLine, Asentinel, and others showed up. Al Subbloie founded Tangoe in Orange, Connecticut in 2000 after a decade inside enterprise telecom procurement (our kind of guy).

He knew where the invoices went to die. Tangoe went public in 2011 and was taken private by Marlin Equity in 2017 for $242.6 million.

The third problem was interconnection.

Once bandwidth was cheap, the scarce thing was no longer just owning fiber. It was being in the right place, with the right networks, under the right operating agreements. The Palo Alto Internet Exchange had been one of the most important neutral peering points in the United States. More than a hundred networks met inside one Bryant Street building.

Metromedia Fiber Network had paid $75 million for it in 1999. When Metromedia filed Chapter 11, Switch and Data bought PAIX out of bankruptcy in 2002 for roughly $40 million. In 2010, Equinix bought Switch and Data for $689 million.

The asset had not changed. The operating context around it had.

These stories are less famous than Cogent because they do not fit the clean bubble narrative. They were not “buy cheap fiber and wait.” They were the systems that made the post-bust network usable.

The bust left behind a fragmented, distressed, mis-billed, mis-provisioned operating layer. Then a generation of companies showed up to fix it, one workflow at a time.

Schaeffer had been close to dark-fiber economics. The Granite engineers had been inside carrier inventory. The Equinix founders had been inside the early internet exchange points. Subbloie had been inside enterprise telecom procurement.

None of them announced themselves as part of the bust.

They announced themselves as people who knew what was broken.

The biggest winners did not look like the asset owners

The next layer was stranger.

By the time Jeff Lawson founded Twilio in March 2008, AT&T had told him three times that integrating telecom into a software product would take two years, cost millions of dollars, and begin with rolling out copper wire.

He had been founding CTO at StubHub. He had built a point-of-sale system at a southern California skate-and-surf retailer. He had been hired by Amazon in late 2004 as one of the first product managers on AWS. He had never worked inside a carrier. He had hit the carrier wall from the outside three times before he started building for it.

That was the workflow.

Twilio’s first product launched November 20, 2008. Lawson rickrolled the TechCrunch editor from a few lines of code. By 2016 the company was public. By fiscal 2025 it generated $5.07 billion in annual revenue on $234 million of pre-IPO venture capital. Twilio’s peak market cap was north of $60 billion.

None of that works in 1998.

The post-bust glut of cheap termination, the dense carrier-neutral interconnects at Equinix, the collapse in wholesale voice and SMS pricing, and the Amazon cloud Lawson had just helped build all had to land first.

Twilio is what you get when the work beneath the telecom boom compounds long enough that the original infrastructure can disappear behind an API.

That same migration happened elsewhere.

Akamai was founded before the bust, but the business only fully made sense once cheap fiber, cheap colocation, and carrier-neutral interconnection made it economical to put servers near users at global scale. Prolexic was built because abundant, always-on connectivity created abundant, always-on attack surfaces. Dyn and ThousandEyes helped enterprises manage the fact that the internet had become mission-critical without becoming internally manageable. Apptio and CloudHealth later did something similar for cloud spend: they helped the enterprise understand a new infrastructure bill that had become too complex to govern by hand.

The important thing is not that these companies were farther away from the physical asset. Distance is the wrong frame.

The important thing is that they sold to a broader buyer.

Granite and Cramer sold to carriers. Tangoe sold to enterprises trying to clean up telecom spend. Akamai, Prolexic, Dyn, and ThousandEyes sold to enterprises whose businesses now depended on the network. Twilio sold to developers who never wanted to talk to a carrier at all. Apptio and CloudHealth sold to CIOs, CFOs, and line-of-business owners trying to govern infrastructure they did not own and could not turn off.

The founder pattern was not identical in the narrow biographical sense. Some had worked inside the buyer. Some had worked inside vendors. Some were academics. Some were developers who kept running into the same wall from the outside.

But they shared something more important than a résumé category.

They were close enough to the workflow to know which abstraction would be trusted.

That is the line between a good analogy and a lazy one. The lesson from telecom is not that every post-bust software company is founded by someone who worked at the original asset owner. That is too narrow, and the history does not support it cleanly. The lesson is that the durable companies were built by people close enough to the operating pain to understand the order in which the mess had to be fixed.

Inventory before automation.

Interconnection before abstraction.

Billing visibility before spend governance.

APIs after the underlying cost structure had collapsed far enough that the API could hide it.

The companies looked obvious only after the layer below them had become stable enough to disappear.

You cannot trade what you cannot provision

There is one piece of the post-bust story that did not produce durable software companies.

In 1999, two industries arrived at the same conclusion at the same time. Bandwidth would soon trade like wheat.

Enron Broadband Services built a trading desk with proprietary network provisioning software and expensive Sun servers. Arbinet ran a NASDAQ-style exchange for international voice minutes and IP capacity, IPO’d in December 2004 at $17.50, and peaked near a $500 million market cap. RateXchange, Band-X, LighTrade, and Global TeleExchange built variations on the same model.

Every one of them failed.

Enron Broadband failed first and most spectacularly: $494 million in losses over the first nine months of 2001. Arbinet held on for six more years before being acquired for parts in 2010.

The founders came from energy trading, commodities finance, and in one case a former military pilot turned deregulation policy director. None of them had run a carrier. None of them had provisioned a circuit. They saw what looked like a commodity and missed that the commodity had not yet become operationally tradable.

The diagnosis was not complicated.

The capacity was not fungible. A circuit between New York and Los Angeles was not interchangeable with one between Chicago and Atlanta. Provisioning was slow enough that the spot-market clock did not work. Asset owners refused to yield pricing control to a third-party exchange. The software required to dynamically measure, route, and trust packets across competing carrier hardware did not exist.

The marketplace required all of that to exist first.

Nobody had built it.

You cannot trade what you cannot provision, measure, route, and trust operationally.

The mistake was not in seeing abundance. The abundance was real. The mistake was in financializing it before the operational middleware was in place.

Bandwidth had its turn in 2000. Compute capacity is having its turn now.

The temptation is older than the asset class. So is the failure.

If someone tells you the opportunity is a GPU marketplace, the first question is not whether GPUs are scarce or abundant. The first question is whether the buyer can trust the capacity, verify the performance, move the workload, understand the security boundary, reconcile the bill, finance the collateral, and get paid when something breaks.

The exchange is the last mile.

The workflow comes first.

Something a little different

The three layers

The telecom cycle did not produce one software market. It produced three.

The first layer sold to the asset operator.

Granite, Cramer, NetCracker, MetaSolv, and parts of the telecom expense management category made the broken operating layer legible to carriers and telecom teams. These companies helped answer basic questions: What do we own? Which physical asset maps to which customer? Which service has been provisioned? Which circuit is still being billed? Which database is wrong?

The exits were real, but the ceiling was set by the buyer base. Carriers were consolidated, indebted, slow to buy, and often still digesting the same wreckage the software was built to manage. Granite sold for roughly $70 million. NetCracker for around $300 million. Cramer for $375 million. MetaSolv for $219 million. Tangoe went private for $242.6 million.

The software mattered. The buyer base was narrow.

The second layer sold to the enterprise that depended on the asset.

Akamai, Prolexic, Dyn, ThousandEyes, and similar companies did not ask enterprises to operate the network. They helped them survive dependence on it. Your application needed to be fast. Your site needed to stay online. Your traffic needed to route correctly. Your security team needed to understand what was happening outside your own walls.

The buyer was no longer the carrier trying to clean up its own plant. The buyer was every enterprise whose revenue, uptime, or customer experience now depended on the network.

The ceiling moved.

Prolexic sold to Akamai for $370 million. Dyn sold to Oracle for more than $600 million. ThousandEyes sold to Cisco for roughly $1 billion. Akamai became the public-company version of the same motion.

The third layer sold to the person who never wanted to think about the infrastructure at all.

Twilio sold to the developer. Apptio sold to the CIO and CFO. CloudHealth sold to the team trying to understand cloud spend. Acme Packet sold into the control plane that made communications infrastructure behave like something software could depend on.

The buyer was no longer “telecom.” The buyer was everyone building on top.

This layer produced the largest outcomes: Acme Packet at $2.1 billion, Apptio at $4.6 billion, Twilio as the public-company version with a peak market cap north of $60 billion.

That is the distinction that matters.

The exit scaled with the breadth of the buyer base, not with the founder’s proximity to the physical asset.

It is easy to tell the story backward and make it sound inevitable. Of course carrier software sold to carriers. Of course enterprise network visibility sold to enterprises. Of course communications APIs sold to developers. But none of this felt obvious from inside the cycle. In 2002, everyone was still staring at bankrupt fiber. In 2004, the interesting company looked like a boring inventory system. In 2008, the interesting company looked like a developer tool for making prank phone calls.

The sequence was not accidental.

The carrier inventory layer had to be partially fixed before enterprises could depend on the network with confidence. The enterprise layer had to mature before developers and CFOs could treat the network as an abstraction. You could not sell the API until the interconnect worked. You could not sell the spend dashboard until the usage existed. You could not make the infrastructure disappear until enough of the work underneath had become stable.

The work moved upward because the layer underneath had become stable enough to vanish.

