Vertical AI: Why the Winners Go Narrow

Every founder now rents the same brain. You, me, and the person building the exact thing you are building all call the same three or four models. The intelligence is not the edge anymore. It is table stakes, priced by the token, available to anyone with a credit card.

You can see it in where the capital is going. The vertical AI companies, the ones built for a single industry like law or medicine or code, are raising at valuations that double in months and pulling in the majority of new AI funding, while the general everything assistants get commoditized by the same labs that make the models. The easy read is that AI is hot. The real read is narrower.

So here is the question that actually decides who wins: once everyone has the same intelligence, what makes yours worth paying for? The answer showing up over and over is not “a smarter model.” It is “a narrower one.” The companies compounding fastest picked one industry, one painful job, and went so deep that a general model pointed at the same problem looks like a tourist.

I run two companies where AI already does most of the first-draft work, and the lesson has been the same in both: the money is not in being able to do everything. It is in owning one thing so completely that switching away from you means tearing out the walls.

What this covers

• The problem: everyone got the same brain
• The framework: the Wedge
• The Labor Line: your TAM is the wage bill
• The Depth Ladder: how deep is deep enough
• What the wedge buys that a horizontal model cannot
• The horizontal exception
• Picking your wedge: a worked example
• What most people get wrong about going narrow
• What to do Monday morning
• FAQ

The problem: everyone got the same brain

For thirty years, software startups competed on capability. You could write code other people could not write, or could not be bothered to write, and that gap was your business. The moat was the building.

AI collapsed that. When a frontier model can draft a contract, read a scan, summarize a deposition, or write a working function, the raw capability stops being scarce. Whatever you can prompt a model to do, your competitor can prompt the same model to do by tomorrow afternoon. I have shipped features whose entire value was clean access to a model, and I watched them get worse every time the model got better, because the model shipped my feature for free inside its own product.

The reflex, when capability goes cheap, is to go wide. Build the everything assistant. Serve every team, every use case, every industry, because the model can technically handle all of them. It feels ambitious. It is actually the most crowded, least defensible place you can stand. You are competing with the labs themselves, who own the model and can undercut you to zero, and with a thousand other founders who had the same wide idea the same week.

The graveyard is already large. Most thin AI wrappers fail, a large share never make a dollar, and the ones that die tend to die the same way: gross margins compress under inference costs while the platform absorbs the feature they were selling. When your product’s core value is “we prompt the model better,” a user replicates that value by pasting your prompt into the model’s own chat box.

Going narrow is the counterintuitive escape. Not narrow as in small ambition. Narrow as in a specific job, in a specific industry, done so completely that the general tool cannot follow you in. The winners of this cycle are not the broadest products. They are the deepest. That is what “vertical AI” actually means, and why it keeps beating horizontal AI in the places money is changing hands.

The framework: the Wedge

A wedge is the opposite of a platform. A platform tries to be wide from day one and hopes depth comes later. A wedge is a sharp point you drive into one narrow crack in a market, and once you are in, you widen from a position you already own.

The shape matters, so here is the picture I keep in my head.

The tip is one painful, expensive, repeatable job inside one industry. Not “AI for law.” One job a lawyer or a paralegal does every week that they hate and that costs real money when it goes wrong. You solve that one job better than anyone alive. Then you widen: the adjacent step in the same workflow, then the whole workflow, then the place all the work and data live, and eventually the layer the industry runs on. Each widening is easier than the last, because you already have the customer, the trust, and the data.

Three things make the wedge defensible, and none of them is the model. First, the domain data you cause to exist by running the job, which no general tool has because the job runs inside you. Second, the workflow you live inside, so removing you means rebuilding how the work happens. Third, a specific buyer who has decided to trust you with an outcome they are accountable for. A horizontal model has none of these. It has capability. You have capability plus a place to put it that took years of domain work to earn.

This is a different axis from the one I wrote about in the Commoditization Clock. That piece is about altitude, how high above the model’s reach you build so a release does not absorb you. The wedge is about width, how narrow you go and which single market you own. You want both: build above the waterline, and drive a narrow wedge into one vertical. Altitude keeps the model from eating you. Width keeps every other founder from copying you.

