AI startup due diligence: how investors check what is real / Decide

Decide

Questions for the management meeting of an AI startup

The management meeting of an AI startup should cover six topics, in this order: the AI claim, the codebase and its evals, provenance and security, margin and model dependency, data and the moat, and the team behind the agents. This page is the full list of thirty questions, grouped by topic, with what a good answer sounds like after each. It is ungated and printable. The questions assume the artefacts were requested before the meeting, so that the meeting tests the answers rather than discovering them; a company that has the artefacts answers each question in a sentence, and the sentence is what you are listening for.

Published September 17, 2026. Editorial.

Key takeaways

Ask for the artefacts before the meeting and use the meeting to test them; a good answer to every question below is one sentence and a pointer to a document already in the data room.
The question that matters most is 'show me the last eval run that failed and what you did about it', because it separates a company that measures quality from one that asserts it.
Three answers end the meeting early: a refusal to run a live input, a gross margin that nobody has rebuilt per customer, and a model inventory the team cannot produce.
Write the answer to each question against the claim table and the flags list, so the memo reads as evidence rather than impression.

The management meeting of an AI startup should cover six topics in order: the AI claim, the codebase and its evals, provenance and security, margin and model dependency, data and the moat, and the team behind the agents. Thirty questions follow, with what a good answer sounds like after each. Print it. The general questionnaire for any software target is on technical due diligence questionnaire; this list adds only what the AI changes.

How should the meeting be run?

Run the meeting after the artefacts have arrived, so that every question tests something already in hand rather than asking for it. Reveneau's diligence protocol sends the artefact request a week before the meeting, and the request is short: the eval suite with its last ten runs, twelve months of model-provider invoices with usage logs, revenue by customer, the model inventory, the code-generation tools and their terms, and the most recent security scan. The meeting is then an hour of listening for the sentence that matches the document.

Two rules. Ask the question and stop talking; the silence after a question about a failing eval run is itself an answer. And write down the answer as given, next to the line in the claim table it supports or contradicts, because the memo is built from the table and the table is built here.

What should you ask about the AI claim?

Ask about the claim first, because everything else assumes the product does what the deck says.

Which sentences in the deck describe what the AI does, and which of them have you tested? Good answer: names the sentences, names the eval cases that cover each, admits which ones are untested. The claim table on what changes when the target is an AI company is the frame.
Will you process an input we supply, now, with the trace visible? Good answer: yes, and the trace shows the model name, the version, the retrieval, the tool calls and the timing. A refusal ends the topic; see how to verify an AI claim with a test.
What does a person do in the loop today, and for what share of requests? Good answer: a number, from the logs, with the trend. The SEC's January 2025 order against Presto Automation found the vast majority of orders through its own AI version needed human intervention, against a claim that the product removed the need [1].
Which parts of the system are rules, and which are the model? Good answer: a diagram with both, and a count of the post-processing rules. "It is all the model" is a flag.
What is the product without the model? Good answer: a list of integrations, workflow, data and controls that would remain. "A text box" is a wrapper; see AI feature or AI wrapper.

What should you ask about the codebase and the evals?

Ask about the evals before the code, because the suite tells you whether the code has ever been measured.

Show me the last eval run that failed, and what you did about it. Good answer: a dated run, the change that caused it, the fix, and the time it took. This is the most important question on the list. A suite that has never failed has never been a gate; the eval suite as a diligence artefact explains why.
How many cases, how are they graded, and what is the pass threshold? Good answer: hundreds, mostly code-graded with model grading by a different model for judgement calls, and a threshold in the pipeline. Anthropic's developer guidance recommends volume with automated grading over fewer hand-graded cases [2].
Where does the specification live, and is it older than the code? Good answer: a versioned document with acceptance criteria that the evals were derived from. "In the prompts" is a flag; see diligence on an AI-written codebase.
What share of the codebase was generated, and who directs the agents? Good answer: a share, a named person or two, and a walkthrough of one recent change from spec to merge.
Which eval cases cover the accuracy figure in the deck? Good answer: the case IDs and the last score. No cases means the figure is untested.
What is not covered by the suite, and who decided? Good answer: a list of known gaps with owners. "It is complete" means nobody has looked.

What should you ask about provenance and security?

Ask about provenance and security together, because both are questions about what flowed where under which terms.