What is forming now

The AI buildout is producing the same kind of work, but not in the same places.

The lazy version is to map telecom terms directly onto AI terms. Fiber becomes GPUs. Bandwidth becomes tokens. Carriers become neoclouds. Bandwidth exchanges become compute exchanges. Twilio becomes “the Twilio for AI.” This is how bad analogies reproduce. They keep the nouns and miss the mechanism.

The mechanism is not “fiber, but AI.”

The mechanism is: capital commits before the operating layer exists, and then the world has to invent the operating layer.

In AI, the operating layer is forming across several balance sheets at once.

It is forming in power procurement, where utilities are being asked to plan around loads that are large, concentrated, and politically sensitive before end-user AI demand is fully proven.

It is forming in data center finance, where leases, special-purpose vehicles, private credit, and hyperscaler commitments are being used to transform uncertain future compute demand into today’s investment-grade paper.

It is forming in GPU-backed lending, where chips become collateral, collateral values depend on utilization and obsolescence, and lenders need to know whether a three-year-old H100 is money-good in a world where the next generation of silicon has changed the economics.

It is forming in tax abatements, where state and local governments are discovering that the fiscal note on a data center incentive can be off by orders of magnitude once the category becomes real.

It is forming in inference billing, where developers and finance teams are trying to understand a cost structure that behaves nothing like SaaS seat pricing and nothing like cloud storage.

It is forming in security, where model access, data leakage, prompt injection, inference abuse, and usage monitoring are becoming line items in enterprise risk registers.

It is forming in procurement, where the enterprise wants the promise of AI but does not want to underwrite a vendor whose gross margin depends on subsidized compute, related-party cloud commitments, or a model provider’s ability to raise the next round.

Most of this will not become startups.

The first version of every new operating layer is people. People in spreadsheets. People in law firms. People in procurement. People in utility regulatory affairs. People in insurance. People in credit. People in finance teams explaining why the bill moved. People who know the one clause in the one contract that everyone else forgot mattered.

Then, sometimes, it becomes software.

The trick is not to ask, “What is the AI version of Twilio?”

The trick is to ask, “Which workflow is painful enough that the person doing it has become the system of record?”

That was true in telecom. It will be true here.

The AI-inventory company may not call itself inventory. It may start as a way to reconcile GPU commitments against actual available capacity, power constraints, contractual priority, security posture, and depreciation assumptions.

The AI-expense company may not call itself FinOps. It may start as a tool for one engineering team trying to understand why inference costs doubled after a model routing change.

The AI-credit company may not call itself lending infrastructure. It may start as a collateral-monitoring workflow for one lender who realizes that GPU marks, utilization, contract tenor, and hardware generation are all moving at different speeds.

The AI-power company may not call itself energy software. It may start as a queue, permit, turbine-slot, interconnection, and load-forecast reconciliation tool for developers who are trying to figure out which announced campuses are real.

The AI-governance company may not call itself governance. It may start as a way for a CFO to answer a very simple question: who is allowed to spend how much money on which model, against which data, for which business purpose?

The categories will sound obvious later.

They do not sound obvious now because the people doing the work still think of it as work.

What speed changes

There is one important difference between telecom and AI.

In telecom, the work beneath the boom became obvious after the bust. The debris was sitting there. The bankrupt networks had already changed hands. The invoices were wrong. The inventory systems were broken. The interconnection points mattered. A founder could spend three years inside a carrier’s billing system after the crash, leave, and still be early.

The debris was not going anywhere.

This time, the work is forming during the buildout.

The Beignet template appeared in October 2024. Virginia adopted a new electric rate class for large power users in November 2025. GPU-backed lending is already a more-than-$10 billion market with collateral values, covenant tracking, and remarketing workflows still held together by spreadsheets. Behind-the-meter power projects are already being sorted into bankable, option-value, and slideware based on permits, offtake, turbines, water, gas supply, and regulatory exposure.

These are not post-crash artifacts.

They are live workflows hardening in real time.

That changes the identification window.

When the work forms after the bust, the market has time to notice the mess. When it forms during the buildout, the window between “someone should fix this” and “someone is already fixing this” compresses.

That is the only claim worth making. Not that the layer sequence will compress. Not that the exits will look the same. Not that AI has to rhyme cleanly with telecom because history enjoys being useful to investors. It usually does not.

Only this: the first software company built for the work beneath the AI buildout to reach a billion dollars in enterprise value is probably already founded.

The founders are not telling the venture press. They are not saying they are building the software layer of the AI bubble. They are not announcing a thesis about capex cycles, salvage value, or historical analogies.

They think they are fixing a workflow.

The window to find them is shorter than last time.

Were you building or investing in the telecom boom while I was listening to Marketplace in the back seat of my dad’s car? Would love to hear your thoughts!

As AI-assisted+ writing is becoming way more common, people are getting more skeptical about what is actually human and what just sounds human. Perfect grammar, clean structure, polished transitions (stuff we used to associate with competence and professionalism) is now a little suspect.

Which has created a strange new problem: people are not just trying to avoid sounding like AI. They are trying to prove they are human.

You can see it everywhere. The Wall Street Journal just wrote about it. Writers adding typos on purpose. Professionals making their emails more aggressively casual. People dropping in extra exclamation points, lowercase openings, fragments, slang, anything that might make the writing feel less generated.

There are even tools now whose whole job is to make AI writing look worse. Sinceerly, an “anti-Grammarly” built with Claude, adds typos and clumsy phrasing back into emails. WriteHuman promises a similar kind of ‘humanity’ at scale.

The pitch is simple: perfect writing reads like AI now, so rough it up. Add a typo. Kill the em dash. Say “hey yo.” Make it messy enough, and maybe no one will notice.

The whole thing has “lol how do u do, fellow humans?” energy.

But the strangest part is not that AI is trying to sound human. It is that humans are trying to sound “human” too.

What a strange shift, right? I guess it makes sense that once polish became suspicious, imperfection became a performance. But think about the consequence: were all suddenly reasoning over the signs of personhood.

Is this too clean? Too balanced? Too formal? Would a real person use this many em dashes? Should I leave the typo in so they know I wrote it? Now I’m self-conscious about the typo in the last paragraph. Am I performing too much?

The problem is not that imperfection does not matter. It does. A typo can mean a lot. It can mean someone trusts you enough to let their guard down. It can mean they wanted to get back to you quickly because they cared more about responding than cleaning up the sentence. It can mean they were tired, distracted, comfortable, rushed.

The typo was never the signal by itself. The situation was.

That is what these attempts at “human” writing miss. They treat imperfection like a texture you can apply after the fact. But real imperfection is not random. It is not decorative. It is the effortlessness that comes from knowing the context, having confidence in what you are saying, and caring the right amount about the person reading it.

Perfect is easy to replicate. It’s weird human imperfections that are hard.

Not because typos are hard to add. Anyone can add a typo. See above. Imperfect is hard because the kind of imperfection that actually reads as human is not something you decide by committee with yourself. It is not, “Should I misspell this word to seem more authentic?” The moment you are reasoning that hard over the signal, the signal is already dead.

Real imperfection happens when you are writing from somewhere.

When generative AI first became mainstream, I wanted to see if it could learn my voice. I built a project in Claude with a detailed voice guide and more than thirty samples of my writing across different formats: emails, texts, LinkedIn posts, longer essays, podcast transcripts. I gave it my professional writing and my casual writing. I tried to teach it my cadence, sentence length, rhythm, humor, preferences, and the phrases I would never use.

The results were bad in a very specific way.

They were not always incoherent. They were not even always obviously wrong. Sometimes they had the surface markers. A sentence would sound vaguely like me. A paragraph would almost have the right rhythm. But the drafts were still off. Too polished in places where I would have been more direct. Too casual in places where I would have been more careful. Too self-aware in places where I would have just said the thing.

It could imitate pieces of my style. It did not know when any of those pieces belonged.

My voice is not a formula. It is not just short sentences, long sentences, conversational tone, occasional self-deprecation, and a few punctuation habits. Those are artifacts. They are the visible residue of something underneath: context, confidence, and care.

The way I write on LinkedIn is different from how I write on Substack. The way I text friends is different from how I email LPs or founders. The way I write when I am excited is different from how I write when I am trying to be precise. The way I soften a point with humor depends on who is reading it. The way I choose not to soften a point also depends on who is reading it.

That is the voice. It’s inconsistent. Imperfect. In need of editing (or maybe not now?). But that’s me.

Voice is not one static style you drag into every room. It changes because the room changes. The audience changes. The stakes change. Your relationship to the person reading changes. AI can copy the artifact, but it does not automatically understand the social logic underneath it.

That is why deliberately “humanized” writing feels so uncanny. It understands the costume of ease but not the conditions that create ease.

A real typo might mean intimacy. Or urgency. Or indifference. A fake typo usually means only one thing: someone wanted the appearance of intimacy, urgency, or indifference without the situation that would naturally produce it. It copies the evidence while missing the cause.

The same is true for casual language. A “hey yo” can work if that is actually how you talk to that person. It can be warm, funny, disarming, or completely natural. But when it is added as a proof-of-humanity tactic, it does not feel effortless. It feels like anxious performance. Anxious that competence now looks synthetic. So instead of asking whether the writing is clear, true, useful, or alive, people start asking whether it has enough human-looking flaws.