The Labor Line: your TAM is the wage bill

The oldest objection to going narrow is “the market is too small.” It killed a lot of good vertical SaaS companies, because they were selling software into a thin slice of the IT budget. If you sell a scheduling tool to dentists, your ceiling is however much dentists will pay for software, which is not much.

Vertical AI broke that ceiling, and this is the part most founders miss. A vertical AI product does not compete for the software budget. It competes for the labor budget. When a personal injury firm pays EvenUp to produce demand letters and case valuations, they are not buying a license. They are buying paralegal output. The comparison in the buyer’s head is not “this tool versus that tool.” It is “this versus another headcount.”

The numbers are not close. US enterprise software spend is roughly $450 billion a year. US labor spend is around $11 trillion. In healthcare alone, administrative staff outnumber physicians, and about $740 billion flows to administrative services each year against roughly $63 billion of IT spend. When your product does the work instead of helping a human do the work, your addressable market moves from the small line on the income statement to the enormous one.

This is why a legal AI company can hit hundreds of millions in recurring revenue without serving every industry. It is charging against what firms already spend on legal work, which is a very deep pool. Bessemer’s read is that vertical AI’s eventual market value will be many times larger than legacy vertical SaaS, precisely because it eats into services spend rather than sitting next to it as another tool.

The retention math backs this up. Vertical software that genuinely learns from its domain data tends to hold net revenue retention north of 130 percent, while broad horizontal tools sit lower, often between 110 and 120. Retention is the tell, because it measures whether customers dig in deeper over time or quietly drift away, and depth is what makes them dig in. Valuations follow the same logic. By 2026, the same revenue at the same stage was getting priced at roughly eight times revenue for a thin wrapper and around three times that multiple for a defensible vertical with real workflow ownership. The market is not confused about the difference. It is paying for the moat, and the moat is the wedge, not the model.

The practical consequence for how you price is large. If you anchor to software, you charge a small monthly seat fee and fight on features. If you anchor to labor, you charge for output, for the letter produced or the claim adjudicated or the note written, and you can charge a meaningful fraction of what the human alternative costs while still saving the buyer money. I wrote more about pricing this way in revenue models for AI products, and it is the single biggest reason narrow does not mean small.

One caution, because I have watched founders get drunk on the $11 trillion number. Selling against labor means you are held to the standard of labor. Nobody audits a software feature the way they audit a person who might be replaced. The bar for accuracy, accountability, and trust is higher, and clearing it is most of the work. The wedge is deep because clearing that bar in one domain is hard, which is exactly why the moat holds once you do.

The Depth Ladder: how deep is deep enough

Not every “vertical” product is defensible. Slapping an industry name on a chat box does not make you vertical AI. The thing that decides whether the model absorbs you or the industry cannot live without you is depth, and depth has rungs.

Rung one is the surface tool. Chat with the tax code, chat with the building codes, chat with the clinical guidelines. It feels vertical because the content is niche, but the mechanism is generic, and the model absorbs it the moment it adds a document upload or a domain mode. This is the rung where the wrapper graveyard lives. I covered the tell in the AI wrapper trap: if your value is that you prompt the model well over the industry’s PDFs, you do not have a moat, you have a demo.

Rung two is the workflow. You do not just answer the question, you run the steps of the job. You draft the letter, route it for the human check, file it, track the response, and handle the exception. Now you are inside how the work happens, and removing you means someone has to rebuild that flow.

Rung three is the system of record. The data the industry needs to run now lives inside you. The matter history, the patient notes, the claims ledger. Leaving you means losing the record, which almost nobody will do. This is where switching costs get real, and where the data you cause to exist becomes the compounding advantage I described in the data moat playbook.

Rung four is the operating layer. The industry runs on you. Your product is where the work starts every morning, the place decisions get made and logged, the rails other tools plug into. Ripping you out is a project nobody wants to sponsor. Very few companies get here, and the ones that do are close to unkillable.

The honest test is simple. If the frontier model shipped a great version of the industry’s chat tool tomorrow, would you lose customers? On rung one, yes, instantly. On rungs three and four, the model’s chat tool is a feature you might even plug in, because the thing customers stay for is the workflow and the record, not the answering. Aim to be at least on rung two before you raise a dollar, and have a clear path to three.

