Two years into the generative-AI investment cycle, the returns picture is starting to sharpen. The horizontal foundation-model arms race has been brutal on capital and gentle on returns; the picks-and-shovels infrastructure layer has produced respectable outcomes for the survivors; and a quieter, faster-compounding category has emerged in the middle: vertical AI.
By vertical AI we mean: software that owns a specific industry workflow (clinical letters, legal contracts, freight invoices, underwriting memos) and uses AI to compress the cost of doing it. These businesses look unremarkable on first read — modest TAMs, boring categories, line-item budgets. But they have the four characteristics that compound:
- Defensible distribution. They sit inside a workflow with high switching cost and aren't competing with a horizontal model for attention.
- Proprietary data flywheel. Every customer interaction produces labelled training data the model improves on — a durable moat that horizontal players can't replicate without buying the workflow.
- Sticky workflows. The customer doesn't evaluate the AI model; they evaluate the completed work product. That changes the diligence question entirely.
- Predictable unit economics. Most vertical AI businesses have ARR multiples and gross margins that look like vertical SaaS — because they essentially are.
The diligence framework we use
When we assess a vertical AI business for a client, the framework starts with one question: what does the workflow look like if the AI disappears tomorrow? If the business is still operating — selling the workflow, the data, the integration layer, the customer relationship — then the AI is a margin expander, not the product. That's the position you want to be in.
If the business collapses without the AI — if the only value proposition is the model — then we are looking at a horizontal model with extra steps. Those businesses are sometimes good outcomes, but they need a different framework, different comparables, and a different posture.
From there, the second-look questions are about data moat depth, workflow switching cost, and regulatory durability. The third is about commoditisation: which layer of the stack — the model, the orchestration layer, the workflow application — will be commoditised next, and is the business positioned to capture the margin that flows away?
What we look for in primary research
In customer interviews we ask three things. Would you switch to a free alternative if one launched tomorrow? The honest answer is the moat. What would have to be true for you to cancel? That's the kill-criterion. What does the alternative buying journey look like — internal build, horizontal vendor, status quo? That's the competitive set as the customer sees it, which is usually not the comparable set the seller's deck shows.
The mid-market opportunity
The underappreciated thesis for the next decade of mid-market software returns: take a $50M ARR vertical SaaS company with 60% gross margins and AI-native its operations down to 80%. The multiple rerating on that EBITDA shift is enormous, and the integration risk is manageable because the workflow already exists.
That's the work we're doing for clients in the enterprise transformation and data-ML-AI space right now — and it's the cleanest application of the vertical AI investment thesis we've seen.