But human writing is not imperfect because humans make mistakes. Human writing is imperfect because humans are writing inside actual circumstances. We are responding to someone. We are trying to be understood. We are managing closeness, authority, warmth, urgency, boredom, humor, risk. We are deciding, often without realizing we are deciding, how much polish the moment deserves.

That is the part you develop through writing. That instinct comes from repetition, reading, editing, overexplaining, being misunderstood, cringing at old work, trying again, and slowly building taste. It comes from writing in different contexts long enough to know that “good writing” is not one thing. Sometimes good writing is clean. Sometimes it is messy. Sometimes it is funny. Sometimes it is blunt. Sometimes the most human thing you can do is send the polished version. Sometimes the most human thing you can do is not.

That is why writing is still thinking.

Not because pen and paper are holy. Not because everyone needs to suffer through every sentence manually or the scourge that is cursive. AI can be useful. It can help organize thoughts, summarize, outline, pressure-test structure, and get you unstuck. But the part worth protecting is not the typing. It is the deciding.

Writing forces you to figure out what you mean, who you are speaking to, and what the moment requires from you. It reveals whether you have something to say or just a tone you are trying to perform.

Perfect is easy now. Imperfect is hard human.

The Builders Series profiles Buyer-Builders in the wild. Start with the thesis.

The Builders Series #001: Benchling

Every well-capitalized team that tried to disrupt home services looked at the same market and found the same customer.

The homeowner.

They saw consumers who could not find a good plumber, did not know which electrician to trust, and had no easy way to compare quotes. So they built search, discovery, lead-gen, booking, reviews, and marketplaces. The problem was obvious because almost everyone had experienced it. Everyone has needed someone to fix a leak, replace an AC unit, clean a house, or show up when something broke.

ServiceTitan looked at the same market and built for the person on the other side of the door.

Not the homeowner looking for help.

The contractor trying to run the business after the job was done.

That choice was made before the first line of code, before the first investor meeting, before there was a company at all. Ara Mahdessian and Vahe Kuzoyan were not running a competitive analysis of home-services software. They were building something for their fathers.

The customer was never in question because the problem was never abstract.

They grew up watching it.

The Market Was Not the Customer

Home services was never small enough to miss.

In its S-1, ServiceTitan estimated that end customers in the U.S. and Canada spend roughly $1.5 trillion annually on trades services, with approximately $650 billion inside the trades and end markets ServiceTitan already served at the time of filing.

This was not some hidden niche waiting to be discovered by a clever founder. It was plumbing, HVAC, electrical, roofing, pest control, garage doors, landscaping, and the rest of the work that keeps homes and commercial buildings usable.

The market was visible.

The right customer was not.

Most companies that attacked the category in the 2010s interpreted the pain from the consumer side. Angi promised contractors access to the “largest homeowner network” and more than 1.5 million monthly opportunities. Handy made it easier for consumers to book home cleaning and handyman services. Porch built around the homeowner lifecycle: buying, moving, maintaining, and improving a home.

They were not wrong that homeowners had a problem.

They were wrong about which problem could become the system of record.

For a homeowner, the drama begins when the sink backs up or the AC dies. For a contractor, the drama began hours earlier and kept going long after the truck pulled away. The phone had to be answered. The call had to be booked. The dispatcher had to know which technician was available, what parts were on the truck, whether the last job was running late, whether the customer was a member, whether financing made sense, whether the quote was profitable, whether the invoice got paid, and whether the owner could make payroll on Friday.

The marketplace version of home services saw the contractor as supply.

The contractor’s son saw the contractor as the customer.

The Business After the Job

Ara and Vahe did not enter home services through a market map.

They entered it through their fathers.

Both were sons of Armenian immigrants in Southern California. Ara’s father ran a residential contracting business. Vahe’s father was a plumber. ServiceTitan’s own company story is direct about the origin: Ara and Vahe originally built ServiceTitan to help their fathers run better contracting businesses.

That detail matters because it changes the shape of the company before the company exists.

The default Silicon Valley interpretation of home services was consumer frustration. Ara and Vahe’s default interpretation was contractor exhaustion.

A marketplace asks:

How do we help the homeowner find someone?

ServiceTitan asked:

Why is the person doing the work still running the business on paper?

That was not positioning. It was inherited context.

Ara and Vahe met in 2004 through Armenian student associations, began working on software for their family businesses, formally founded ServiceTitan in 2007, and launched the cloud platform in 2012. The early version was not a grand enterprise transformation platform. It was a toolkit for people they knew.

Vahe later described the original motivation simply: he wanted to be useful to his dad. Ara described the early plan as a summer project they could build, hand to their parents, and leave behind before getting jobs at Google or Facebook.

That is not what happened.

Other contractors heard about it. Word spread through the local trades community. The first demand signal did not come from paid acquisition or a polished sales motion. It came from contractors telling other contractors that someone had finally built software that understood how their businesses actually worked.

That is the buyer-builder thesis in its most compressed form:

Insider knowledge does not just determine how you build.

It determines who you build for.

And who you build for determines everything else.

The Search Bar Was a Decoy

The outsider mistake was understandable.

Home services is one of the few categories where almost everyone has felt the consumer pain personally. Everyone has needed a contractor. Everyone has wondered whether a quote was fair. Everyone has heard a story about someone who never showed up.

But personal experience with one side of a transaction can create false intimacy with the whole market.

From the homeowner’s side, the problem looks like discovery.

From the contractor’s side, discovery is just one variable inside a much larger operating machine.

Leads are not useful if the office cannot answer the call, book the appointment, dispatch the right technician, generate a quote, collect payment, and follow up. Demand is not the same as capacity. A booked job is not the same as a profitable job.

This is where the marketplace model quietly loses the plot.

A lead-generation business can help a contractor get more demand. But a contractor drowning in administrative work may not need more demand. He may need a way to stop losing margin in the handoffs.

ServiceTitan built for the handoffs.

By the time it filed to go public, ServiceTitan described its platform as an end-to-end operating system across advertising, scheduling, dispatching, estimates, invoices, payment processing, CRM, field service management, ERP, human capital management, and fintech.

That product surface was not random bloat.

It was the shape of a trades business when viewed from the inside.

Bessemer’s Series A memo described ServiceTitan as sitting “at the core” of daily operations, from inbound call management to mobile dispatching and invoicing to back-office management.

That is the difference between a marketplace and an operating system.

One sends you work.

The other changes how work moves through the company.

The First Moat Was Customer Selection

ServiceTitan did not discover contractors after trying homeowners first.

There was no clever pivot from consumer marketplace to vertical SaaS. The customer was selected before the business model.

That choice shaped everything.

Selling to contractors meant ServiceTitan could become mission-critical. It could touch revenue, scheduling, technician productivity, payments, accounting, customer reviews, marketing attribution, and management visibility. It could grow as the customer grew. It could become harder to rip out with every workflow it absorbed.

Selling to homeowners would have meant fighting for traffic, brand, SEO, conversion, and take rate.

Selling to contractors meant owning the place where revenue actually got created.

By early 2015, before ServiceTitan had become a venture-scale brand, the signs were already there. Bessemer’s memo said the company had grown from roughly $1 million ARR to more than $3 million ARR, landed a $4 million ACV franchise deal, done it with around 20 employees, spent essentially nothing on sales and marketing, and remained cash-flow positive.

Those are not the metrics of a company persuading the market into existence.

They are the metrics of a company the market is pulling out of the founders.

The same memo emphasized that most new customers were coming through referrals and word of mouth. Contractors were not being convinced that ServiceTitan understood them. They were recognizing themselves in the product.

The strongest version of vertical software often starts this way. The wedge does not look like strategy from the outside. It looks too specific, too unglamorous, too close to the ground.

But the specificity is the point.

Ara and Vahe were not trying to “own home services.” They were trying to make their fathers’ businesses easier to run. In doing so, they picked the customer everyone else had reduced to supply.

The Marketplace Tax

The consumer marketplaces created real value.

They aggregated demand, made service providers more discoverable, and gave homeowners more ways to compare options. But their structural incentive was different from ServiceTitan’s.

A marketplace wants more transactions.

A contractor operating system wants healthier contractors.

That difference sounds subtle until you follow the economics.

If a platform monetizes leads, the contractor is often paying for a chance to win work. If a platform runs the contractor’s business, the contractor pays because the system improves booking, dispatch, close rates, collections, reporting, and customer retention.

The first model taxes demand.

The second model expands capacity.

This is why the wrong customer can lead to the wrong product even in the right market.

Homejoy is the extreme version of the marketplace trap. It attacked home cleaning from the consumer side and shut down in 2015 after struggling to raise capital while facing worker-classification lawsuits. The company had already faced growth and revenue challenges, but its CEO said the “deciding factor” was the lawsuits over whether workers should be classified as employees or contractors.

ServiceTitan avoided that entire axis.

It did not need to intermediate labor. It did not need to own consumer demand. It did not need to decide whether the technician was an employee or contractor.

It sold software to the business already responsible for the work.

That was the cleaner entry point and the deeper one.

