AI due diligence for private equity: what to test before you sign

AI due diligence, done properly, has three jobs. It should tell the deal team how AI changes the target's market, whether the AI the target sells or depends on is real, and what a new owner can make from AI, costed and in the model. Most reports I read do the second job, call it done, and leave the first and third to guesswork.

That is backwards. The claims test is the easiest part and it rarely moves the price. Exposure and opportunity do. A target can have no AI at all and still carry more AI risk than one that sells it.

A target can have no AI at all and still carry more AI risk than one that sells it.

Search for "AI due diligence" and half of what comes back is software that reads a data room faster. That is a real category, and a useful one. The Bain & Company and StepStone Group 2026 GP Outlook found that GPs report their highest returns from generative AI in deal sourcing and due diligence itself. Funds using these tools are getting quicker at reading contracts and summarising management accounts.

This piece is about something else: diligence of the target's AI, and of the target's exposure to AI. The tools question is about your own process and your own cost base. The target question is about what you are buying, and what it will be worth at exit. A fund can have excellent diligence tooling and still buy a business whose pricing model AI is about to take apart.

Start with the market, before anyone mentions the target's product. The question is how AI changes demand, pricing power and the moat over a normal hold period, and it applies to every deal, including the ones with no AI story at all.

Two patterns show up most often in the lower mid-market.

The first is seat-based software. Early 2026 brought a sharp sell-off in listed software stocks, driven by fear that AI agents undercut per-seat SaaS pricing, and J.P. Morgan Asset Management pointed out the knock-on for private equity and private credit, where software is one of the largest sector exposures. If public investors now discount seat-based revenue, the exit buyer for your software asset will too.

Picture a vertical software business with £12m of ARR, 60% of it billed per seat to customer service teams. That is £7.2m tied to the headcount of the target's customers. If AI lets those customers run their service teams with 25% fewer people over three years, and seats follow people, £1.8m of ARR is under pressure. That is 15% of the revenue base, before a single competitor does anything clever. The numbers are illustrative. The arithmetic is not optional.

The second pattern is services businesses that sell hours. Accountancy, legal support, marketing services, recruitment, IT support, engineering consultancy. If the target bills by the hour, and AI compresses the hours needed to do the work, either the client captures the saving through lower fees or the target captures it through better margin. Which one happens depends on contract structure, competitive intensity and how sophisticated the clients are. Diligence should give a view on which way it goes, per service line, with reasons.

Exposure work is mostly commercial reasoning with technical knowledge underneath. You need someone who knows what current models can actually do in that workflow today, as opposed to what a vendor demo implies.

If the target sells AI, or says its operations depend on it, you are testing three things: is it real, is it defensible, and can the business afford to run it at scale.

Real is the part most reports cover. Is the "proprietary model" a prompt sitting on top of a third-party API? Is the "automation" a team of people in another time zone? I've written separately about AI washing in due diligence and how to spot it, so I won't repeat that here.

Defensible is harder. If the feature is a well-made wrapper around a general-purpose model, a competitor can ship the same thing in a quarter, and so can the model provider. The defensible parts are usually the boring ones: proprietary data the target has the rights to use, deep integration into a customer workflow, or a distribution position. Check the rights. Customer contracts often say nothing about using customer data to improve models, or say something the product team has quietly ignored.

Affordable is where I find the most surprises. AI features have a running cost that scales with usage, which most SaaS finance teams are not used to modelling. Take a target charging £400 a month for an AI assistant add-on. At pilot usage, the model costs might be £30 per customer per month. A heavy user running ten times the volume costs £300, and the gross margin on that customer drops from over 90% to 25%. Ask for twelve months of AI provider invoices and plot cost per customer against usage. If nobody in the business has done this, that is a finding in itself.