Which tools generated code, from when, under which account type and terms? Good answer: a record by tool and date. Anthropic's commercial terms, for instance, state the customer owns its outputs and that Anthropic may not train on customer content, and consumer terms differ [3]; the answer should name which applied when.
Was public-code matching on, and what did the logs show? Good answer: the setting's history and the licences found. GitHub's Copilot documentation describes logging the matching file and its licence when a match is accepted [4]; see code provenance and licence exposure.
When was the last security scan, and what did you do with the findings? Good answer: last quarter, findings triaged, fixes merged, report available. Veracode's July 2025 test found 45 percent of AI-generated samples failed security checks [5], which is why "we will run one" is a flag; see security of AI-written code in diligence.
What untrusted text reaches the model, and what can the model do as a result? Good answer: a list of input sources, a list of tools the model can call, and the injection cases in the eval suite.
What customer data goes to which provider, under which terms, with what retention? Good answer: a data-flow diagram with the provider, the account type and the retention on each hop.

What should you ask about margin and model dependency?

Ask about the margin with the invoices open, because the answers are checkable on the spot.

What is inference cost per customer per month, and which customers are negative? Good answer: the table, sorted, with the bottom rows named and the renewal plan for each. Bessemer's August 2025 data puts the fastest-growing AI companies at 25 percent gross margin and often negative [6]; see rebuilding inference gross margin from invoices.
Does priced usage reconcile to the invoices? Good answer: yes, within a stated tolerance, with the negotiated discount explained.
Which requests are cached, which are batched, and why not the rest? Good answer: shares from the logs and a reason for each. Both major providers list cache reads at a tenth of the input price and batch at half price [7].
What is the cap on model calls per user, and where is it in the code? Good answer: a number and a file. No cap is both a margin and a security flag.
Which models and versions are in production, and what is the retirement date for each? Good answer: the inventory, with the providers' published rules against each entry. Anthropic states at least 60 days' notice and retired Claude Sonnet 4 and Opus 4 on 15 June 2026; OpenAI states at least 6 months for generally available models [8] [9]; see model dependency risk.
What happened on your last forced migration? Good answer: a date, a duration and what broke. "We have not had one" from a company more than a year old on either major provider's models is unlikely.
Will you run the eval suite on your second-choice model now? Good answer: yes, and it finishes in the session with scores you can read. ICONIQ's July 2026 survey found companies running 3.3 models on average [10]; a company that cannot swap is behind its own customers.

What should you ask about the data and the moat?

Ask about the data last among the business questions, because the answers depend on the eval suite from earlier.

Show me product quality against data volume from your own eval runs. Good answer: a chart that rises across the range the company operates in. A flat chart means an asset without a moat; see data moat claims.
What would this data cost a competitor starting today, in money and in time? Good answer: a figure and a reason. a16z's 2019 analysis argued the marginal unit of data is worth less and costs more as the corpus grows [11]; the good answer engages with that.
Show me one dated example of product use changing a later release. Good answer: a release note, the data that drove it, and the eval run that measured the change.
Which parts of the data loop does the model provider see? Good answer: a hop-by-hop list, and an argument for why the provider could not rebuild the loop from it.

What should you ask about the team?

Ask about the team in terms of the roles the agents created.

Who writes the specifications and who writes the evals, and are they the same person? Good answer: two names, and a reason they differ.
Give me an example of an agent decision that was wrong and how it was caught. Good answer: a specific incident, caught by the suite or by review, with the fix.
What happens to the codebase if the person who directs the agents leaves next month? Good answer: a second name, a documented process, and a spec the next person could generate from.

The three answers that end the meeting

Three answers mean the rest of the diligence has to be priced as if the claims are unsupported. A refusal to process a live input: every capability claim is untested. A gross margin that exists only as a blended figure: the plan's cost of goods is unknown, and Bessemer's finding that the fastest growers are often negative is the base rate [6]. A model inventory the team cannot produce: the supplier risk is unmanaged and the migration is a future project with no date.

Each of the three leaves the company standing and ends the version of the deal that was priced before the meeting. The SEC's March 2024 press release on Delphia and Global Predictions, and the FTC's five Operation AI Comply actions of September 2024, are the reminder that the first of the three is also the one a regulator may ask about later [12] [13].

Write the thirty answers against the claim table and the flags list from AI startup diligence red flags by stage, and the memo is most of the way written. If you would rather have this run for you on a live deal, technical diligence for VC and private equity describes the engagement, and the AI startup due diligence guide has the method behind every question above. If you are the founder about to sit on the other side of the table, how to prepare for technical due diligence is where to start.

Best for

A deal team with a management meeting scheduled and artefacts already requested
An investment committee member who wants the questions and the good answers on one page
A founder preparing for the meeting

Avoid if

The artefacts have not been requested yet; send the request first and use the meeting to test them
The target's AI is internal tooling; use the general questionnaire

Verify before you commit

Ask question 6 and record the answer word for word
Have the invoices and the per-customer table open during questions 17 to 20
Run question 2 and question 23 live in the session rather than accepting a description

Common questions

What is the single most important question to ask an AI startup?