The outside-in company tries to abstract away the messy operator.

The inside-out company makes the operator more powerful.

The Product Could Only Expand One Way

Once ServiceTitan owned the operating layer, expansion was not a separate act of imagination.

It was the natural consequence of following the workflow.

A job does not stop at dispatch. It turns into a quote, then an invoice, then a payment, then a review, then maybe a membership, then maybe a financing opportunity, then maybe another job.

• The technician needs mobile tools.

• The CSR needs call tracking.

• The owner needs margin visibility.

• The office needs accounting.

• The larger customer needs multi-location controls.

• The consolidator needs standardized processes across acquired shops.

ServiceTitan’s S-1 makes the product logic explicit: CRM, field service management, ERP, HCM, and fintech are the “core centers of gravity” inside a trades business.

The company’s expansion into payments, financing, inventory, accounting, payroll integrations, and AI products followed the actual path of work rather than a generic SaaS land-and-expand diagram.

This is why the numbers began to look less like a niche SMB tool and more like infrastructure.

In the 12 months ended July 31, 2024, ServiceTitan processed $62 billion in gross transaction volume through the platform. In fiscal 2024, its customers completed approximately 109 million jobs through ServiceTitan, and the company had about 8,000 active customers. Its gross dollar retention had been above 95% and net dollar retention above 110% for each of the prior ten fiscal quarters.

By fiscal 2026, the company reported $82.1 billion in GTV, $961 million in revenue, about 10,800 active customers, gross dollar retention above 95%, and net dollar retention above 110%.

Those are not the metrics of a product that helps contractors find jobs.

Those are the metrics of a product that contractors run jobs through.

The Investor Lesson Was Not Just “Vertical SaaS”

It is tempting to reduce ServiceTitan to the obvious category lesson:

Vertical SaaS can produce enormous outcomes in overlooked industries.

That is true, but incomplete.

The deeper lesson is that vertical SaaS only works when the company chooses the right vertical truth.

“Home services” was too broad to be the insight.

“Homeowners need contractors” was too shallow to be the insight.

“Trades businesses are operationally complex, under-softwared revenue machines run by people who have been ignored by modern SaaS” was the insight.

Bessemer got there early from the product and traction. Its memo emphasized that ServiceTitan had strong domain knowledge, that the founders came from plumbing and residential contracting families, and that they had started because they could not find good software for Vahe’s father’s business.

Battery arrived through a broader labor-market lens. Its “blue-collar software” research started with the observation that a large share of the workforce did not sit behind desks and had not received the same software upgrades office workers had. That research led Battery to ServiceTitan.

Different frameworks.

Same company.

The reason is that ServiceTitan sat at the intersection of both truths: a huge underserved labor market and a founding team with lived access to the customer’s operating pain.

The customer selection came first.

The category followed.

The IPO Was the Receipt

ServiceTitan went public on Nasdaq under the ticker TTAN in December 2024.

The company priced its IPO at $71 per share, above the marketed range, raised approximately $624.8 million, and was valued at roughly $6.3 billion at the offering price.

On its first day of trading, the stock rose roughly 42%, with a market capitalization approaching $9 billion around the first-day trading price.

The public-market story was about scale: revenue, GTV, retention, market size, AI, payments, and expansion into more trades.

But the company’s origin still explains the outcome better than the numbers do.

ServiceTitan did not win because it noticed that home services was large.

Plenty of people noticed that.

It won because it understood where the software belonged.

Not in front of the homeowner.

Inside the contractor’s day.

Why We Like This

The software opportunity in overlooked markets is often misdescribed as a digitization problem.

Paper needs to become software.

Calls need to become workflows.

Invoices need to become payments.

Spreadsheets need to become dashboards.

That framing is directionally right and strategically thin.

The real question is:

Digitization for whom?

In home services, the consumer-facing answer produced marketplaces. The contractor-facing answer produced ServiceTitan.

Both looked at the same market.

Only one became the operating system.

This is why buyer-builder fit matters. It is not just a founder credential. It is a market selection mechanism. Outsiders can study a market and still choose the wrong customer because the most visible pain is not always the most valuable pain.

Insiders often begin with the right customer because the pain was never theoretical.

They watched it happen at the kitchen table.

Ara and Vahe did not need to be told that contractors were underserved. They had seen the second shift. They had seen the paper invoices. They had seen the missed dinners. They had seen competent operators lose hours to work that software should have absorbed years earlier.

The insight was not that contractors needed more leads.

The insight was that the contractor was the platform.

That is the pattern we care about in regulated, fragmented, operationally dense markets. The best founders are not always the ones with the cleanest market map. They are the ones who know which user has been forced to carry the system on their back.

In ServiceTitan’s case, that user was not the homeowner searching for help. It was the contractor, after the job was done, still awake with the invoice.

tl;dr: AI productivity is real; AI margin expansion is still unproven.

fast framework: Don’t ask whether AI replaces labor. Ask what input stack produces the most output per dollar.

open question: What if the long-term savings never arrive?

Rightsizing. Streamlining. Flattening. Refocusing. Increasing operational efficiency.

The corporate layoff memo has always been an act of linguistic jujitsu: minimize that you’re firing people and maximize the impact of not paying those people anymore.

AI has given the memo a new vocabulary.

We’ve seen a rash of these over the past few months, but there were two that leapt out at me because of the stark contrast between them.

Coinbase & Cloudflare.

Very different companies, very different financial profiles, same family of claim: intelligence tools have changed what it means to run a company. Fewer people, more software, more leverage.

The temptation is to read these as the same story. Or at least as part of the same larger story about AI. I don’t think they are.

Coinbase looks like the old layoff memo in new clothes: a cost correction the company already needed, narrated in the language of AI productivity.

Cloudflare is harder. Cloudflare is making a more ambitious claim: that AI changes the input mix of the company itself. Fewer people in some roles. More compute. More tooling. More infrastructure. More leverage from the people who remain. The abstract AI debate asks whether the technology is good or bad, liberating or destructive, acceleration or doom. The corporate version is more concrete and more useful: does AI actually let a company produce more output per dollar of input?

Not per employee. Per dollar of input.

Because labor does not simply disappear. It gets replaced by something. Tokens. GPUs. Inference. Internal tools. Data pipelines. Evaluation layers. Monitoring. Security review. Vendor lock-in. And the expensive humans still required to make all of it work.

That is the AI margin paradox.

AI may make workers more productive. It probably already does. But productivity is not the same thing as margin expansion. A company that cuts 1,000 employees and spends the savings on compute, chips, tooling, infrastructure, and equity packages for the remaining engineers has not eliminated cost; rather, it has simply changed its shape.

For now at least, it appears the arrival of a free lunch has been greatly exaggerated.

Jack the narrative

In February, Jack Dorsey announced that Block would cut more than 4,000 jobs, nearly half the company, as part of an AI-driven overhaul. Dorsey’s line was the kind of thing public markets were ready to hear:

“Intelligence tools have changed what it means to build and run a company.”

That is a beautiful sentence if you are holding a bloated cost structure and need to make it sound like a strategy. In the past, we might have called this cost control. In the language of the new world, this is AI-driven innovation.

It moves the cut from the past tense to the future tense. The company is not correcting something. It is adapting to something.

Do not let a great media narrative go to waste, right?

Want to read 6,000 words on AI doomerism? Yeah you do.

Coinbase: the admission

Coinbase is the cleaner version of that story.

In May, Coinbase said it would cut about 700 jobs, roughly 14% of its workforce, as it trimmed costs amid crypto-market volatility and repositioned for the AI era. The restructuring would cost $50 million to $60 million, and Brian Armstrong cited AI tools allowing smaller teams to ship code and automate work.

That sounds like the future.

The filings sound like the past.

Coinbase had already lived through the first version of this cycle. It hired into the 2021 crypto boom, cut during the 2022–2023 downturn, then re-expanded. There was no AI productivity restructuring hiding in the 10-K. There was no “we have discovered a new operating model.”

There was hiring.

The income statement had already made the next move obvious. Revenue grew, but operating expenses grew faster. The company had restructured, told the market it had learned discipline, and then quietly rebuilt the cost base. The AI letter gave it a better way to reverse itself.

This is cost correction.

And it is not necessarily bad. Sometimes the company really did need to confess. Sometimes the right move is to cut the thing you should not have built back up.

The market can understand that trade because it is not really about AI. It is about taking costs out of a volatile, transaction-sensitive business and giving investors a cleaner path to profitability.

The AI framing matters because it makes the cut feel offensive rather than defensive.

Without AI, Coinbase is cutting because crypto volumes are volatile and management over-hired again. With AI, Coinbase is becoming lean, fast, and AI-native.

Same action. Better tense.

Cloudflare: the investment

Cloudflare is harder to understand, and I think that is the point.

In May, Cloudflare said it would cut about 20% of its workforce, more than 1,100 employees, after internal AI use increased more than sixfold in three months. The memo was not written like a cost-cutting memo. It was written like an operating-system migration.

Cloudflare said the cuts were not about individual performance or short-term cost pressure. The company said it was reimagining internal processes, teams, and roles for the agentic AI era. Matthew Prince later clarified that very few engineers or customer-facing salespeople were affected, and that Cloudflare would continue hiring in those roles.

