AI due diligence for private equity: what to check before close.

Standard technology diligence asks whether the systems are sound: architecture, security, scalability, technical debt, team. Those answers still matter, but they do not tell you whether AI changes the economics of the business.

AI diligence asks a different set of questions. Which work in this company is repetitive, rule-bound and expensive? Is the data needed to automate it owned, clean and usable? Is the AI that management describes running in production, or sitting in a demo? And could a competitor or a customer use the same tools to take margin away?

A company can pass technology diligence with a clean report and still have no credible route to AI value. The reverse is also true: an untidy mid-market business with a large back office is often the better AI opportunity.

1. Where does AI move EBITDA after close? Start from the P&L, not from the technology. Find the cost lines dominated by people doing document, data and review work: intake, processing, reconciliation, quality checks, reporting, first-line support. For each one, estimate the volume, the unit cost and how much of the work follows rules. A defensible answer names three to five use cases, gives a range for each, and shows the assumptions.

2. What is real in the target’s AI story? Ask to see it running. For every AI capability management claims, establish whether it is in production, how many customers or staff use it weekly, who owns it, what it costs to run, and what happens when the model is wrong. A feature that relies on one engineer and one vendor API key is a prototype, whatever the sales deck says.

3. What are the risks to price or fix? Four come up repeatedly. Data rights: does the company have the contractual right to use customer data in the way the plan assumes? Vendor dependence: how much of the product sits on one model provider, and on what terms? Unowned models and shadow tools: what are staff already using without approval? Regulation: does the product make decisions in an area where AI is, or soon will be, regulated?

4. What happens in the first 100 days? A finding without a plan is not much use to the operating team. Good diligence ends with a sequenced first wave: which use case goes first, who inside the company owns it, what has to be true about the data, and how the result will be measured.

Use this as a request list and an interview guide. It is deliberately short.

• Economics. Headcount and cost by function. Volumes for the main repetitive processes. Error and rework rates. Current use of offshore or outsourced labour.
• Data. Where operational data lives. Who owns it under customer contracts. How much is structured. Whether history is retained and for how long.
• Product. Every AI feature in market or on the roadmap. Usage figures. Which models and vendors sit underneath. Unit cost per transaction.
• People. Who leads AI and product. Whether engineers have shipped anything with current models. How much the wider staff already use AI tools.
• Governance. An AI policy, if one exists. Vendor contracts and data processing terms. How model output is reviewed before it reaches a customer.
• Exposure. Which customers could do this work themselves with AI. Which competitors already do. How pricing would hold if the unit cost of the service fell.

The work runs in two passes. The first is outside-in and can start before exclusivity: the product, public materials, job postings, customer reviews, the vendor stack and how comparable companies are using AI. This is usually enough to form a view on exposure and to draft the questions.

The second pass needs management access: process walk-throughs, a look at the data, the contracts, and a demonstration of anything described as AI. Two to three weeks is a realistic length for both passes on a mid-market target.

The output is a memo an investment committee can read in ten minutes. It should hold:

• A one-paragraph conclusion: does AI make this a better or worse deal, and by roughly how much.
• The value cases, each with a range, the assumptions and the dependency that could break it.
• A plain statement of what is real and what is claimed in the target’s AI story.
• The risks, sorted into price it, fix it in year one, or accept it.
• A 100-day plan the operating team can start on day one.

• AI revenue or savings in the plan with no named owner inside the company.
• A product described as proprietary AI that is a thin layer on one vendor’s model.
• Customer contracts that are silent on, or prohibit, the data use the plan relies on.
• A services business whose customers could automate the same work themselves.
• A management team that cannot say what its AI features cost to run.

None of these should end a deal by itself. Each one changes the price, the plan or the first hire.

AI diligence earns its place when the target’s cost base is mostly people and paperwork, when AI appears in the equity story, or when the business sells work that AI can now do. For an asset-heavy company with a small back office, a few pointed questions inside technology diligence are usually enough.