Same models. Different economics.
Vertical AI companies are going to beat horizontal ones in the enterprise. Not because their models are better. They are running the same frontier models everyone else is running.
Vertical wins for three reasons:
Easier to sell. They can put a number on your competitive gap. Nobody else can.
Easier to transform with. They have the standing to tell a client CEO he is wrong about his own operation.
Cheaper to roll out. The industry semantics and agent libraries arrive already built instead of getting constructed on your budget.
All three come from the same source. And to be clear about what that source is, because the obvious objection is a good one: nobody is carrying your proprietary anything to your competitors. What moves between clients in a vertical vendor is the industry's shared substrate. Ontology. Document conventions. Regulatory posture. The standard shape of a claims workflow or a credit memo. That is plumbing. Every firm in your sector already has it and none of you differentiate on it, any more than you differentiate on using GAAP. The vendor amortizes the plumbing so you can spend your money on what actually distinguishes you.
Easier to sell
The generalist pitch is "we can transform any company because we have AI." It is unfalsifiable, and it is the sixth version of that claim your CEO has heard this year. Unfalsifiable claims produce fatigue, not urgency.
The vertical pitch is different in kind. "Here are the benchmarks for your industry. Here is what four of your competitors deployed. Here is your gap, quantified." That is falsifiable, which is exactly why it lands. It converts a discretionary technology investment into a competitive defense decision, and those get funded on different terms and different timelines.
The tell is in the case studies. Every published Distyl case study is anonymized: F20, F50, F100, seven of them, no company named. Rogo names Rothschild, Jefferies, Lazard, Moelis, and Nomura. Legora names Linklaters, Dentons, Goodwin, and Bird and Bird. That is not a marketing preference. For a vertical player the reference is the product, because the named customer becomes the benchmark that wins the next deal.
Easier to transform with
You do not capture AI value by deploying a tool. You capture it by changing roles, decision rights, gearing ratios, and accountability. Which requires a leadership team to accept being told they are wrong about how they run their company.
That permission is expensive. A CEO with twenty years in specialty pharma distribution does not take operating model advice from someone who learned the industry during discovery. He takes it from someone who has been inside three other distributors' exception handling and can name which of his assumptions stopped being true.
This is structure, not intelligence. I spent years at McKinsey and I know what a good generalist team can do. But a partner with no independent view of the industry has no basis to challenge the process you described to them. So they optimize it. They make the existing thing faster. They optimize the drive. They cannot cancel the trip.
MIT's NANDA study is the evidence I would put in front of a board. Ninety five percent of GenAI pilots produced zero measurable return. Underneath that headline: capabilities bought from specialized external vendors reached production about two thirds of the time, versus one third for internally built systems. Aditya Challapally, who led the research, put it plainly: "Unlike consumer tools such as ChatGPT, which adapt flexibly to individual use, enterprise AI requires careful integration into existing systems. Without that, pilots rarely scale."
Cheaper to roll out
This is the reason that compounds, and the one buyers underweight.
A horizontal player shows up with a control plane: development studio, evaluation scaffolding, InfoSec guardrails, audit trails, an agent harness that swaps frontier models as the leaderboard changes. Genuine IP. Also the layer deflating fastest, being absorbed into the model providers, the hyperscalers, and open source.
A vertical player shows up with the work already done. Rogo, which sells only into institutional finance, maintains an evaluation suite built from real valuation and thesis tasks by people out of Lazard, J.P. Morgan, and Barclays, with native integrations to CapIQ, FactSet, LSEG, and PitchBook. Legora, which sells only into law, is at 800 customers across 50 markets with the drafting conventions of the firms it serves built into the product. Institution number 251 and customer number 801 inherit all of it on day one.
Distyl arrives at Fortune 500 client eleven knowing how to build agents. That is not nothing. It is also not the same thing, and the difference shows up as months on your implementation timeline and headcount on your invoice.
Kleiner Perkins, explaining why they led Rogo's round: generic AI tools "fall apart the moment they encounter the precision and auditability that institutional finance demands."
The best argument against me
Three facts cut the other way, and the version of this argument that ignores them is not worth reading.
Vals AI benchmarked legal research accuracy. Legal-specific tools scored 80 percent. ChatGPT, with no legal specialization, also scored 80 percent. Lawyers scored 71 percent.
Stanford's RegLab found purpose-built legal tools from Lexis and Westlaw still hallucinate 17 to 33 percent of the time.
Working analysts are brutal in public. One investment banking director on Rogo: "totally useless from VP upwards."
If benchmark accuracy were the unit of work, I would be wrong. It is not the unit of work. Daniel Ho and colleagues at Stanford made the point precisely in PNAS this year, in a paper titled "There is no free benchmark": "Legal AI lacks what we call legibility. We know surprisingly little about the performance of legal AI systems, the kinds of mistakes they make, and the likelihood of error."
The score tells you the model got the question right. It tells you nothing about whether the output was usable. An analytically correct valuation in the wrong template is worthless, and every managing director reading this knows it. Which goes back to - "it's not the model, it's the end to end workflow that matters" (from an earlier article)
What to ask your next AI vendor
How much of the industry work arrives already built? Ask to see the evaluation suite. If there isn't one, you are the evaluation suite.
Am I buying infrastructure or transformation? You may need both. They have different half-lives and should not sit in one contract at one price.
Who on your team has run this business before? Not studied it. Run it.
The point
Strip out the positioning and the argument is simple. A vertical AI vendor walks in the door already knowing your industry's language, already holding the competitive benchmarks that make the business case, and already carrying the agent libraries that would otherwise take your team eighteen months to build. The horizontal vendor walks in with excellent tooling and a blank page where your industry should be.
Horizontal players are confined to the layer deflating fastest. That is not new. We have paid for this translation layer before, under a different name, for about thirty years, and next week I want to talk about who used to do it.