That makes Cloudflare different from Coinbase.

Coinbase is shrinking the human-cost bucket. Cloudflare is changing the mix of inputs. It is not just saying fewer people. It is saying fewer of some kinds of people, more AI tooling, more infrastructure, more internal agent usage, more product surface around inference, and more operating leverage through a different company architecture.

Cloudflare’s five-year picture looks nothing like Coinbase’s. Revenue compounded more than 3x from 2021 to 2025. Headcount rose every year and more than doubled over the same period. Cash from operations improved meaningfully. Revenue per employee rose. Cloudflare looked like a company expanding the market, not one cleaning up a prior hiring mistake.

And yet 2025 exposed the tension.

Operating margin had improved for years, then stalled. Gross margin compressed. Capex jumped. SBC remained large. Cloudflare’s AI references in the 2025 10-K were about product and network capacity, not internal headcount transformation. Then, weeks later, the company described a 20% cut as part of an agentic AI-first operating model.

The market responded with a 20% cut to its market cap.

A Startup Claim With a Public P&L

The Cloudflare selloff is not necessarily a verdict on Cloudflare’s strategy.

It is what happens when a startup claim shows up inside a public company.

Early-stage founders make this kind of claim every week.

We do not need ten implementation managers because agents handle onboarding. We do not need a large support team because the product resolves tickets itself. We do not need a services-heavy deployment model because AI turns messy customer inputs into structured workflows. We do not need the old headcount shape because the company is built around a different operating system.

Maybe.

But that claim is not proven by saying “AI.” It is proven only when the product ships faster, customers get served better, gross margins hold, and the compute bill does not eat the savings.

That is why Cloudflare is interesting. It is not just a public-company layoff story. It is an AI-startup operating claim appearing on a public-company income statement.

And that’s why as an early-stage vertical AI investor, I’m paying attention.

Every AI-native startup is making some version of Cloudflare’s claim. Cloudflare is just large enough that the claim has to touch the P&L.

AI vs. Margins

The AI margin story is usually told as if costs associated with labor simply disappears. Perhaps there are very short term dislocations, but the financials show how costs shift. The post-AI cost base has three buckets:

AI does not simply reduce the first bucket. It expands the other two. Sometimes that is a great trade. Sometimes it is a terrible trade. Most of the time, right now, nobody knows.

A software company can announce that AI tools let one engineer do the work of five. Great. What does that mean for Codex usage? Claude Code? Internal tooling? Model-router spend? Fine-tuning? Inference? Evaluation? Security review? Monitoring? Data pipelines? Prompt caching? The GPU bill? The vendor lock-in? The human review layer that appears after the agent fails in production?

Payroll is boring, which is another way of saying it is predictable. Compute is not.

A salary is a number you know in November when you set the budget. Compute is a monthly surprise party hosted by your most enthusiastic users, most ambitious engineers, and whichever lab dropped the “most capable model ever” that billing cycle.

This is where the public market data is starting to get interesting.

The Information’s recent survey of public tech earnings calls showed the spread. Spotify said headcount was down slightly, but compute per employee was rising, including spend on Claude Code and Codex. Shopify’s AI assistant Sidekick was popular with merchants, but merchant usage raised LLM costs enough to partly offset customer-support savings.

When AI works, people use it more.

When people use it more, the bill goes up.

The thing replacing support cost also creates inference cost. The thing helping engineers ship faster also creates tool spend. The thing letting the company slow headcount growth also creates an entirely new class of variable expense.

The labor savings are recognized this quarter. The compute bill arrives later.

Why buyer-builders matter

Every AI deal at pre-seed is making a small Cloudflare-shaped claim: the company can use AI to serve customers with a different mix of humans and machines.

The risk is that the founder spends compute on a workflow that does not actually matter.

That is where the buyer-builder thesis becomes relevant.

We think the future of software will be written by former customers. Read the thesis here:

Not because buyer-builders magically solve the compute-cost problem. They don’t.

They solve a different problem: they reduce the odds that you spend compute on the wrong workflow.

A generic AI founder starts with the tool. A buyer-builder starts with the broken process.

That difference matters more now, not less.

If compute is cheap and predictable, you can afford to spray automation at vague pain. But if compute is variable, usage-sensitive, and capable of eating the margin you thought you were creating, then workflow precision becomes more valuable.

You need to know exactly where the work breaks, exactly who pays for the fix, and exactly how much human cost, delay, error, or risk the software is replacing.

The buyer-builder has already lived that.

They know which task is annoying but worthless, and which task is ugly but budgeted. They know which workflow looks small from the outside but controls the whole day. They know where the customer will tolerate imperfection and where the customer will not. They know which “AI automation” will actually save money and which one will just create a new review queue.

That is the private-market version of the Cloudflare question.

Cloudflare is asking investors to believe that a new mix of humans, agents, and infrastructure will make the company more efficient. AI-native startups are asking customers to believe the same thing, one workflow at a time.

The safest version of that claim does not start with “AI lets us use fewer people.”

It starts with:

I know this workflow well enough to know where AI is worth paying for.

That is why buyer-builders matter. The context they bring through direct industry knowledge makes the value of AI-augmented workflows more legible.

That is better for investors, sure.

It is also much better for acquiring customers.

Public companies are large enough to show the trade. Startups make the same trade before the numbers are visible.

At pre-seed, the AI margin story is usually still a sentence in a deck. The founder says the company will need fewer people because agents will do the work. Maybe that is true. But somewhere underneath that sentence is an input mix: model calls, evals, monitoring, review, edge cases, and the cost of making the system reliable enough to charge for.

The open question is which one compounds faster: AI productivity or AI cost.

That is not a public-market question or a private-market question.

It is the AI company-building question.

These are the businesses that run on materials, labor, scheduling, invoices, dispatch, compliance, and physical work. They have been promised a software revolution every five years and have rarely received one that survived contact with their day.

We think the next version is finally arriving, and that the right way to invest in it is meaningfully different from what venture capital tried to do in these markets last decade.

Our view here comes up in LP conversations so much that we decided to share how we think about the category.

The marketplace mistake

The 2010s playbook for these industries was the demand-side marketplace. Companies aggregated fragmented supply, connected it to consumers, and took a percentage of the transaction. Some of those companies worked, and most of the ones we care about did not.

The structural problem has two parts. The first is disintermediation. The value created by introducing a homeowner to a roofer, or a diner to a restaurant, or a shipper to a carrier, dissipates almost immediately after the introduction. Your roof is not going anywhere, your favorite restaurant is around the corner, and your trucking carrier already has your number. The marketplace eats the acquisition cost on the first transaction and rarely gets a second crack at the customer. The take rate gets crushed by direct repeat business that the platform paid to manufacture.

The second is geography. These are physically constrained markets. You cannot import a roofer from another town, and you cannot deliver a meal from across the country. The marketplace cannot use scale to compress unit economics, because the economics are determined by the local supply and the local demand, both of which the platform has to keep buying.

Marketplace vs. Operations

There is also a funnel problem inside many of these businesses. Residential services contractors, for instance, spend somewhere around 10 percent of revenue acquiring customers and present the estimate at the kitchen table only after the entire acquisition cost has already been paid, which is why the homeowner experience can feel like a gentle hostage negotiation. We are not interested in fixing that funnel. The contractor’s funnel is the contractor’s problem. The opportunity for us as investors is not in front of the funnel. It is underneath it.

An ops-focused software wedge.

What we have actually seen work in these categories is a different shape. The company starts as software, captures the operator’s daily workflow, and then expands into the financial layer because the operator’s binding constraint is almost always economic rather than informational. Toast generated approximately $5 billion of fintech revenue in its most recent fiscal year against $936 million of subscription revenue. Shopify reported 76 percent of its revenue from Merchant Solutions, with $248 billion of GMV flowing through Shopify Payments. ServiceTitan reported approximately $82 billion of gross transaction volume across roughly 10,800 active customers, and integrated Affirm into the technician workflow last year so financing happens at the moment of estimate.

We look for a few specific things inside that pattern.

The first is buyer-builders. The founders most likely to identify the right workflow are the ones who already lived inside it, and whose first ten customers are reachable through a phone tree rather than a paid funnel. We would rather back a founder who spent a decade inside the industry being served than one who spent a decade building software somewhere else and is now shopping for a vertical to apply it to.

The second is what we call single-player utility. The product has to be useful to the first customer the moment they turn it on, with no network effect required. If the company has a cold-start problem, the company has the wrong product. The right shape is software that solves a specific operator’s pain on day one and incidentally creates the data layer that lets later products attach to the same surface.

The third is a path to a B2B marketplace through the supplier side rather than the consumer side. Suppliers in these categories are not geographically locked the way contractors are. A regional aluminum supplier sells across multiple states, and a restaurant equipment distributor serves dozens of chains. Going to the supplier means going to the participant in the chain with budget, with pain, and without the geographic lock that kills the demand-side version.

Read our investment thesis

Why now

We have been hearing some version of “AI changes vertical software” for two years, and most of the framings collapse on contact with the actual ground reality. These are the primary drivers we’ve identified.

