AI/IT due diligence: how to assess a company's "AI" before you sign

Why AI needs its own track

Standard due diligence covers finance, legal, and technology. "AI" slips through all three, because a company can show a working demo, a few customer logos, and a deck full of charts while, underneath, running a thin wrapper on someone else's model with no durable advantage at all. The risk isn't that the AI doesn't work. The risk is that you're paying for a capability the company doesn't control.

An AI audit is a time-boxed review that answers one question for the board: is what the company calls AI an asset, or an operating cost dressed up as one. You run it alongside the rest of due diligence, in two to three weeks, and you finish with a risk map, not an essay. Separately: who can run AI due diligence.

Three layers you have to separate

Most misleading assessments come from blending three things into the single word "AI."

• The model. Usually an off-the-shelf LLM from an outside provider. On its own it's not an advantage: competitors have the same model.
• The data. This is usually where the real value sits: proprietary, clean, well-labeled training data or a knowledge base no one can recreate overnight.
• The system around it. The pipeline, RAG over company documents, evaluation, guardrails, integrations. This shows whether the team can turn a model into a product.

A company with a strong data and system layer on a weaker model is usually safer than a company with the latest model and nothing underneath. You can swap the model in a week. You can't swap the data and the pipeline.

What to check, and what's a risk signal

This is the heart of the review. Each row explains what you're looking for and what should raise a flag.

What to check	Risk signal
Where the model comes from	Solely a third-party API, with no proprietary data layer and no exit plan
Provenance of the training data	No documented sources, consents, or licenses; data "from somewhere"
Whether evaluation exists	Quality judged "by feel," no fixed test set and no metrics
Cost per response	Margin tied to a provider's pricing the company doesn't negotiate
Dependence on a single provider	Full lock-in, no abstraction over the model, no alternative
Who built the system	All the knowledge in one person, or with an outside contractor
What happens on an error	No guardrails, no human in the loop on high-stakes decisions

If the company answers most of these questions with an architecture slide instead of evidence, the slide itself is the answer.

Five questions for management before you sign

The table above tells you what to look for. These questions help you draw it out in a conversation with management before a price is on the table. For each one we describe what a strong and a weak answer look like.

1. Which processes are genuinely automated? Strong answer: the process has a documented map, an owner and a before-and-after measurement, and automation is part of day-to-day work, not a demo for investors. Weak answer: a slide of vendor logos and a few scripts maintained by one person. Ask: "Walk us through one process end to end. Where does a person decide, where does an agent act, and who signs off the result?"
2. Does AI run inside a supervised architecture, or next to it? Strong answer: data classification, a human in the loop for sensitive decisions, logging of model calls and a clear line between production and experiments. Weak answer: API keys in a spreadsheet and customer data sent to a public model with no policy. Ask: "Where does the experiment end and production begin, and who formally approved it?"
3. Who owns governance? Strong answer: an inventory of AI systems with a risk level (in the spirit of the AI Act classification), a register of training and test data, a retention policy and an audit trail from request to response. Weak answer: no register, no owner, and "it's in our security policy" with no detail. Ask: "Which AI systems run today, who owns them, and how would you show a regulator that customer data is safe?"
4. Is the knowledge in the company, or in a few heads? Strong answer: prompts, evaluations and maintenance runbooks sit in the company's repository, with clear rights and licences. Weak answer: everything works because of one engineer, and there's no documentation. Ask: "If two specific people left tomorrow, what would stop working, and how long would it take to rebuild?"
5. What is the margin thesis built on? Strong answer: the thesis is broken down by process, with assumptions about the cost of running the model, a fallback model and a scenario for vendor price changes. Weak answer: "AI will increase margin", with no breakdown and a single vendor hard-coded into the architecture. Ask: "What happens to the margin if your main vendor doubles prices or changes the licence tomorrow? What's your plan B?"

Red flags in the conversation with management

• "AI-native" in the deck, Excel in operations. The slogans aren't backed by the process map or by the tools the team actually uses.
• Nobody owns AI. It's not about a "Head of AI" title. It's about a person with a budget, decision rights and responsibility for the systems register.
• Customer data in public models without classification. An early signal of regulatory and reputational risk, in any jurisdiction.
• One vendor, one contract, one person. Each on its own is a risk. All three at once is a concentration of risk that should affect the deal terms.
• A pilot running for more than a year. No decision to deploy or retire shows the organisation can't close experiments, and that usually comes back in the first year after the deal.

Data debt: the most expensive thing you buy

Training data is often the largest hidden liability. A model trained on data without documented consent, on scraped content, or on customer data used in breach of contract is not an asset but deferred legal risk, and it lands on the buyer's side after close.

Ask directly: where does each dataset come from, on what basis, and can it be proven in an inspection. Treat a non-answer the way you'd treat missing title to a property, because economically it's the same thing.

How to fit this into a board decision

The output of the review has to be board-readable: short and comparable. A simple format works.

1. A one-sentence verdict. Asset, wrapper, or asset with conditions.
2. The three biggest risks, with an estimated cost to fix or close each one.
3. What needs to happen in the first 100 days to turn risk into value.

Operator's rule: don't ask "does the company have AI." Ask "what happens to the margin if the model provider raises prices or changes terms tomorrow." The answer to that second question tells you what you're actually buying.

Where to start

The first step is cheap: ask for one concrete piece of quality evidence: a fixed test set, an evaluation result, a documented data source. A company that genuinely controls its AI will show it within a few days. A company with a wrapper will start explaining why it's complicated. That difference in reaction is often more telling than the deck itself.