Why I'm Betting on Vertical AI in Regulated Industries
A year ago, I was on a Teams call at Lazard advising a European venture firm on the sale of several minority stakes. It was my daily routine – models, marketing materials, due diligence. What stuck with me wasn't the work. It was watching how the partners made decisions.
They weren't debating discount rates or terminal multiples. In fact, they didn't even have a model. They were debating beliefs. What mattered wasn't precision, it was conviction. Where the world was heading; they backed their conviction with capital.
That was the first time I seriously asked myself a question I haven't stopped thinking about since: if I were deploying capital today, where would I actually put it – and why?
The answer I keep returning to is vertical AI in regulated industries.
This isn't a bet on AI as a buzzword. It's a play on where durable advantage forms when technology collides with real-world constraints.
The Problem with Horizontal AI
Most horizontal AI products share the same underlying weakness. They sit too close to the model layer and too far from the workflow.
You take a general-purpose model, wrap it in a clean interface, and sell it to a broad audience. Early adoption looks strong. Revenue comes quickly. But the product is fragile. Competitors can replicate the functionality. The underlying models improve and commoditise. There's no moat – distribution platforms integrate similar features natively.
The result is predictable: shrinking differentiation and compressed margins.
This isn't because founders are lazy or talent is lacking. It's structural. If your product doesn't require deep domain understanding to build or maintain, it can be competed away faster than it compounds.
Horizontal AI often becomes a feature, not a business. And features don't defend themselves for long.
Why Vertical AI Compounds Differently
Vertical AI companies look different from the start. They are narrower, slower to build, and harder to explain in a pitch deck. That's precisely the point. Go deep in a specific industry – healthcare, legal, veterinary – and build a product that actually understands the workflow.
So, why does this work?
Let's start with domain depth. If you're building software for litigation, dentistry, or veterinary clinics, you need to understand workflows that aren't documented anywhere cleanly. You need to know how decisions are actually made, where responsibility sits, and what happens when something goes wrong. That knowledge takes years to accumulate and can't be hired overnight.
Then there's regulatory complexity. In healthcare, legal services, or finance, compliance isn't an edge case. Most founders see this as friction. I see it as filtration. Regulation reduces the number of credible entrants and raises the cost of mistakes. That slows competitors down while increasing the value of trust once it's earned.
Now we can turn to workflow lock-in. Vertical AI tools don't just assist. They replace steps in mission-critical processes. Once a product is embedded in daily operations, switching costs are real. Retraining staff, migrating data, rebuilding procedures – all of that is far more painful than marginal product improvements are worth.
Finally, there's data gravity. Vertical products generate domain-specific data that improves the system in ways competitors can't easily match. A litigation platform trained on thousands of case chronologies doesn't just get faster. It starts to reason differently. That gap widens over time.
These advantages don't show up in early traction charts. They compound quietly.
Where This Thesis Breaks
This is not a universal rule. Many vertical AI companies will fail.
Some regulated industries are too fragmented to sell into efficiently. Others have incumbents with entrenched distribution and closed data. Some workflows are too bespoke to automate meaningfully. And in a few cases, foundation model providers may successfully move far enough up-stack to compress vertical margins.
The thesis only works where three things align:
Regulation that raises barriers without freezing innovation
Workflows that are repeatable enough to standardise
Customers who feel the pain acutely enough to switch
Miss any one of those, and "vertical AI" becomes a slogan rather than a strategy.
Failure modes I'm actively watching for:
Foundation model leapfrogging. If GPT-6 or Claude 5 can handle domain-specific reasoning out of the box, the "vertical depth" advantage compresses. The moat becomes the workflow integration, not the intelligence layer. I'm watching how quickly general models close the gap on specialised fine-tuning.
Incumbent response at Series B. Vertical AI companies often look safe at seed because incumbents are slow. By Series B, the best ones have gotten the incumbent's attention. The question is whether workflow lock-in compounds faster than enterprise sales cycles allow incumbents to respond. Some verticals have stickier workflows than others.
Horizontal re-bundling. Salesforce, Microsoft, and the hyperscalers will eventually offer "AI for legal" and "AI for healthcare" as platform features. The question is whether they can match domain depth or just good-enough automation. I think they'll struggle with true vertical complexity, but I've been wrong about platform power before.
What Most People Get Wrong About Vertical AI
The mistake I see most often is focusing on the label instead of the mechanism.
"AI for X" isn't a thesis. It's a category. The real question is why this workflow, in this industry, at this moment creates defensibility.
Many investors over-index on TAM and underweight adoption velocity. Others confuse compliance risk with moat, assuming regulation alone protects returns. It doesn't. The protection comes from how deeply the product reshapes daily work.
That distinction only becomes clear when you spend time with users, not decks.
What This Looks Like in Practice
This thesis isn't abstract. It's the lens I've used to analyse companies in practice.
Three examples – all VC-backed – illustrate the pattern:
La Fraise is building AI agents for dental clinics that handle admin, explain treatment plans to patients, and surface financing options. Dentistry is a forgotten healthcare vertical: large, highly regulated, and severely under-digitised. In France alone, roughly half of treatment plans are abandoned, not because patients don't want care, but because admin complexity and opaque pricing break trust upfront. The founding team came from Doctolib and has already scaled regulated healthcare software, with deep familiarity around reimbursement, compliance, and provider trust dynamics.
Wexler is attacking litigation, one of the most manual and time-intensive workflows in legal services. Lawyers still spend days constructing case chronologies from thousands of documents. Wexler compresses that into hours. The founder isn't a lawyer, but grew up immersed in the legal system, with a father who spent decades as a barrister and High Court judge, and a brother who is now a partner at Cleary Gottlieb. He has domain proximity without incumbent thinking – a rare combination.
Lupa is building an AI-native operating system for veterinary clinics. The same pattern repeats: a large under-digitised market, professionals drowning in admin, and software that works with clinicians rather than trying to replace them.
Each of these companies is explored in more detail in the case studies.
This Isn't Theoretical
This lens shaped how I first analysed Lupa Pets, a veterinary software company rebuilding practice management from scratch. It's the same framework I've since applied to legal AI and other regulated verticals.
In each case, the question wasn't "is this AI-powered?" It was: if this works, how hard will it be to displace five years from now?
Horizontal AI is a race to commoditisation. Vertical AI is where durable value gets built. That's the bet I'm interested in making.