What the wedge buys that a horizontal model cannot

Look at the vertical AI companies actually working in 2026 and the pattern is consistent. None of them won on a smarter model. They won on the three things the wedge buys.

The first is proprietary domain data. Harvey, in legal, built on the tacit reasoning of senior partners and the shape of real legal work, none of which sits in a general training set. It crossed a few hundred million in recurring revenue serving law firms and legal departments, roughly tripling revenue in under a year, because the product understands legal work at a depth a general assistant does not. Abridge, in healthcare, sits inside the clinical conversation and turns it into documentation, accumulating the domain data that makes the next note better. The data comes from running the job, and running the job requires being trusted inside one industry, which is the whole point of the wedge.

The second is workflow lock-in. Cursor, in coding, is not just a chat window over a model. It lives inside how code gets written, with retrieval and multi-file editing tuned for that one job, which is why it reached around two billion in recurring revenue while general chat assistants existed the whole time. The model was available to everyone. The workflow was not.

The third is a specific buyer who trusts you with an outcome. Sierra, in customer service, crossed significant recurring revenue by owning the outcome of a support interaction for a specific kind of buyer, not by being generally helpful. EvenUp, in personal injury law, produces the paralegal work product a firm is accountable for. The buyer is not shopping for intelligence. They are buying a result they can put their name on, and they will pay far more for that than for a clever tool.

By 2026 the market was rewarding that depth in a way it never rewarded thin tools. The legal wedge reached an eleven figure valuation while tripling revenue in under a year. A medical answer engine built only for doctors and a documentation company built only for the clinical visit each crossed multi billion dollar valuations by owning one job in one industry. What the capital is underwriting in every one of these is not model access, because model access is the one thing every competitor already has. It is the accumulated domain data, the workflow, and the trusted seat next to a buyer who is accountable for the result. That is the wedge, and it is the part a general model cannot buy its way into.

Notice what is not on that list: a proprietary model. You almost certainly will not out-train the labs, and you do not need to. You need the model to be good enough, then you win on everything around it. This is the same reasoning I use for build versus buy in the AI era. Rent the intelligence, own the wedge.

The horizontal exception

I want to be honest about where this thesis bends, because it does. Horizontal AI companies are winning too. Enterprise search sits across every department and is doing very well as a single, broad product. General assistants from the labs are used by hundreds of millions of people. So “always go narrow” is too clean.

Here is the real line. Horizontal wins in two situations. The first is when the “vertical” is actually all of knowledge work, and the job is genuinely horizontal by nature, like searching across everything a company knows. The second is when you have a distribution machine most founders do not have. The labs win horizontally because they own the model and the front door. A large incumbent can win horizontally because it already sits on every desktop. Their moat is not the wedge, it is reach, which they already paid for.

You, most likely, have neither. You do not own a frontier model and you do not have a hundred million users. For a founder without a distribution engine, horizontal is a fight you enter with no weapon, against opponents who own the battlefield. The wedge is the one entry where your disadvantages stop mattering, because a giant with a billion users still has not done the unglamorous domain work to win one narrow job in one industry. That is the crack you drive into.

The other honest caveat is that horizontal platforms will keep trying to absorb verticals from above, and sometimes they will. The defense is depth. A general platform can add a legal mode. It cannot easily become the system of record that a firm’s entire matter history lives in. If you are on rung one when the platform comes, you lose. If you are on rung three, the platform becomes something you might integrate. This is the same dependency risk I mapped in platform risk for founders, and the answer is the same: own something the platform structurally cannot, and do not build your company inside the blast zone of a feature toggle.

Picking your wedge: a worked example

Abstract advice about “going narrow” is easy to nod at and hard to act on, so let me run a real narrowing on a generic idea and show where it lands.

Start with the idea most people would pitch: “AI for HR.” That is not a wedge, it is a category. HR spans hiring, onboarding, payroll, benefits, compliance, performance, and offboarding, across every company on earth. A general model already does a passable version of most of it. If you build “AI for HR,” you are on rung one competing with everyone.

