Tous les articles

20 questions to ask an AI startup before you invest

If you only get two hours with an AI startup before an investment decision, the questions you ask matter more than the data room. These are the twenty AI due diligence questions we ask in fund engagements, grouped by theme. For the ones that decide deals, we have added what a good answer sounds like, because the questions are easy to find and the calibration is not.

How to use these AI due diligence questions

One principle above all: ask for evidence rather than descriptions. Every question below has a version that ends in "show me", and that is the version to use. The grouping matters too, because answers in one theme calibrate the others. A team that is fuzzy on economics but precise on evaluation is disorganized; a team that is fuzzy everywhere is managing you.

Product and production

  1. What share of your product’s value is delivered by AI in production today, as opposed to in the demo?
  2. Can you show me a trace of one real production request, end to end?
  3. What happens when the model gives a wrong answer, and where do those cases go?
  4. When did you last change your prompts or your model, and how did you know quality held?
  5. How many users touch the AI features weekly, and what is retention on those features specifically?
  6. Which items on the roadmap are engineering, and which are research?

Question 4 is the sharpest of the six. A good answer names an eval set, its size, and a score that moved: "we run 800 cases on every change, the last model swap dropped extraction accuracy four points and we held the release." An evasive answer describes a philosophy of quality. Teams with real evaluation practice answer from a dashboard; teams without one answer from conviction.

Data

  1. What data do you train or fine-tune on, and what do the contracts behind it permit?
  2. If a customer demands deletion, does deletion reach backups, embeddings, and fine-tuned models?
  3. What personal data ends up in prompts and logs, and under what legal basis?
  4. Your enterprise contracts promise data isolation: can you show me where the architecture enforces it?

The stakes here are legal more than technical. A confident answer to question 7 cites contracts; a worrying one cites common practice. We wrote up the details, including what clean looks like, in training data rights, the diligence question with litigation attached.

Economics

  1. What does one typical request fully cost, including retries, context, and failed attempts?
  2. What is your gross margin today, and what is it in the model at ten times the volume?
  3. What happens to that margin if your model provider doubles prices, or halves them?
  4. What did you spend on inference last month, and who looks at that number weekly?
  5. Which single customer or workload drives the most model spend, and is that priced in?

Good teams answer 11 and 12 in minutes, because margins are the question every AI deal should start with and they track them internally too. If the answer to 14 is "finance sees it quarterly", the company does not yet run AI as a business.

Team

  1. Who can rebuild the pipeline from scratch if your best engineer leaves tomorrow?
  2. Who has changed the prompts and evals in the last three months?
  3. What would you lose, concretely, if your top two technical people left in the same month?
  4. Which parts of the system does the founding team no longer understand?
  5. Who gets paged at 2am when the AI misbehaves, and how often has that happened this quarter?

A good answer to 18 is specific and a little uncomfortable. An evasive one is "we document everything". Nobody documents everything, and in AI systems the undocumented part is usually the judgment that makes the product work. The mapping exercise behind these questions is covered in key person risk.

How to read the answers

Across all twenty, the pattern to watch is the same. Strong teams answer with numbers, artifacts and named owners, and they volunteer the unflattering parts: the incident, the regression, the module nobody likes. Weak teams answer with adjectives. An evasive answer rarely contains a lie. What it contains is a gap, and finding gaps is what diligence is for.

You will not get through all twenty in one meeting, and you do not need to. Pick the eight that touch your thesis hardest, ask the "show me" version, and follow the discomfort. If you need the full review list rather than a meeting script, start from the AI due diligence checklist we use on live deals.

When the answers raise more questions than they settle, that is usually the moment to bring in someone who can read the system directly. Our AI due diligence work starts where the question list stops: in the repos, the logs and the eval sets, checking that the good answers are also true.