AI startup due diligence: a checklist for angels and operators

Putting money into an early-stage “AI startup” is one of the highest-risk things an investor can do — and one of the easiest to do badly, because the word “AI” short-circuits scrutiny. This checklist is a structured way to keep your scrutiny switched on. It’s aimed at angels and operators looking at private deals, but the questions work for anyone trying to tell a real AI business from a pitch deck.

This is educational, not investment advice (see the disclaimer). Use it to ask better questions, not as a score that makes the decision for you.

1. The data

AI lives or dies on data, so start here.

• Provenance. Where does the training data come from? Can they show you?
• Rights. Do they have the legal right to use it? Scraped or licensed? Unclear answers are a legal liability, not just a quality issue.
• Moat. Is the data proprietary and hard to replicate — a real data moat — or could any competitor with an API key reproduce the result? If it’s the latter, much of the “AI” isn’t defensible.

2. The model

• Real or wrapper? Is there genuine model work, or is the product a thin layer over someone else’s API rebranded as “our AI”? Both can be businesses — but they’re priced very differently.
• Performance. How is accuracy measured, on what test set, and will they show you the error rate? “It works great” is not a metric.
• Dependency risk. If they rely on a third-party model, what happens when that provider changes pricing, terms or capability? That’s a single point of failure on someone else’s roadmap.

3. The team

• Verifiable experience. Do the founders and technical leads have relevant, checkable backgrounds — not just freshly-minted “Head of AI” titles?
• Build vs. buy capability. Can they actually build what they claim, or are they assembling off-the-shelf parts? Neither is disqualifying; conflating the two in the pitch is a flag.

4. The moat and the market

• Defensibility. Beyond data, what stops a larger, better-funded company copying this in a quarter? Distribution, regulatory approval, switching costs?
• Real demand. Are customers paying, or are these unpaid pilots and letters of intent dressed up as traction?

5. The commercials

• Revenue quality. Recurring and paid, or one-off and free? Is “AI revenue” broken out or blended into vague totals?
• Burn and runway. AI compute is expensive. How fast is cash going out, and what does the next round assume?
• Terms. Lock-ups, liquidation preferences, dilution, SPV fees — the structure can quietly eat your return even if the company does well. We cover these in the hidden risks of AI startup investing.

The AI-washing red flags (quick scan)

Run the same checks as for public companies — they apply doubly to startups, where there’s no 10-K to keep anyone honest:

• “AI” with no named model, technique or dataset.
• A demo that’s always pre-recorded or quietly human-in-the-loop.
• Accuracy claims with no error rate or test set.
• A valuation that re-rated on an AI rebrand, not on the business.
• More AI on the homepage than in the product.

Our AI-washing guide goes deeper on each, with the CFA Institute framework behind them.

How to use this

Don’t treat the checklist as a pass/fail score. Treat it as a map of where to dig. A startup can miss several points and still be a great investment; what matters is that you know where it’s weak and are pricing that risk in deliberately. The failure mode isn’t backing an imperfect company — it’s backing one whose weaknesses you never checked because “AI” made you stop asking.

---

Educational content, not investment advice. Early-stage investing can result in total loss of capital. Do your own due diligence and seek professional advice where needed.