Founders and finance lead preparing a data room for an AI funding round

AI Startup Due Diligence: What Investors Ask Before They Fund

Investors now test model dependency, data rights, margins after inference costs, evaluation evidence and AI-washing risk. Here is how founders and CFOs can prepare.

By Shayan Salehi 8 min read

Image: ColdAI

When an AI company raises capital, investors now test six things: how much of the product depends on someone else's model, whether the company has the right to the data it uses, what gross margin is left after inference costs, whether the company can prove the product works, which regulations apply, and whether every claim in the deck is true. Founders and CFOs who assemble that evidence before the first meeting shorten diligence and protect the valuation.

Why the diligence list changed

Investors used to ask whether the technology worked and whether the market was large. Those questions remain, but the evidence they want has shifted. Models are available to every competitor, so a pitch that says "we use AI" proves little. What investors can examine is the layer around the model: the contracts, the data, the unit costs, the test results and the accuracy of the story.

Regulators have also given investors a reason to look harder. Since 2024, the SEC and the FTC have brought cases over false or unsupported AI claims, and those cases describe what regulators treat as misleading. A fundraising deck is a set of claims about a company, and the same discipline applies.

This article is written for founders and CFOs preparing a raise. It covers what an investor's team will ask, what evidence answers the question, and the mistakes that slow a round.

Dependency on third-party models

Most AI startups build on a model they do not own. That is normal. The diligence question is how much risk it creates and whether the company has described it accurately.

NIST's Generative AI Profile, NIST AI 600-1, lists value chain and component integration among its risk categories, and describes it as non-transparent or untraceable integration of upstream third-party components, including data that was improperly obtained. An investor reading that framework will ask what sits upstream of your product.

Prepare a dependency map that answers these questions:

  • Which models, APIs and datasets does each feature use, and who owns them?
  • What do the provider contracts say about pricing changes, rate limits, service levels, data use, termination and model retirement?
  • How long would it take, and what would it cost, to move a feature to another provider or to an open model, and has anyone tested it?
  • Does one provider account for most of your inference spend?
  • Does the pitch describe the third-party component accurately?

The last question has enforcement history behind it. On January 14, 2025, the SEC announced settled charges against Presto Automation over statements about its AI drive-thru product. The SEC said the company did not disclose that, for a time, a third party owned and operated the AI speech recognition in all deployed units. A founder who calls a licensed model proprietary invites the same problem.

Data rights and licensing

Investors ask where the training data, fine-tuning data and retrieval data came from, and whether the company may use it for this purpose. The answer must hold up under a lawyer's review.

The US Copyright Office has published a report on this subject. Part 3 of its report on Copyright and Artificial Intelligence, covering generative AI training, was released in pre-publication form in May 2025. The Office said a final version would follow without substantive changes expected in the analysis or conclusions. The report examines where copying can occur during training, how fair use applies, including the question of unlawful access, and licensing options. It does not settle your position, and courts are still working through these questions, but it shows the issues an investor's counsel will raise.

Build a data provenance register. For each dataset, record the source, the license or terms under which it was obtained, whether the terms allow commercial model training, any opt-out or deletion obligations, and who approved the use. Do the same for customer data. If your product learns from customers' inputs, the customer contracts must permit that, and the privacy notice must say so.

Gaps are common, and a disclosed gap with a remediation plan is better than one found by the investor's counsel. Replace or license questionable data before the raise if you can. If you cannot, say what the exposure is and how you would address it.

Unit economics after inference costs

Software investors have long looked at gross margin. AI products add a cost that scales with usage: every request consumes model capacity. Investors now ask for gross margin after inference, by product and by customer segment.

The FinOps Foundation's paper on FinOps for AI recommends measuring cost per successful outcome, reported as a range, rather than raw token cost, and tying unit economics to results the business owns. That is also a good template for the data room. A table that shows, for each workflow, the volume, the cost per attempt, the success rate and the cost per successful outcome answers most of an investor's questions in one page.

Consider a hypothetical startup that charges a flat monthly fee per seat. If its heaviest users send many times more requests than its average user, the average margin can look healthy while the top segment loses money. Show margin by usage band, and show what changes if the provider changes its pricing or if usage grows faster than revenue.

