Most AI startups are a thin layer over someone else's model. Here's a practical framework for telling durable companies from demos — whether you're investing, joining or buying.
The AI boom has produced thousands of startups, and a large share of them are a clever prompt wrapped around a model someone else built. Some of those will become real businesses; many won’t. This guide is a practical checklist for evaluating an AI startup — useful whether you’re considering investing, taking a job, or buying the product.
Most AI startups sit on top of a foundation model from OpenAI, Anthropic, Google or an open-weight provider. That’s fine — but it raises the key question:
If a model provider shipped this as a feature tomorrow, would the company still exist?
The answer reveals whether you’re looking at a product or a temporary gap in the market. Durable companies have something the model alone doesn’t provide.
A thin wrapper is easy to copy. Look for at least one of these moats:
The model is rarely the moat. What surrounds it usually is.
AI products have a cost structure traditional software doesn’t: inference is expensive. Every query may call a paid model, so usage costs money in a way a static web app never did.
A company growing fast while losing money on every query is buying revenue, not building a business.
The best AI startups improve as they’re used: more usage produces more data, which makes the product better, which attracts more usage. Ask whether such a flywheel exists, or whether the product is static.
On the team, you want people who understand both the domain and the technology. A brilliant ML team with no sense of the customer’s problem tends to build impressive things nobody needs.
AI demos are dangerously good. The gap between a polished demo and reliable production is where most startups stall.
Building on one model provider means inheriting its pricing, availability and policy changes. Strong startups stay model-agnostic where they can, so they can switch providers as the market shifts. Total dependence on a single vendor is a structural risk worth pricing in.
| Signal | Healthy | Worrying |
|---|---|---|
| Defensibility | Data, workflow or distribution moat | Pure prompt wrapper |
| Margins | Improving with scale | Eroded by inference cost |
| Demo vs reality | Works on messy inputs | Only shines in demos |
| Model dependence | Provider-flexible | Locked to one vendor |
The right question is never “is the AI impressive?” — most of it is. It’s “what does this company own that a model and a weekend hackathon can’t replicate?” Companies with a clear answer are worth your time. The rest are riding a wave, and waves recede.