The single most important question to ask an AI startup is 'show me the last eval run that failed, and what you did about it', because the answer separates a company that measures product quality from one that asserts it. A good answer is a dated run, the change that caused it, the fix and the time it took. Anthropic's developer guidance describes suites with many automatically graded cases and a threshold; a suite built that way has a failure history, and a suite with none has never gated a release.

What should I ask an AI startup before the management meeting?

Before the management meeting, ask an AI startup in writing for its eval suite with the last ten runs, twelve months of model-provider invoices with usage logs, revenue by customer, the model inventory with versions, the code-generation tools with their terms, and the most recent security scan. The meeting then tests documents rather than discovering them. Veracode's July 2025 finding that 45 percent of AI-generated code samples failed security tests is why the scan is on the list.

Which questions should I ask about an AI startup's gross margin?

Ask an AI startup for inference cost per customer per month with the negative rows named, whether priced usage reconciles to the invoices, what share of requests is cached or batched, and where the cap on model calls per user sits in the code. Bessemer's August 2025 data gives the fastest-growing AI companies a gross margin of 25 percent, often negative, so the bottom rows of the per-customer table are the finding. Have the invoices open while asking; the answers are checkable on the spot.

What questions reveal model dependency risk?

Three questions reveal model dependency risk: which models and versions are in production with a retirement date against each, what happened on the last forced migration, and whether the team will run its eval suite on its second-choice model in the session. Anthropic states at least 60 days' notice and retired Claude Sonnet 4 and Opus 4 on 15 June 2026; OpenAI states at least 6 months for generally available models. A company that cannot swap in an afternoon has an unmanaged supplier.

How do I ask about AI washing without accusing the founders?

Ask about AI washing as a request rather than an accusation: 'Will you process an input we supply now, with the trace visible?' and 'What does a person do in the loop today, and for what share of requests?' A founder with a real product welcomes both. The SEC's January 2025 Presto Automation order found that the vast majority of orders through its own AI version needed human intervention, which is exactly what the second question surfaces from the logs rather than from a debate.

What answers should end an AI startup management meeting early?

Three answers should end the meeting early, in the sense that the deal must be repriced as if the claims are unsupported: a refusal to process a live input, a gross margin that exists only as a blended figure, and a model inventory the team cannot produce. The FTC's September 2024 Operation AI Comply and the SEC's March 2024 Delphia and Global Predictions settlements are the reminder that the first of those three is also what a regulator may later ask the investor about.

What should I ask about the people who direct the coding agents?

Ask who writes the specifications and who writes the evals and whether they are the same person, for an example of an agent decision that was wrong and how it was caught, and what happens to the codebase if the person who directs the agents leaves next month. A generated codebase looks like a team built it, so the bus factor has to be asked rather than read. Anthropic's guidance on test cases describes what a suite should contain; a second name who could maintain that suite is the answer.

How do I check a data moat claim in the meeting?

Check a data moat claim in the meeting by asking for product quality plotted against data volume from the company's own eval runs, what the data would cost a competitor in money and time, one dated example of product use changing a later release, and which parts of the loop the model provider sees. a16z's 2019 analysis by Casado and Lauten argued that the marginal unit of data is worth less and costs more as the corpus grows, so a good answer engages with where the curve flattens.

Should I ask about the code-generation tools' terms?

Yes, ask which tools generated code, from when, under which account type and terms, and whether public-code matching was on. Anthropic's commercial terms state that the customer owns its outputs and that Anthropic may not train on customer content from the services, while consumer terms differ, so the dates matter. GitHub's Copilot documentation describes logging the matching file and licence when a suggestion matching public code is accepted; the logs are the evidence for the second half of the question.

Can I use this question list if I am the founder?

Yes, a founder should use this list to prepare, because every question has a one-sentence good answer and a document behind it, and a company that has both walks through the meeting in an hour. Prepare the eval failure history, the per-customer margin table, the model inventory with retirement dates, the provenance record and the scan before the request arrives. ICONIQ's July 2026 survey found companies running 3.3 models on average; being able to demonstrate a swap is now ordinary practice.

References

More in Decide

AI startup diligence red flags, by stage

AI startup diligence red flags are stage-specific because the same fact means different things at different cheque sizes: a missing eval suite is normal at pre-seed and disqualifying at growth, and a claim with no test behind it is the flag at every stage. This page lists only the flags that exist because the target is an AI company. The general red flags, a bus factor of one, no tests, a rewrite in progress, an unpatched dependency, live on the technical due diligence guide; this page adds to that list. Each flag below names the stage it matters at, what it looks like in the evidence, and what to do about it.

Previous: AI startup diligence red flags, by stage Back to the guide