The Equal Access Problem
Foundation model access is becoming available to everyone. Any team with $20/month and a credit card can call GPT-4. The same model that powers enterprise applications is available to any developer on the planet. When everyone can access equivalent model capability, differentiation migrates away from the model and toward everything around it.
That "everything around it" is where vertical AI wins. Domain knowledge that took years to accumulate. Workflow integration that required hundreds of customer implementation hours. Accumulated training data from real deployment at scale. Compliance infrastructure that took 18 months to build. Switching cost that accumulates with every week of use.
The Compounding Mechanics
Vertical AI companies compound in three distinct ways that horizontal tools cannot replicate.
Data flywheel: Every transaction processed generates labeled outcomes that improve the model for the next transaction. A healthcare revenue cycle AI that processes 10 million claims has training data that a new entrant can't acquire by reading documentation or scraping the web. The only way to get it is to process the claims — which requires customers — which requires the data to exist first. This circular advantage is self-reinforcing.
Workflow deepening: Every integration point adds switching cost. A product that sits inside the EHR, the billing platform, and the payer submission portal isn't one product anymore — it's three integrations that would all need to be rebuilt with a replacement. Buyers rationally avoid that disruption.
Account expansion: A company that starts with underwriting can expand into claims, fraud detection, and pricing. Each new module increases total switching cost while leveraging existing data and relationships. NRR above 120% becomes achievable when expansion is natural and structurally enabled.
From the Book
“The first vertical AI investment that really worked for me had nothing particularly impressive about the model underneath it. What it had was three years of domain-specific training data from real deployment, a team from inside the industry they were serving, and integrations so deep that the largest customer told me they'd rewritten their entire workflow around the product. That's a moat. The model is just the engine.”
— Trace Cohen, The Value Add VC
The Breadth Trap
Horizontal AI tools face a strategic contradiction. To grow, they need to serve more users in more industries. But serving more users in more industries requires generality — the ability to work well enough for everyone — which conflicts with the depth that creates defensibility. The result is a product that's good enough for many use cases but exceptional for none.
Vertical AI companies avoid this trap by design. They choose a beachhead — one industry, one workflow, one buyer type — and go impossibly deep. Once the beachhead is owned, they expand from adjacent workflows and use cases within the same buyer relationship, not by adding new verticals.
What This Means in Practice
When evaluating a vertical AI company, the core question is: does the team understand this industry at the level required to build something that an insider would trust? Not "did they read about it." Not "do they have an advisor from the industry." Do they have the deep domain expertise to make product decisions that a generalist team would get wrong?
The model isn't the moat. The moat is what you build around it while everyone else is staring at it.