Buyers pay for AI that has already changed a number in the accounts: a lower cost per unit of delivery, a higher gross margin, revenue from an AI feature customers use. They discount pilots, roadmaps and "AI-powered" labels, often to zero. To build an AI equity story for exit that survives due diligence, you need production usage data, measured margin impact, clear rights to the data your AI runs on, and governance a buyer can check, and you need to start 12 to 24 months before a sale so the results show up in the trading history.
This is written for sponsors and portfolio CEOs who are a year or two from a process. I've been on the sell side of diligence as a founder and CEO, and on the buy side as an investor. The pattern hasn't changed much, but the AI questions have got sharper in the last year.
Hours saved that nobody can find in the P&L don't count.
How buyers value AI in M&A in 2026
They ask about it on nearly every deal, and they arrive sceptical. EY's Global PE Exit Readiness Study 2026, based on 100 PE executives and 100 executives from recently exited portfolio companies, found the share of GPs naming AI as a challenge in preparing companies for sale rose from 7% to 18% in a year. EY's reading is that buyers now probe not only how a company uses AI but how exposed it is to disruption from it, and that they distinguish AI activity from AI strategy. "A list of pilots or isolated productivity tools may not be enough to support valuation," the report says.
Grant Thornton's July 2026 article, Buyers are asking questions about AI, and it's impacting valuations, puts it more bluntly. Chen Liu, a partner at Grant Thornton Stax, is quoted: "Buyers aren't paying for AI potential." The same piece notes that AI diligence is now part of standard product and tech diligence on nearly all deals.
In practice AI reaches the price through three routes, and buyers treat each differently.
| Route into the price | What the seller usually claims | What a buyer credits |
|---|---|---|
| EBITDA | "AI will save £1.2m a year" | Savings already in the last twelve months, plus a run-rate adjustment only where the costs have demonstrably left |
| Growth | "Our AI product will add 20% to revenue" | Revenue from AI features customers already pay for, at retention rates the buyer can see |
| Multiple | "AI makes us a technology business" | Evidence the business is protected from AI disruption, or owns something AI makes more valuable |
The third route cuts both ways. A buyer who decides AI threatens your pricing will take a turn off the multiple whatever your AI features look like. That's the disruption question, and it's covered in the single-company AI disruption risk assessment. Sellers who prepare only the upside story are often blindsided by the downside one.
Why buyers discount AI claims
Because most of the AI claims they see don't survive contact with the data room. Sellers describe pilots as deployments, licences as adoption and forecasts as results. A buyer's tech adviser asks for usage logs and finds that 12 people in a 400-person business use the tool weekly. Every other claim in the information memorandum then gets read with the same suspicion.
Governance is the second reason. Grant Thornton's 2026 AI Impact Survey private equity findings found only 9% of private equity leaders very confident they could pass an AI governance audit within 90 days, against 22% across industries. If the sponsor can't evidence governance at fund level, a buyer won't assume it exists in the portfolio company.
The third reason is that buyers have been slow to diligence this properly and are catching up. KPMG's 2025 Technology M&A Survey found 66% of dealmakers discussed technical and AI debt during deal planning, but only 33% prioritised investigating it before the deal. That gap is closing, and sellers selling in 2027 should expect the buyer's adviser to go further than one selling in 2025 did.
My own view: discounting is the rational default. Absent evidence, a buyer has no way to tell your AI apart from the ten other "AI-enabled" businesses in their pipeline this quarter. The seller's job is to make the evidence so easy to check that the default doesn't apply.
What AI evidence buyers actually credit
Four kinds of evidence move a buyer. Most sellers have one, partly.
Production usage
Usage data from the systems themselves, not a survey of staff. Who uses each tool, how often, on what volume of work, over at least six months. For an AI feature in a product, the share of active customers using it and whether usage is rising. Give the buyer the raw export as well as the chart.
Measured margin or growth impact
A baseline taken before the change and a measurement after it, on the same definition. Cost per claim processed, hours per audit report, first-response time, quote turnaround, gross margin on the service line. The method is in measuring AI ROI. The test a buyer applies is simple: did the number move, and did cost leave the business or get redeployed into something that shows up elsewhere? Hours saved that nobody can find in the P&L don't count.
Data rights
If the story is that your data makes your AI better than a competitor's, the buyer will read your customer contracts to check you're allowed to use it. Many businesses hold years of client data with no right to reuse it for training or for building products. The Grant Thornton article highlights the questions buyers now ask when third-party models are used: whether proprietary data trains external models, and whether it is segregated from other customers. Answer both in writing, with the vendor terms attached.
Governance
An AI policy staff have actually been trained on, a register of every AI tool in use with what data goes into it and who approved it, a record of what happens when an output is wrong, and the business's position under the EU AI Act if it sells into Europe. None of this needs to be elaborate in a mid-market business. It needs to exist and to match what people do. The UK AI policy template is a starting point.
How to build an AI evidence pack for exit
The evidence pack is a folder in the data room that backs every AI sentence in the information memorandum. I'd build it in this order, with the timings counted back from launch of the process. The wider technology fixes on the same timeline are in our exit readiness checklist.
- 24 to 18 months out: baseline and choose. Score the business on the disruption question (our AI disruption risk scorecard sets out the method), take baselines on the three or four processes where AI could plausibly change the numbers, and pick two or three use cases to put into production. Pick ones where the result will show in the accounts within a year.