AI is genuinely good at the repetitive, awkwardly shaped tasks that have historically been too messy to digitize at a reasonable cost. AI can reconcile six thousand SKUs across paper catalogs and emailed PDFs without a custom integration. It can extract line items from a hand-drawn site sketch that a contractor walked in with. It can read a supplier’s tier-pricing table directly into an estimate without anyone retyping anything in between. None of this is glamorous AI, and all of it used to require a custom build on every customer.

Voice and vision are now usable input modalities in the field. The technician can describe a job out loud while walking the property and get a structured estimate back. The dispatcher can take a picture of the work site and have the system read the equipment list off the photograph. These input modalities capture data that previously did not exist, because the operator was never going to type it into a form on a phone at the end of a fourteen-hour day.

AI lets the software adapt to the operator instead of forcing the operator to adapt to the software. Most legacy vertical SaaS in these industries was built around a data model designed for someone in a corporate office, and most of the operators we are trying to serve assume CRUD describes CRM because every CRM they have ever used was, in fact, terrible. Software that meets the operator in their own vocabulary, with their own workflow, on their own schedule, was not really possible before. It is now.

This is the corner of vertical software we expect to spend most of the next decade in. It is where the actual money moves through real-world operations and services, and the founders who understand the chain are the ones we want to be in business with. If you or someone you know is building in this space, we’d love to hear from you.

Benchling was built by the right user.

The Problem Was Not the Paper

Walk into a molecular biology lab and look at the bench. You will find pipettes, centrifuges, gel boxes, and freezers stacked with labeled samples. You will also find a notebook. Not a laptop. A physical notebook with numbered pages, bound so none can be removed, with entries written in ink. Next to every entry, a date and a signature. In many labs, a second signature from a witness.

This was not an artifact of old habits. It was a deliberate system, built around specific legal and regulatory requirements that made paper more defensible than software for most of the history of biomedical research.

The carbon copy lab notebook was the institution’s answer to two problems: intellectual property documentation and regulatory compliance. U.S. patent law required contemporaneous evidence of inventive activity. A properly maintained lab notebook, with bound and numbered pages that prevented insertions, ink entries that prevented erasure, and dated witness signatures that corroborated the sequence of work, was one of the strongest forms of that evidence. FDA regulations under 21 CFR Part 58, governing Good Laboratory Practice, required that all research data be recorded directly, promptly, and legibly in ink. The carbon copy gave researchers an immediate duplicate. One copy stayed in the lab. One copy went with the researcher. Neither could be altered without detection.

This is why paper persisted not just through the 1990s but through the 2000s and into the 2010s, long after the rest of the scientific world had adopted computers for everything else. The notebook was not a failure of imagination. It was an answer to a real problem: how do you create a trustworthy, tamper-evident, legally defensible record of the work that generates billions of dollars in drug development and intellectual property? For decades, paper solved that problem in a way that software did not.

As recently as 2012, the vast majority of academic scientists were still using paper lab notebooks. Most major research universities had no enterprise-wide electronic lab notebook solution at all. The software that existed to replace paper had not convinced most researchers it was worth the switch.

Sajith Wickramasekara, the founder of Benchling, understood why. He had used both.

Why the Software That Existed Didn’t Work

Electronic lab notebook (ELN) software was not new when Benchling launched. The market had been developing since the late 1980s. What it had been developing was the wrong product.

IDBS, founded in 1989, became the dominant incumbent in pharmaceutical and large biotech environments. Its E-WorkBook platform combined electronic lab notebook functionality with laboratory information management into a single large, configurable system built by database consultants for pharmaceutical IT departments. Implementing it required months of deployment, dedicated IT infrastructure, custom configuration, and ongoing IT support. It did not launch a cloud version until 2017, nearly three decades after founding.

It was built to satisfy compliance officers and IT directors, not to make a researcher’s workday faster.

This mattered because the lab notebook problem had two very different versions. The first was the pharma enterprise version: a large organization with IT departments, regulatory affairs staff, and existing vendor relationships needed a solution that satisfied 21 CFR Part 11 electronic records requirements, integrated with existing LIMS (Laboratory Information Management System) infrastructure, and generated audit trails that could survive FDA inspection. IDBS was built for that version of the problem.

The second was the version Wickramasekara had lived: a research lab at a university or an early-stage biotech, where the actual work of science was being done, where there was no IT department, no enterprise budget, and where a researcher trying to find data from three months ago was still searching through a paper notebook or an Excel file on someone’s laptop. For this version of the problem, the existing software was worse than paper. It required institutional procurement, IT setup, ongoing maintenance, and licensing fees that most academic labs simply could not afford.

When researchers did try legacy ELN systems, they encountered auto-logout sessions that timed out after fifteen minutes of inactivity, rich text editors that could not paste from Word documents, unfixed bugs that had been reported for months, and interfaces that looked and felt like database applications from the late 1990s.

LabArchives, which launched in 2009, was the closest prior attempt at the middle ground. It was cloud-based, designed for academic researchers, and did not require institutional IT infrastructure to deploy. It eventually reached more than 350,000 researchers at universities worldwide before being acquired by Dotmatics in 2021. But LabArchives spread primarily through institutional enterprise agreements: universities purchased licenses and researchers used it because it was available.

Benchling took a different approach. It spread because individual researchers chose it for work they were already doing. The distinction between institutional adoption and researcher-led adoption is not semantic. It produces fundamentally different products.

Scientists did not reject electronic lab notebooks because they preferred paper. They rejected the specific software that existed because adopting it meant trading a problem they had learned to manage for a different set of problems that were harder to manage. What Wickramasekara later characterized as roughly 40% of researcher time being lost to administrative work, data entry, and hunting through fragmented records was a cost they paid daily. An enterprise ELN with a six-month implementation timeline and a five-figure annual license was a cost they could not justify.

Nobody had built the product that sat between those two failure modes: modern, cloud-native, designed for the researcher rather than the IT director, and accessible enough that a graduate student could start using it today. That product required understanding both failure modes from the inside. Wickramasekara had lived in the second one.

The Man Behind the Bench and Screen

Sajith Wickramasekara grew up interested in both computers and science. He attended the North Carolina School of Science and Mathematics, a residential high school for students with particular aptitude in science and math. He arrived at MIT to study electrical engineering and computer science, but his interest in biology research pulled him toward the lab.

Before founding Benchling, he worked as a research assistant at Duke University Medical Center and then at Liquidia Technologies, a biotech in Durham building microfluidic drug delivery systems. These were working research environments with functioning wet labs: experiments, samples, notebooks, results that needed to be recorded, reproduced, and built upon. He used the tools that existed. He spent enough time inside those environments to understand them from the inside.

The contrast he kept returning to was specific. As a software engineer, he had version control. He had searchable documentation. He had real-time collaboration with teammates across different locations. He had infrastructure that evolved continuously and improved every year. In the biology lab, conducting work he considered more consequential than most software engineering projects, he had a paper notebook and Excel.

Most software entrepreneurs looked at biology labs as a niche market with complex procurement cycles and slow adoption rates. Most biologists had internalized the limitations of their tools as structural features of the work rather than solvable problems. Both were wrong, but neither had the vantage point to see it clearly.

Wickramasekara had both perspectives simultaneously. He knew the tools that existed in modern software development well enough to understand that what biology labs were running on was not just outdated but structurally broken. And he had spent enough time in those labs to understand exactly where it was broken: not in the compliance layer, not in the audit trail, but in the daily workflow of a researcher designing an experiment, running it, recording what happened, and trying to build on it the next day. He had been that researcher. He knew which parts were broken because he had experienced them breaking on him.

That combination was rare. Biology graduate students who spent their careers in wet labs had stopped seeing the paper notebook as a problem worth solving. Software engineers who had never spent time at a bench had no way to understand which features would actually matter. Wickramasekara sat at the intersection of those two populations, and he was young enough that the contrast between modern software tools and the state of lab infrastructure felt like an obvious failure rather than an accepted condition.

He dropped out of MIT and co-founded Benchling in 2012 with Ashu Singhal, whom he had met in his freshman year. Singhal brought the engineering depth the product would require. Wickramasekara was 21. Paul Graham accepted Benchling into Y Combinator’s Summer 2012 batch and offered a framing that captured the opportunity precisely:

“People who work with biology are important and so to be the most important software for them will surely turn out to be valuable.”

The First Product Was Not a Lab Notebook

Benchling’s initial release was not an electronic lab notebook. It was a free, web-based DNA sequence design and plasmid mapping tool.

Why? Because every molecular biology lab needed to design plasmids and sequences. Every researcher doing that work was using desktop applications that required installation, licensing, and a specific operating system, or older web tools that had not been updated in years. Benchling offered the same capability in a browser, immediately, for free. No institutional procurement. No IT approval. A graduate student could be using it in ten minutes.

This was the entire strategy in one product decision. Not an enterprise ELN sale to a pharma IT department. A free tool that researchers actually needed for their daily work, with zero friction to access. Once a lab adopted Benchling for sequence design, the workflow was already inside the door. The same platform that managed their sequences could manage the experiments that used them. The data that had lived in five disconnected places could live in one.

The free academic tier was not a loss leader with a later plan to charge. It was the distribution mechanism. Wickramasekara understood something the incumbent ELN vendors had never needed to think about: the graduate students and postdocs using the product today were the scientists entering pharma and biotech companies in two to five years. They would bring their workflows with them.

