How to Evaluate an AI Startup Claim
Every startup claims an AI advantage; most don't survive scrutiny. A framework to evaluate AI startup claims: real moat, model dependency, unit economics, and genuine traction — for investors, buyers and partners.
Table of contents
Every company is an "AI company" now, and every pitch deck promises a defensible AI advantage. Most of those claims don't survive scrutiny. Whether you're an investor, a partner, or a buyer, you need a quick framework to separate a real AI business from a thin wrapper around someone else's model. Here's how to evaluate an AI startup claim.
Ask: what's actually defensible?
The core question is moat. Many AI startups are a nice interface on top of a foundation model anyone can call. That's not necessarily bad — but it's not, by itself, a durable advantage. Probe for what they have that competitors (and the model provider) can't easily copy:
- Proprietary data. Do they have unique, hard-to-get data that makes their AI better? Data is the most common real moat.
- Workflow integration. Are they embedded in a customer's process in a way that's painful to rip out?
- Distribution. Do they have a channel or customer base others can't cheaply reach?
If the answer to all three is "no," you're looking at a feature, not a company.
Test the model dependency
- Whose model are they using? If their product is mostly a prompt over a third-party model, ask what happens when that provider raises prices, changes terms, or ships the same feature natively.
- Could the model provider eat them? The most dangerous competitor is the platform they're built on.
- What's their plan if the underlying model changes? Good teams have an answer; thin wrappers don't.
Scrutinize the unit economics
AI has a cost structure SaaS didn't:
- Margins. Inference costs money on every query. Ask about gross margin — heavy model usage can quietly erode it.
- Cost per user/task versus what they charge. A product that loses money on usage doesn't scale; it bleeds faster.
- Pricing model. Flat pricing on a variable-cost product is a red flag unless they've solved the cost side.
Check product-market fit signals
- Retention and usage, not just signups. Are customers using it repeatedly and renewing?
- Willingness to pay for the AI specifically, versus using it because it's free/bundled.
- Real outcomes. Can they show the AI changing a metric the customer cares about?
A quick checklist
| Question | Green flag | Red flag |
|---|---|---|
| Moat? | Proprietary data / deep integration | "Better prompts" |
| Model dependency | Multi-model, a plan | Pure wrapper on one provider |
| Margins | Healthy, cost understood | Loses money per query |
| Traction | Strong retention + WTP | Signups only |
Who this is for
- Investors doing fast diligence on AI pitches.
- Enterprises choosing an AI vendor they'll depend on.
- Partners assessing whether a startup will still exist in two years.
Bottom line
Evaluate an AI startup the way you'd evaluate any business, then add the AI-specific tests: is there a real moat (data, integration, distribution), how dependent are they on someone else's model, do the unit economics survive inference costs, and is there genuine retention and willingness to pay? A great demo proves the model is good. None of those four answers being strong proves the company is.