Narrow by industry first. Not HR for everyone, HR for one industry with real pain and real money. Multi-state restaurant franchises, say. They hire constantly, churn constantly, and operate across states with different labor laws, which makes compliance genuinely painful and expensive to get wrong.

Narrow by job second. Not all of HR for restaurants, one job. Onboarding compliance: making sure every new hire across every location has the right documents, certifications, and state-specific paperwork completed correctly before their first shift. It is repetitive, it is expensive when it fails, and a franchise owner loses real money and real legal exposure when it does.

Now check the depth. Rung one would be a chatbot that answers questions about restaurant labor law. Absorbable. Rung two runs the onboarding workflow end to end: collects the documents, checks them against the right state’s rules, flags gaps, routes the exceptions. Rung three becomes the system of record for every hire’s compliance status across the whole franchise group, the thing an auditor asks for. That is a company. Same starting idea, but narrowed down the ladder until it is defensible.

Price it against the labor line, not the software line. The comparison is not “another HR app.” It is the cost of a compliance person, plus the cost of the fines and lawsuits when onboarding is done wrong. Charge a fraction of that and you are cheap to the buyer and rich to yourself. Then, and only then, do you widen: scheduling compliance, wage-and-hour, the next state, the next franchise vertical. You expand from a place you already own, which is the whole promise of the wedge. And you find your first ten customers before you write one word of generic AI marketing, because the wedge is proven by a buyer paying, not by a landing page. That first-customer discipline is the same one behind audience-first distribution.

What most people get wrong about going narrow

The most common mistake is treating “vertical” as a market size instead of a depth of integration. Founders pick a small industry, build a shallow tool, and think the niche itself is the moat. It is not. A shallow tool in a small market is just a horizontal wrapper with fewer customers. The narrowness that protects you is not the size of the logo pool, it is how far down the workflow you go. You can be narrow and shallow, which is the worst place to be, or narrow and deep, which is the best.

The second mistake is picking a vertical you do not understand because the market looks attractive from a spreadsheet. Domain depth is the whole advantage, and you cannot fake it. The founders winning in legal, in healthcare, in field service, either came from the industry or embedded in it long enough to know where the bodies are buried. If you pick a vertical you have no feel for, a founder who lives in it will out-narrow you, and your only edge, the domain knowledge, belongs to them.

The third mistake is the one that feels smartest and hurts most: staying broad to keep your options open. Optionality feels safe. It is the opposite. When you serve everyone, every customer wants something slightly different, nothing reaches production cleanly, and you never accumulate the data or the workflow depth that compounds. Breadth spreads you across a hundred shallow puddles. Depth digs one well. The wells are where the water is.

Here is the reframe I had to internalize. Going narrow feels like leaving money on the table, all those other industries and use cases you are saying no to. But you are not leaving that money behind. You are earning the right to it. You get the adjacent market by first owning one market so completely that expanding is a downhill walk. The founders who tried to grab all of it at once got none of it. The ones who took one narrow job and refused to be pulled wide are the ones who now have the whole industry calling them the default.

And the last thing people miss: narrow does not mean unambitious. The wedge is a strategy for getting big, not staying small. Amazon started with books. The narrowest, most owned entry point is the fastest path to the widest eventual footprint, because you expand from strength instead of fighting on every front at once with no home base to defend.

What to do Monday morning

If you are staring at a broad AI idea and wondering how to sharpen it, run it through the Wedge Test. Every honest answer on the right means you have a real wedge. Every answer on the left means you have a wrapper wearing a costume, and you should keep narrowing.

Then take these five steps. Pick one job, in one industry you actually understand, that a person is paid to do today and hates doing. Write down the exact cost of that job going wrong, in dollars and in liability, because that number is your pricing. Go get the first ten paying customers before you build anything general, because a wedge is proven by a buyer, not a deck. Instrument how often a human has to step in on your core task, and treat anything above a couple of percent as a product that is not done, because an agent that needs a human every few tasks is headcount with a nicer interface, not a scaling business.

And climb the ladder on purpose. Know which rung you are on now and which rung you are building toward next. If you are on rung one, your only job is to get to rung two before you raise or hire. Do not add a second vertical until the first one runs on you. The discipline of refusing to go wide before you have gone deep is the entire game. For the larger operating picture this sits inside, I keep coming back to the AI-native founder playbook.

