Two years ago, “we use AI” was a differentiator. Today it appears in nearly every acquisition target’s materials, from genuine machine learning products to a chatbot wired into a third-party API over a weekend. Both kinds of company may be worth buying. They are not the same company, and the difference is verifiable if you know what to ask for.
The short answer: AI due diligence verifies four things. What the product’s AI actually does, what it depends on (data, third-party models, key people), what it costs to run, and whether customers demonstrably use it. Everything else is the pitch.
The mistake is evaluating the claim. The work is verifying the machinery behind it. Here is what I would want to see.
Verify what the product actually does
Start with the workflow, not the demo. Ask where, exactly, AI sits in the product: which customer action triggers it, what it produces, and what happens when it is wrong or unavailable. A feature that drafts a suggestion a human reviews carries a different risk and value profile than one that acts on customers’ behalf.
Then ask who built it and what it is built on. There is a wide and legitimate spectrum between proprietary models trained on proprietary data and a thin interface over someone else’s API. Neither is automatically a problem. What matters is whether the company’s position on that spectrum matches its story, its pricing, and its claims about defensibility.
A useful test: ask the team to describe the last three meaningful improvements to the AI capability, who made them, and how long each took. Real capability leaves a trail of real work.
Verify what the AI depends on
Every AI capability rests on dependencies, and in an acquisition you inherit all of them. Three deserve direct attention.
The first is data. What data trains or grounds the system, what rights the company has to use it, and whether customer agreements support how it is being used. The exact legal questions belong with counsel, but the operational questions belong in diligence: where the data comes from, whose it is, and what happens to the product if a source withdraws.
The second is third-party models and providers. If the product depends on an external AI service, understand the commercial terms, the version and behavior-change exposure, and what the company would do if pricing, terms, or capabilities shifted. A product one API change away from a different personality is a dependency to price, not a detail to skip.
The third is people. If the AI capability lives in two engineers’ heads, the acquisition is partly an acqui-hire whether anyone calls it that. Know which individuals matter and what keeps them.
Verify what it costs to run
AI features have a cost curve that traditional software does not. Every customer query can carry a real, metered cost, and costs can scale with success.
Ask for the current monthly cost of operating the AI capability, how it is measured, and how it has moved as usage grew. Then ask the harder question: at the volumes in your investment case, what does the cost structure look like, and who has modeled it? You are not looking for a precise forecast. You are looking for evidence the company understands its own economics well enough that your model can trust theirs.
A target that cannot describe its AI operating costs has not finished building its business model, whatever the product looks like.
Verify what customers actually use
The strongest evidence in AI diligence is boring: usage. Which customers use the AI features, how often, in which workflows, and what happened when the features were unavailable? A capability customers rely on shows up in usage patterns, support tickets, renewal conversations, and references.
Ask to see adoption and retention for the AI features specifically, separated from the product overall. Then ask two or three customer references a direct question: which part of this product would you actually miss? The gap between the marketed capability and the answered question is the diligence finding.
A one-page AI verification worksheet
Complete this before treating any AI claim as part of the deal thesis.
Product reality. Where AI sits in the workflow, what it produces, who built it, and what it is built on.
Data foundation. What data the system uses, the company’s rights to it, and the exposure if a source changes or withdraws.
External dependencies. Third-party models and providers, their commercial terms, and the plan if they shift.
Key people. Who the capability depends on and the retention decision for each.
Operating economics. Current and projected cost of running the capability, and who owns the model.
Customer evidence. Usage, reliance, and reference answers specific to the AI features.
Frequently asked questions
What is AI due diligence in an acquisition?
It is the verification of what a target’s AI actually does, what it depends on, what it costs to operate, and whether customers rely on it, so the deal thesis rests on evidence rather than marketing.
How do you verify a company’s AI claims?
Ask where AI sits in the workflow, what data and providers it depends on, who built it, what it costs to run, and what usage and customer references show.
What are the risks of AI built on third-party models?
Pricing changes, terms changes, and behavior changes at the provider can all reach the product. Diligence should identify those dependencies and the company’s plan if one shifts.
Does an AI feature automatically increase a company’s value?
Not by itself. Value follows verified usage, defensible data and dependencies, and operating costs the company understands.
Price the machinery, not the adjective
AI is a real capability with real dependencies, real costs, and real evidence, and it deserves the same verification discipline as any other part of the technology stack. The targets that welcome these questions tend to be the ones with something underneath the claim.
CTO Input helps investors and acquirers verify technology and AI claims during diligence, in business language a deal team can use. If AI is load-bearing in your next deal’s thesis, run the worksheet before you run the model, and bring us in if you want the verification done properly.