Regulation belongs here too, if the target sells into the EU. Under the EU AI Act, prohibited practices applied from February 2025 and transparency duties under Article 50 from August 2026. The Digital Omnibus deferred high-risk obligations for standalone systems to 2 December 2027 and for AI in regulated products to 2 August 2028, as Gibson Dunn sets out. Deferred is not cancelled. If the target's product touches hiring, credit or similar decisions, those deadlines sit inside a normal hold period. UK GDPR applies regardless.

The opportunity case is the one that gets waved through with a slide. "AI will drive efficiencies across the back office." Nobody can underwrite that, and the IC knows it.

Scepticism is reasonable. The same Bain and StepStone outlook found nearly 40% of GPs do not expect material financial impact from AI in their portfolio companies in 2026. That is the reason to cost the case properly, not a reason to skip it.

Take a £25m services business with 180 staff, 30 of them in finance, operations and client administration. Each works about 140 hours a month, so that is 4,200 hours. Walk the processes with the ops leads and say 30% of those hours sit in work current tools can handle well: invoice matching, onboarding paperwork, report assembly, triage. That is 1,260 hours. Assume you capture 40% of it in year one, because change takes time and some of it won't work. Around 500 hours a month returned to the business. At a loaded cost of £35 an hour, roughly £210k a year, against perhaps £60k to build and £2k a month to run.

That is a modest number. It's also one a CFO will sign up to, with named processes, a baseline and a sequence. Build conservative, base and stretch cases and let the deal team choose which one goes in the model. The longer version of this argument is in AI due diligence as a value-creation plan.

On a lower-mid-market deal you have weeks, not months, and management is already stretched by financial and legal diligence. The scope has to be tight.

Four or five sessions of 60 to 90 minutes are usually enough. The CTO or head of engineering, to walk the architecture and the AI stack live, with a screen share rather than slides. The product lead, on roadmap, pricing and what customers actually use. Two or three operations leads, on where hours go. A short session with the CFO on AI costs and how they are accounted for. If the target sells AI, ask to see the product used on real data by someone who is not the founder.

In one week you can deliver a red, amber and green view on exposure and claims, good enough to kill a deal or confirm there's nothing alarming. Two weeks adds hands-on testing and a proper cost-to-serve analysis. Three weeks adds the costed opportunity case, sequenced into the first hundred days. The fixed quote for our AI due diligence work comes back within 48 hours, which matters when the clock is already running.

The IC does not need forty pages. It needs a one-page summary with a rating per area, the reason in a sentence, and what it means for the deal.

Behind the summary sit the workings: the arithmetic, the invoices, the interview notes. Anyone on the IC should be able to follow a rating back to its evidence.

Each job tends to land in a different part of the term sheet.

Exposure findings move price. If 15% of ARR is tied to a seat model under pressure, the growth assumption in the model should reflect it, and so should the exit multiple. Sometimes the right response is a lower price. Sometimes it's a repricing programme in the value-creation plan, with the cost of doing it counted.

Claims findings usually move structure. An earn-out tied to AI product revenue puts the vendor's confidence to work. A specific indemnity can cover data rights the contracts don't clearly grant, since a known issue is unlikely to be covered by W&I insurance. Conditions can go into the 100-day plan: cost-to-serve monitoring in place by day 30, contract templates updated by day 60.

Opportunity findings change the equity story. A costed, sequenced AI case gives the IC a reason to back the deal and gives the exit buyer evidence rather than adjectives. Bain's 2025 Global Private Equity Report found nearly 20% of portfolio companies had operationalised generative AI use cases with concrete results. Being able to show that at exit, with numbers, is worth planning for at entry.

The answers matter less than how quickly and specifically they come back. A management team that knows its cost per AI query to the penny is a different investment from one that says it will "look into it".

Good AI due diligence gives the IC three things it can act on: a view on the market, a tested view on the claims, and a number it can put in the model. If you have a deal in exclusivity and want that done properly in the time you have, the free first call is a good place to start.

Brannan Coady

Founder of Anyway Ventures. Exited founder, former CEO of YourParkingSpace, investor and board member; now building AI inside PE-backed and mid-market businesses. More →