Include a plan for cost control: model routing to cheaper models where quality allows, caching, batching of non-urgent work, and pricing changes that tie revenue to usage. The paper's guidance to choose the least expensive model that meets the quality bar is the kind of discipline investors want to see.

Evaluation evidence

An investor cannot see inside the model, so they ask for test results. A demo shows the best case. An evaluation shows the distribution.

Evidence worth preparing includes a documented test set that reflects real customer inputs, results over time, the share of outputs that need human correction, the failure modes found and fixed, and the process that runs before each model or prompt change. Independent measurement carries more weight than vendor claims. A peer-reviewed study in the Journal of Empirical Legal Studies tested AI legal research tools whose providers had described them as avoiding hallucinations, and found that the tools hallucinated between 17 percent and 33 percent of the time. The study concerned one product category, but it shows why investors discount unmeasured claims.

The Presto case shows the same point from the enforcement side. The SEC said the company described its rate of orders completed without human intervention in a misleading way, and that the vast majority of drive-thru orders on one version still required human intervention. If humans correct your outputs, state how often, and include the cost of that labor in your margin.

Regulatory exposure

Start by working out your role. The EU AI Act sets obligations for providers of general-purpose AI models. The European Commission states that those obligations entered into application on 2 August 2025, that its enforcement powers, including fines, begin on 2 August 2026, and that providers of models released before August 2025 have until 2 August 2027. Providers must publish a public summary of training content using the Commission's template. If you build on someone else's model, those duties mainly sit with your provider, but you should know which of your features would make you a provider yourself, and ask your counsel.

Investors will also ask about sector rules that attach to your use case, privacy law for the data you process, and any customer contractual terms on AI use. List them, name the owner for each, and keep the evidence of compliance in the data room. Confirm current dates with counsel before you rely on any of them.

AI washing enforcement

On March 18, 2024, the SEC announced settled charges against two investment advisers, Delphia and Global Predictions, over false and misleading statements about their use of AI. The SEC ordered civil penalties of $225,000 for Delphia and $175,000 for Global Predictions, $400,000 in total. It said Delphia overstated its use of AI and machine learning that incorporated client data, and that Global Predictions made false claims about its use of AI on its website and social media.

On September 25, 2024, the FTC announced Operation AI Comply, a set of five actions against companies that allegedly made deceptive claims about AI, including DoNotPay, which allegedly marketed an AI lawyer service that could not deliver on its claims. FTC Chair Lina Khan said at the time that "there is no AI exemption from the laws on the books."

These cases involved advisers, a public company and consumer-facing businesses, not private fundraises. They still show how regulators read AI claims: they check whether the technology exists, does what is claimed, and is what the company says it is. Investors and their counsel read decks the same way.

A claims review costs little. Take every statement about AI in the deck, website and data room, and attach evidence to each: a test result, a contract, a log, a customer reference you are allowed to use. Remove or soften any statement you cannot support, such as "proprietary," "autonomous," "fully automated," or an accuracy figure without a method.

Where to start

Work through this sequence over four to six weeks before you open the data room:

  1. Draw the dependency map and read every model and data contract for termination, pricing and data-use terms.
  2. Build the data provenance register and fix or disclose any gaps with counsel.
  3. Produce the unit-economics table by workflow and usage band, with inference cost and human-review cost included.
  4. Assemble the evaluation pack: test set, results, failure modes and change process.
  5. List the regulations that apply to you, your role under each, and the owner and evidence for each.
  6. Run the claims review, and have one person who did not write the deck check each AI statement against its evidence.

Treat the preparation as a rehearsal of the investor's diligence, run by someone who has not seen the pitch. The gaps it finds are cheaper to fix before a term sheet than during one.

Tags:ai-startup-due-diligencefundraisingai-washingdata-rightsinference-costsmodel-dependencyinvestor-readiness

Your starting point

What would you like to move forward?

Start with the outcome that feels closest. We'll help give the next step a useful shape.

Let's find your next move

One conversation. A clearer direction.

Tell us what you want to improve. A few sentences are enough to start.

shayan@coldai.org