The academic base was the sales force.

Legacy vendors had approached the market from the top down, selling to pharmaceutical IT departments with enterprise contracts. Benchling approached it from the bottom up, earning the trust of researchers before they had any purchasing authority, so that when they did, Benchling was already the tool they knew.

By 2015, more than 6,000 scientists across 1,000 research institutions were using Benchling. Andreessen Horowitz led a $5 million Seed round, with Thrive Capital and SV Angel also participating. The investment thesis, as Balaji Srinivasan wrote at the time:

“Software may be eating the rest of the world, but it has failed science.”

Eighteen months later, Thrive returned to lead the $7 million Series A: a signal that the academic-to-industry pipeline was already working.

The View From the Bench Was the Edge

The pharmaceutical and biotech industry adopted Benchling faster than it had adopted anything the ELN market had previously produced because the product was designed from the researcher’s seat outward, not from the compliance requirement inward.

Wickramasekara could see things that the database consultants who built IDBS in 1989 could not. He understood that what he had characterized as roughly 40% of researcher time being lost to administrative work was not an abstraction. He had experienced it. He understood that compliance requirements were real but secondary for most research workflows, and that building compliance-first produced a product that was effectively useless for the vast majority of researchers not yet in a regulated development stage. He understood that the obstacle to cloud adoption in labs was not a genuine security concern but institutional inertia, and that a tool built for individual researchers would bypass institutional purchasing cycles entirely.

He also understood something that did not surface in any sales conversation with an IT director. The fragmentation of lab data was not just an inconvenience. It was a structural contributor to the reproducibility crisis running through biomedical science.

A 2016 Nature survey of more than 1,500 researchers found that 70% had failed to reproduce at least one peer’s study, and more than 50% had failed to reproduce one of their own experiments. When experimental conditions lived in personal notebooks, when the chain of custody between a biological material and the results it produced was not tracked, when a researcher left a lab and took their notebook with them, the work they had done became effectively unreproducible. One study, published in the journal Molecular Brain, found that when editors requested raw data from submitted manuscripts, more than 97% of authors did not provide it when asked. While narrow in scope, the finding pointed at a structural pattern across biomedical research.

Benchling addressed this not by selling reproducibility as a compliance argument but by making the research workflow itself better. The version control, the linked records, the searchable history: these were product decisions that made daily research faster and easier.

The reproducibility benefit was a consequence of building the right product, not a feature pitch.

That distinction is only available to someone who had done the work. The compliance-first approach produced products that researchers avoided. The workflow-first approach produced a product they actually used, which also happened to address the reproducibility problem.

When You Build for the Right User

In 2017, five years after launch, Benchling introduced the Bioregistry and Inventory modules, connecting physical biological materials, plasmids, strains, antibodies, cell lines, to the experimental records that used them. The expansion followed the same logic as the original product: researchers working with sequences were also working with physical samples, and the connection between the two was entirely unmanaged. The product told you what was missing, if you had spent enough time in the lab to understand what you were seeing.

The valuation trajectory reflects how completely the market validated that logic. From $60.5 million pre-money at Series B in April 2018 to $3.8 billion pre-money at Series E three years later, the company’s growth rate was among the sharpest in life sciences software of that era. The category it had entered had existed for decades. What had not existed was a product researchers would actually choose to use.

In January 2026, Benchling announced a partnership with Eli Lilly’s TuneLab to embed Lilly’s proprietary machine learning models, trained on more than a billion dollars of research data, directly into the platform. The product has moved from a research notebook to an AI-augmented R&D infrastructure. The orientation has not changed. The scientist is still the user the product is designed for.

The full arc of that progress, across fourteen years of company-building, is below.

Key Metrics (2024)

• Customers: 1,300+ biotech companies; more than half of the top 50 global biopharma

• Scientists: 200,000+ active users globally

• Revenue mix: 70%+ enterprise; estimated $200M+ ARR

• Average ACV: ~$175K blended across all tiers (Sacra, May 2024); enterprise plans range to $1M+ for largest customers

• Total raised: $417.88M across 8 rounds (per PitchBook)

• Peak valuation: ~$6.1B post-money (Series F, October 2021)

Where the Moat Actually Came From

The electronic lab notebook market had incumbents with enterprise relationships, deep compliance infrastructure, and in IDBS’s case more than two decades of pharmaceutical deployment experience when Benchling launched. None of that protected them from a 21-year-old Electrical Engineering and Computer Science (EECS) student who had spent time at a research bench.

The protection the incumbents had was real: long contracts, switching costs, regulatory compliance depth, and relationships with pharma IT departments.

What they did not have was a product built for the person actually using it.

Every feature decision in every incumbent product had been made by someone who understood the purchasing side of the market. None of them had written entries in a carbon copy notebook, searched for an experiment result in an Excel file from eight months ago, or experienced the specific daily friction of conducting important research on infrastructure that had not changed since the 1990s.

Wickramasekara was not a biologist who learned to code. He was an engineer and computer scientist who had worked in biology labs, which gave him something neither category had on its own: he could see the researcher’s experience with precision, and he had the technical background to understand what building something better would require. The buyer-builder credential is not a single observation. It is the accumulation of enough time inside a broken system to understand where it is actually breaking, combined with the specific knowledge to do something about it.

That is what the incumbents could not replicate. The ELN vendors who had been selling to pharmaceutical IT departments for two decades had deep relationships with the people who controlled the budget. What they had never built was the product the person at the bench would actually choose to use. Benchling built that product first because Wickramasekara had been that person.

Why We Like the Life Sciences Industry

As a biology major at Rice, I used carbon copy notebooks. I hated writing in them. I always thought it was so tedious and made the labs twice as long as they needed to be.

But I understood why they existed. As an college student, it was to prevent plagarism. In a professional context, it was to ensure compliance and traceability. A properly maintained carbon copy notebook is legally defensible in a way that a digital file on someone's laptop is not. That is not bureaucratic inertia.

It is the product of decades of patent litigation and FDA enforcement shaping how research institutions protect their work. That experience made Wickramasekara's insight immediately legible: the compliance requirement was legitimate, but it had become an excuse to build products that served the institution's legal department and nobody else in the lab.

The ELN vendors building from the top down were answering the right question for the wrong customer. Their question was: how do we satisfy audit requirements and enterprise compliance at the C-Suite level? That is a legitimate problem. It produced real products with real enterprise contracts. What it did not produce was anything a graduate student at a bench would choose to use voluntarily.

Benchling answered a different question: how do we make the daily workflow of the researcher better than paper? Those two questions look like they are competing in the same market. They are not. They serve different customers entirely. The top-down compliance vendors were selling to the executive managing regulatory exposure. Benchling was earning the trust of the person doing the actual work before anyone in that organization had a procurement decision to make. By the time those researchers moved into industry and their companies needed an ELN, the answer was already the tool they knew.

The bottom-up approach built continuity and traceability into the workflow itself rather than bolting it on for compliance. That is a structural difference in product philosophy, and it produced a structural difference in adoption. The researcher did not need to learn a compliance system. The compliance value was embedded in the workflow, not added on top of it.

Life sciences is a space Daring invests in directly, and the reason is exactly the pattern Benchling demonstrates. Regulation runs through the entire industry, not just the research bench. It governs how data gets recorded in the lab, how products get approved, and how they get sold: sales reps operating under FDA promotion guidelines, training programs built around compliance documentation, commercial teams managing audit requirements alongside revenue targets. The same structural gap exists at every layer.

PraxisPro, one of our portfolio companies, operates in this space. Different part of the industry, same dynamic: compliance requirements that shaped how the software got built, and practitioners left working around tools that were never designed for them.

The founders who will move regulated markets are the ones who stopped accepting the friction as a fixed cost and started seeing it as the product. That is the credential Daring looks for. Not the market analysis. The lived experience of the broken tool.

The YC launch post for Legion Health opens with a claim: “the first mental health company ever to have approval to let AI provide actual medical care.”

Their core thesis: psychiatric medication renewals are too slow, too expensive, and too dependent on clinician time that does not exist. The solution is to automate the clinician out of the renewal loop. It is written with the confidence of a founding team that has cracked a distribution problem.

Their solution is simple.

All they need is a “stable” patient on one of fifteen medications. And $19 a month.

AI handles the rest.

The claim is accurate. It is indeed much easier.

What if that process wasn’t supposed to be easy? Like what if it was designed not to be that simple in the first place?

How Legion Health Got Here

Legion was founded by three Princeton roommates: Yash Patel, a former Medicare/Medicaid policy analyst at the Congressional Budget Office; Danny Wilson, who leads AI and engineering; and Arthur MacWaters, who runs product and operations.

They entered Y Combinator’s Summer 2021 batch with a B2B marketplace model. Legion’s first product connected mental health clinicians to healthcare organizations that needed clinical capacity on demand. Hospitals, digital health companies, value-based care providers. By early 2022, they had 800 clinicians across all 50 states and eight enterprise customers. The product had no prescribing authority. It was staffing infrastructure.