- 18 to 12 months out: ship and measure. Get the use cases live with real users, logging usage from day one. Fix the data rights in customer and supplier contracts as they renew. Write the AI policy and tool register if they don't exist.
- 12 to 6 months out: let the numbers accrue. Monthly reporting on each use case against its baseline. If savings mean roles aren't backfilled, make sure that shows in headcount and payroll so a quality of earnings provider can trace it.
- 6 to 3 months out: assemble and test. Build the pack, write the one-page AI section of the information memorandum, then have someone independent try to break it. This is the point at which a vendor technical due diligence report covers AI alongside the rest of the technology.
- At launch: make it easy. Raw data, not only summaries. A named person who can demo every AI system on live data and answer questions without a script.
A complete pack usually contains an AI inventory (every system, the model or vendor behind it, the data it uses, its owner), usage data per system, before-and-after measurements with the method written down, unit costs including inference and vendor fees, the relevant contract terms on data rights, the AI policy, training records and tool register, an incident log, and the written answer to the disruption question.
AI washing in due diligence describes how a buyer's adviser will test each claim. Run the same tests on yourself first.
Writing the AI equity story
The AI equity story should fit on one page. It should say what AI already does in the business, with the numbers, what it will do under the next owner and why that's credible, where the business could be exposed and what is being done about it, and how it's governed. Each statement points to a document in the pack.
Including the exposure paragraph feels counterintuitive to most sellers. Include it anyway. The buyer will ask the question regardless, and the version they write themselves will be less generous than yours. A seller who says "around 30% of our revenue is billed hourly for drafting work AI will compress; we have moved 18% of clients to fixed fees and are on track for 40% by the end of next year" has turned the risk into a management credential.
The forward-looking part is where buyers apply the heaviest discount, and EY's study reflects the difficulty: management teams find it hard to build a credible plan that covers both AI's risks and its opportunities. A plan that buyers credit has named use cases, costs, owners and a track record from the ones already delivered. "We will apply AI across operations" earns nothing.
AI exit readiness mistakes I see most often
The first is starting too late. Six months before a process is enough time to write a story, and not enough to generate the evidence for one. EY found 86% of GPs said exit preparation improved valuations, with the strongest results when it began 12 to 24 months before a sale.
The second is counting licences. A Microsoft Copilot rollout to 300 staff is a cost line. Unless usage and output data show it changed something, a buyer will treat it as one.
The third is claiming "proprietary AI" for a thin layer over a third-party model. There's nothing wrong with using third-party models, and buyers expect it. The problem is the word proprietary, which invites a test the business will fail.
The fourth is run-rate adjustments with nothing behind them. Adding back £800k of "AI savings" that haven't yet left the cost base will be challenged by the buyer's financial advisers, and losing that argument costs credibility on every other adjustment.
The fifth is ignoring governance until the buyer asks. Writing an AI policy during diligence is visible to everyone, and it reads as a business that hadn't thought about it.
The sixth is having one person who understands the AI. If the head of data leaves during the process, or simply can't attend the management presentation, the story goes with them.
What AI exit readiness costs
Our exit readiness baseline costs £15k to £25k and takes two to three weeks. It scores the business the way a buyer will, on the disruption question, the AI evidence and the wider technology, and produces a prioritised plan of what to fix and what to leave, timed to the planned process. Closer to launch, a vendor technology due diligence report that buyers can rely on costs £25k to £40k, depending on the size and complexity of the technology estate.
If a sale is 12 to 24 months away and AI is part of the story you want to tell, the exit readiness baseline is the place to start, and the first conversation is free.
Questions
Do AI cost savings count as EBITDA adjustments in a sale?
Only if they can be evidenced. Savings already in the last twelve months' accounts count without argument. A run-rate adjustment for savings delivered part-way through the year can be accepted if the before-and-after data is clean and the costs have actually left the business. Planned savings belong in the forecast, where buyers will discount them. Expect the buyer's quality of earnings provider to test every AI-related adjustment line by line.
Should the information memorandum include a section on AI?
Yes, as one plain page: where AI touches the business, what is in production, what it has changed in the numbers, where the business could be exposed and what management is doing about it. Every sentence should point to something in the data room. A longer AI chapter full of capability claims tends to invite more scrutiny than it repays.
What if our business has done very little with AI before a sale?
Say so plainly and show a credible, funded plan instead. Buyers can price a business that hasn't started more easily than one whose claims fall apart in diligence. With 12 months or more before a process, two or three production use cases with measured results are achievable and change the conversation. With less time, focus on the disruption question and the plan.
Do buyers care about AI governance in a mid-market business?
Increasingly, yes, because it decides whether they can trust the AI claims and whether they're buying a liability. They'll ask what tools are in use, what data goes into them, who approved them, whether customer data trains third-party models and what happens when outputs are wrong. A short AI policy, a tool register and evidence that both are followed answer most of it.
How early should a business start preparing its AI story for exit?
Twelve to 24 months before a sale. That gives enough time to get two or three AI use cases into production, collect six to twelve months of measured results, fix data rights and governance gaps, and let any savings show up in the accounts. EY's 2026 exit readiness research found the strongest outcomes came from preparation started in that window.