What using AI should mean at pre-seed

"We use AI" appears in most pre-seed decks now, and in most of them it means one of three things: the product calls a model owned by a large vendor, the founder used AI tools to write the code, or both. None of those is a problem. The problem is when the angel cannot tell which it is, and the founder has never checked whether the AI part works. This page is the ten-minute version of the check. The fuller method, for a seed or Series A deal where the AI is the company, is in the AI startup due diligence guide.

What should the claim mean?

At pre-seed, "we use AI" should mean that the founder can answer four questions, on screen, in plain words.

1. Which model? A name. The founder should know which model the product calls, and whether it is one model or several.
2. Whose model? Almost always a vendor's. That is fine, and it means the product's AI capability is rented. Ask what the product does that the vendor's own product does not, because that is where the company's value has to live. AI feature or AI wrapper is the page for that question on a bigger deal.
3. What happens when it is wrong? Every model is wrong some of the time. The founder should be able to describe what the user sees when it is: a source they can check, a confidence note, a human review step, a way to correct it. A product with no answer to this has not been used by anyone who noticed.
4. How often is it wrong? A number, from a test the founder ran, that they can show on screen. It does not have to be a large test. Twenty examples with a pass or fail against each is a test. "It is accurate" is a claim.

A founder who answers all four has an AI product. A founder who answers the first two has a vendor integration. A founder who answers none has a line in a deck. All three can be investable; they should be priced as what they are, and the deck should describe the one the company actually has.

Reveneau builds software only after an evaluation suite has been written from the specification, so its angel review asks every AI startup for the same thing: the test the founder ran, the number it produced, and the cases it failed. A founder who has that file has done the work regardless of what the model is.

Why does the claim need a check at all?

Because the claim is cheap and the regulators have started charging for it.

On 18 March 2024 the SEC announced charges against two investment advisers over statements about their use of artificial intelligence. Delphia had said it used AI and machine learning on client data to make predictions, and did not have the capability it described; it paid $225,000. Global Predictions had called itself the "first regulated AI financial advisor" and made claims about "expert AI-driven forecasts" it could not support; it paid $175,000. The SEC chair's statement named the practice "AI washing".

On 25 September 2024 the FTC announced five actions under "Operation AI Comply". One was against DoNotPay, which had marketed itself as "the world's first robot lawyer" and said it would replace the legal industry with AI. The FTC's complaint said the company had not tested whether its output matched a lawyer's, and did not employ any attorneys. DoNotPay agreed to pay $193,000 and to tell subscribers about the limits of its service.

Both regulators found the same thing: a company describing AI it had not built or had not tested. An angel's four questions are the same check, run before the cheque rather than after the complaint.

The claim is also everywhere now. The ACA's 2026 Angel Funders Report, published on 13 July 2026, said nearly two-thirds of reporting angel groups had completed at least one AI-related investment in 2025. If two-thirds of groups are doing these deals, the check belongs in every group's standard process rather than with whichever member follows AI.

What does a good answer look like on screen?

Ask the founder to show the test. What you are hoping to see is a file, a spreadsheet, or a dashboard with rows in it. Each row is an input the product was given, what the model answered, and whether that answer was right. At the bottom is a number.

A founder who has this will also have the failures, and the failures are the useful part. Ask to see three. A founder who can say "here is one where it invented a date, so we now show the source next to every date" has a product that has been used and improved. That sentence is worth more than the number.

What you should not accept as the test: a demo of the AI feature working on an input the founder chose; a vendor's published accuracy figure for the model, which describes the model rather than the product; or a statement that "users have not complained". How to verify an AI claim with a test is the method for building one if the founder has none, and the eval-driven development guide explains what an evaluation suite is and why it matters more once AI writes the code.

What if the AI is in the code rather than the product?

Ask separately. "We use AI" often means the founder wrote the product with AI coding tools, which is normal at pre-seed and changes one thing: the security question.

Veracode's July 2025 GenAI Code Security Report tested more than 100 language models on coding tasks across Java, Python, C# and JavaScript and found that 45 percent of the generated code samples failed security tests. Java was the worst at 72 percent, followed by C# at 45, JavaScript at 43 and Python at 38. The report also found that newer and larger models did no better on security than older ones, even as they got better at producing code that runs.

You cannot check that yourself. You can ask whether anyone has run a security scanner over the code and what it found, which is the five-minute security block of the one-hour technical check. For the fuller treatment, security of AI-written code in diligence covers what a firm would look for, and the AI-generated code guide holds the arguments about when AI-written code is safe to ship.

What are the red flags specific to the AI claim?

Five, in the order they appear.

• The founder cannot name the model. This usually means a contractor built the AI part and the founder has not looked.
• The failure rate is a vendor's number. The vendor tested its model; nobody tested this product.
• There is no fallback. When the model is wrong, the user sees the wrong answer with nothing to check it against.
• The AI is the whole pitch and the product does nothing the vendor's own product does not.
• The claim changed between the deck and the call. "AI-powered" in the deck became "we plan to add AI" in the meeting.

The last one is the SEC and FTC gap in miniature: a capability described in the fundraising material that the product does not have on the day of the call. Technical red flags at pre-seed and seed puts these next to the non-AI flags, and AI startup diligence red flags by stage is the full list for when the AI is the company.

How does this fit the rest of the angel check?

As one ten-minute block of six. The AI claim is the third of the three technical risks that decide pre-seed deals, after the demo-versus-production gap and the single-developer risk, and the pillar page puts them in that order for a reason: a product that does not run for real users and cannot survive its founder's absence has a bigger problem than its AI claim. Check those first. If they pass, ask the four questions, ask to see the test, and write down the model, the number and whether you saw it.

If you are checking a software vendor's AI claim rather than a startup's, the questions are different, and questions that expose a fake AI-native claim is the page for that.