Frequently asked questions

What is vertical AI, and how is it different from horizontal AI?

Vertical AI is built to do one specific job inside one specific industry, deeply enough that it owns the workflow and the data around that job. Horizontal AI is built to be broadly useful across many industries and tasks, like a general assistant or an enterprise search tool. The difference is not the model, which is often the same. It is depth. Horizontal AI answers questions for anyone. Vertical AI runs a job for someone in particular, and holds the data and workflow that make the next run better.

Isn’t a single vertical too small to build a big company?

That was true for vertical SaaS, which sold into a slice of the IT budget. It is not true for vertical AI, because vertical AI is priced against labor, not software. The comparison in the buyer’s mind is another headcount or an outsourced service, not another app. US software spend is around $450 billion a year, but US labor spend is around $11 trillion. When your product does the work instead of assisting it, a single vertical can support a very large company. Bessemer expects vertical AI to be worth many times more than legacy vertical SaaS for exactly this reason.

How do I choose which vertical to go after?

Pick an industry you genuinely understand or can embed in, because domain depth is the advantage and you cannot fake it. Inside that industry, find one job that is painful, expensive, repeatable, and costly when it goes wrong. Confirm three things: a person or a spreadsheet already does this job today, fewer than a handful of funded competitors target the exact same buyer, and you can name the specific role you are replacing. If you cannot name whose job it is, you have a category, not a wedge, and you need to narrow further.

Won’t ChatGPT or a big platform just absorb my vertical?

They will try, and if you are shallow they will succeed. A general platform can add an industry mode over a weekend. What it cannot easily do is become the system of record that an industry’s history lives in, or earn the trust to be accountable for a regulated outcome. The defense is depth. If your value is answering questions over the industry’s documents, you will get absorbed. If you run the workflow and hold the record, the platform’s new feature becomes something you might plug in rather than something that kills you.

Is vertical AI just a wrapper with an industry label?

It can be, and that version fails. A wrapper wearing a costume chats with the industry’s PDFs and calls itself vertical. The real thing goes deeper: it runs the steps of the job, holds the data the job produces, and takes responsibility for an outcome a buyer is accountable for. The test is what happens if the underlying model ships the same capability for free. A wrapper loses its customers that day. A real vertical AI product keeps them, because they stay for the workflow and the record, not for access to the model.

How narrow is too narrow?

Narrow enough that one process owner can decide to buy, that you can name the exact role you replace, and that you can reach the first ten customers by hand. Too narrow is when the total number of buyers who do this exact job cannot support the company you want to build even after you expand. In practice, most founders err wide, not narrow. If your instinct says the wedge feels almost embarrassingly specific, you are probably close to right. You widen later, from strength, once you own the first job.

What changed between vertical SaaS and vertical AI?

Vertical SaaS showed information and assisted a human who still did the work. Vertical AI reasons and executes, so it does the work. That single shift moves the buyer’s budget from the software line to the labor line, which is far larger, and it changes the moat from workflow familiarity to accumulated domain data plus outcome accountability. It is why a narrow AI company can reach revenue that a narrow SaaS company in the same industry never could, and why the defensibility runs deeper once you are established.

When should I expand beyond my first wedge?

When the first job runs on you, not before. The signal is that customers treat your product as the default for that one job, your data and workflow make you hard to remove, and expansion means selling an adjacent job to buyers who already trust you. Expanding earlier, before you own the first wedge, throws you back into the broad, undefended fight you were trying to escape. Widen from a position you already hold. The whole point of the wedge is that the second market is a downhill walk once the first is yours. For the mindset behind these calls, see how founders should think about AI and the human edge that has no clock in the taste moat.

The cycle everyone is living through rewards the opposite of what it looks like it rewards. It looks like a race to the widest, smartest, most general product. It is actually a race to own one narrow thing so completely that the general product becomes a commodity you buy by the token and point at the job only you know how to do. Intelligence is cheap now. The wedge is the moat. Go narrow, go deep, and expand from a place that is already yours.