When large language models matured, the founders pivoted. In late 2024, they raised a $6.3 million seed round and relaunched as a full-stack, direct-to-consumer telepsychiatry clinic. Legion hired its own clinicians, contracted with insurers, and began seeing patients directly in Texas, focused on medication management. The AI ran the operational layer: scheduling, intake, documentation, risk-scoring, clinical summaries. The clinicians made the prescribing decisions. The founders watched how psychiatric care actually functions and used that data to identify which workflows could be automated.

By early 2026, they received regulatory authorization from Utah to take the next step: autonomous AI renewal of fifteen psychiatric maintenance medications for patients the system classifies as stable, with no physician reviewing individual decisions.

The founders described their approach as the Tesla model. Start with a world-class clinic staffed by real providers, instrument everything in real conditions, automate step by step. A play to encode domain expertise by having models learn from real practitioners.

Legion Health has not published outcome data from its supervised Texas clinic. The only real-world evidence base for removing physician oversight from individual prescribing decisions is internal and undisclosed.

The question is whether the clinical experience they accumulated as operators of a supervised clinic is the same credential that qualifies them to remove the clinician from the loop.

It is not. And the reason it is not reveals everything about how this product was built.

Who Built This, and What They Cannot See

When asked why AI is central to the product, Patel’s answer was not about clinical outcomes. It was about economics:

“The AI is not just a cool thing to do; it’s really meaningfully creating a margin profile.”

That framing is honest. It is also the framing of someone whose primary credential is understanding how healthcare systems pay for care, not what happens inside the clinical relationship when that care is delivered.

Legion’s founding team has no clinical experience. No one is a psychiatrist, a nurse practitioner, or a pharmacist. The company has a medical advisory board with credentialed psychiatrists and is hiring a collaborating physician. That clinical scaffolding is real. But running a supervised clinic as an operator is not the same as having built it as a clinician. The product reflects what the founding team understood when they made the core architectural choices: autonomous AI in the prescribing loop, clinical advisors consulted around those choices rather than driving them.

From Patel’s vantage point, the provider in the renewal loop looks like a cost problem. Clinician time is scarce, expensive, and poorly distributed. The renewal appointment consumes that time for a process that, on paper, looks routine. The logical move is to automate it.

The problem is that psychiatric medication renewal is not routine. It looks routine from the outside. From inside the clinical relationship, it is one of the last places where a clinician can detect that the patient in front of them is less stable than the form suggests.

That is the category error at the center of this product. Legion looked at the clinician in the renewal loop and classified that role as operational. Everything that follows, the product architecture, the pricing, the regulatory strategy, the expansion plan, flows from that single misclassification.

Not All Friction Is the Same

It’s no secret that American healthcare is fundamentally broken. Much of that comes from poorly designed infrastructure. But some of it is a speed limit in a school zone. Both prevent you from getting to work as fast as you like. But treating them as one problem is a critical mistake.

Prior authorization delays, billing overhead, scheduling systems that have not been modernized, documentation that consumes clinician time without adding clinical value. That friction is worth removing. The AMA estimates the prior authorization burden alone consumes hours of physician time per week. That is the friction Legion Health should be targeting.

The provider reviewing a medication renewal is something different. A psychiatrist or nurse practitioner on a fifteen-minute telehealth renewal is not performing an administrative function. They are observing things no symptom questionnaire captures. Is the patient’s affect flat, elevated, incongruent with what they are saying? Are they speaking differently than last time, faster or slower, more agitated or more withdrawn? These are the signals that clinical training builds over years. They are not in the text box.

And the text box is all the AI has.

A patient renewing an SSRI for depression may have recently started a blood pressure medication their cardiologist prescribed. They may be taking a supplement that creates a serotonin interaction risk. They do not mention these things because they do not know they are relevant. Serotonin syndrome can be fatal and is triggered by exactly these kinds of undisclosed interactions. A clinician who knows the patient’s chart looks for this. The renewal checklist does not.

The pilot’s eligibility criteria are meant to address some of this. Only patients with no psychiatric hospitalization and no medication changes within the past year qualify as “stable.” That guardrail sounds meaningful until you consider what it actually screens for. It screens out the recently hospitalized and the recently adjusted. It does not screen for the patient whose stability is about to break down, which is the patient the clinical relationship exists to detect. A patient can meet every stability criterion on intake and be in crisis two weeks later. That gap is not a scheduling inefficiency. It is the reason the renewal visit exists.

From the policy side, the renewal visit and the prior authorization delay look like the same category of problem. They are both friction. They both consume time. They both cost money. The distinction between them, that one is waste and the other is the safety layer, is only visible from inside the clinical relationship.

Legion automated both because from where the founders stood, they could not see the line between them.

The access gap in mental healthcare is real

More than 122 million Americans live in areas where providers are in short supply. Wait times stretch weeks. The system is failing the patients who need it most.

But the access gap is not a product problem. It is a reimbursement and enforcement problem that has been running for seventeen years.

A psychiatrist charges $550 for an initial consultation out of pocket. Medicare reimburses $216. Medicaid reimburses $177. Private insurers pay an average of 13 to 14 percent below Medicare rates for behavioral health.

As of 2024, only 18% of psychiatrists listed in Medicaid provider directories accept new patients. The Mental Health Parity and Addiction Equity Act was passed in 2008. It requires that mental health coverage be no more restrictive than coverage for other medical conditions, but has no enforcement mandate.

Thus, there is still no incentive to fix the access gap. Payers are structured to manage medical loss ratios. PBMs are structured to maximize rebates from drug manufacturers, which frequently does not align with the medications that produce the best clinical outcomes.

I spent time at the corporate venture arm of one of the largest payers and PBMs in the country. That position gave me a direct view of the incentive misalignment that the healthcare system runs on. Not as a policy abstraction. As financial decisions made in actual budget meetings.

Legion Health is not solving this. A $19 monthly subscription removes the provider from the loop without addressing why the provider is scarce. The reimbursement problem remains. The parity enforcement failure remains. What changes is that the patient now has an AI chatbot instead of a clinician, at a price point that the broken system made possible precisely because the human was too expensive under the broken system’s own terms.

The product does not begin with the most underserved patients, the most complex cases, or the least connected communities. It begins with the easiest patients to serve: already diagnosed, already prescribed, already stable enough to qualify, already capable of navigating a digital intake. That is not where the system is most broken. It is where automation is easiest to deploy.

Regulatory Arbitrage Isn’t Healthcare

Utah was not chosen because it has the worst access problem. It was chosen because it has the most permissive regulatory environment for this kind of pilot.

Doctronic, the state’s other AI prescribing experiment, is already in talks with regulators in Arizona, Texas, and Wyoming about comparable sandbox frameworks. Congress introduced the Healthy Technology Act of 2025 to formally allow AI to qualify as a practitioner eligible to prescribe drugs, timed as the FDA was cutting its AI regulatory staff.

The Utah oversight agreement requires monthly reporting of accepted and denied renewal counts, physician concordance rates, and adverse health outcomes. What it does not specify is any performance threshold that would trigger suspension.

The company’s claimed 99.2% accuracy rate is self-reported, based on internal simulation, and measures agreement with clinician decisions rather than patient outcomes. No publicly documented metric defines when the pilot has failed. The pilot runs for one year, but the agreement does not specify who evaluates the outcome, on what criteria, or what finding leads to shutdown rather than renewal.

The expansion map is a regulatory map. It is not a clinical-needs map. The strategy is to establish precedent in permissive jurisdictions, generate proprietary performance data, and expand into additional states before the federal regulatory conversation catches up. That is not a strategy critique. It is a category observation.

Companies that choose where to operate based on which regulators will ask the fewest questions are not doing healthcare. They are doing policy entrepreneurship. The two look similar from the outside, but they produce very different outcomes.

Who holds the bag

In traditional medicine, the accountability structure is legible. The clinician who prescribes holds professional liability. They can be sued for malpractice. Their license can be revoked. The threat of those consequences shapes clinical behavior. It forces judgment to remain attached to a human being.

Legion’s model breaks that chain. In the Utah pilot, no individual physician reviews an individual renewal decision. The AI makes the decision. The company can be sued. But professional accountability becomes diffuse. Financial risk may transfer through insurance. It does not replicate the accountability structure of a licensed clinician whose career depends on getting it right. A physician who repeatedly makes poor prescribing decisions loses their license. An AI that repeatedly makes poor prescribing decisions generates a dataset.

Legion Health’s product works by reclassifying clinical judgment as operational friction and then removing it. That reclassification is what makes autonomous prescribing look efficient. It is also what makes it unsafe.

The strongest version of Legion’s argument is that some care is better than no care. For the patients who genuinely have no access to a psychiatric provider, an imperfect system may be better than nothing. That argument has force. But it only works if the product is actually designed for those patients. Legion’s eligibility criteria select for the opposite: patients who are already stable, already diagnosed, already prescribed, already navigating the system well enough to find a digital intake and fill out a questionnaire.

The patients with no prior diagnosis, no established medication, no clinical relationship, and no digital fluency do not qualify. The product reaches the patients who are easiest to serve and calls that solving access.

The broken incentive structure created the access problem. Removing the provider does not fix the incentive structure. It removes the safety layer on top of it and calls the result innovation.