When a target says it has AI, it is making a claim about revenue, cost and rights. A buyer’s finance team can test all three before anyone opens the code.
When a company for sale says it has AI, the claim tends to be made by the founder and checked by a technologist. Finance is then asked to model whatever premium is left over.
That is the wrong way round. If a company’s AI is worth paying for, it shows up in the numbers: which customers pay for it, what it costs to serve them, and who owns what the model learned from. Those are finance questions, and finance can answer them first.
My first job was in investment banking finance at Barclays, and I later worked in finance roles in wealth management and private equity. Today I run a UK accountancy firm for small businesses, and we look at small practices to buy. The rule I apply is simple: a claim is worth what you can verify.
Why the label is not enough
The AI label has carried a premium for years, which is why it gets stretched. In 2019 MIT Technology Review reported that MMC Ventures had looked at 2,830 European start-ups that claimed to use AI, and found that 40 per cent of them were not using any machine learning.
Regulators have since caught up with the marketing. In March 2024 the US Securities and Exchange Commission settled charges against two investment advisers, Delphia and Global Predictions, over false and misleading statements about their use of AI. The penalties came to $400,000 between them.
A wrapper is the subtler case. The AI is real, but it belongs to someone else: a general-purpose model, rented by the token, with a product built on top. That can be a perfectly good business. It is just not one that has earned a premium for its AI.
Test one: revenue that depends on the AI
Start with the switch-off question. If the AI feature disappeared tomorrow, which customers would cancel, which would ask for a discount, and which would not notice?
Ask for revenue by product and by customer, then set it against the usage logs for the AI feature. Revenue from customers who barely use it is revenue the AI did not win. Look as well for services dressed as product: set-up fees, managed services and staff doing work the pitch describes as automated.
Pilots and proofs of concept need their own line. They are evidence of interest, not of recurring revenue, and they do not belong in an annual recurring revenue figure.
Test two: who owns the data
A proprietary data moat is only proprietary if the target has the right to use the data. That right sits in three places: its contracts with customers, the law on personal data, and the terms of any model or dataset it licenses from someone else.
Read the customer contracts for what they allow on training and product improvement. Where personal data is involved, the UK’s Information Commissioner’s Office has published guidance on AI and data protection that sets out what it expects. A data room that cannot show the permission trail is showing you a risk, not a moat.
The downside is not theoretical. In 2021 the US Federal Trade Commission finalised an order requiring Everalbum, a photo app, to delete “the models and algorithms it developed by using the photos and videos uploaded by its users”. When the rights fail, the model can go with the data.
Then test whether the data actually compounds. Does each new customer make the product better for the next one? Could a competitor with the same base model buy or collect the same data? If the honest answers are no and yes, the moat is a word in the deck.
Test three: cost to serve
Software buyers are used to the next customer costing almost nothing to serve. AI breaks that habit, because every query carries a bill: model usage priced by the token, cloud compute, and often people checking the output.
Ask for twelve months of model provider and cloud invoices and map them to customers. Heavy users on flat fees are the obvious place to look. Ask, too, whether any of those costs are currently covered by free credits, and when the credits run out.
Falling model prices help, but less than the deck suggests. Stanford’s 2025 AI Index found that the inference cost of a system performing at the level of GPT-3.5 dropped more than 280-fold between November 2022 and October 2024. Every competitor got that price cut too, so cheap inference is a tailwind for the sector, not a moat for the target.
Test four: gross margin with the model costs in
Gross margin is where a wrapper shows itself. In 2020 Andreessen Horowitz wrote that AI companies often ran gross margins of 50 to 60 per cent, against 60 to 80 per cent or more for comparable software businesses, and named cloud costs and human support as the main reasons. The figures are six years old. The question they raise is not.
The test is where the costs are booked. Model usage and the staff who review its output belong in cost of sales. If they sit in research and development or general overheads, the reported gross margin flatters the business, and a multiple applied to it flatters it twice.
Then rebuild the margin customer by customer. A healthy blended figure can hide a handful of large accounts that cost more to serve than they pay.
Test five: the quality of the books
None of this works if the numbers underneath are wrong. Builder.ai, a UK start-up once valued at about $1.5bn, collapsed in 2025 after its lenders found its sales were far below what they had been told. The Financial Times reported, as summarised by Silicon UK, that lenders had been given a 2024 sales projection of $220m, and that after an internal audit actual sales came to about $55m.
With a small UK target, the public record is often thin. A company that qualifies as small can choose not to send its profit and loss account to Companies House, so the filed accounts may say almost nothing about trading. Ask for the management accounts, the trial balance and the bank statements, and reconcile them to each other yourself.
Look hard at anything the target expects to receive from HMRC. If the cash forecast includes a research and development tax credit, check the claim. When HMRC examined a random sample of claims made by small and medium-sized companies in 2021 to 2022, it found that around half were non-compliant in part or in full.
Filed does not mean right, in either direction. One client came to us having paid £18,000 in VAT when it should have been £6,000, on the work of their previous accountant. A buyer who had taken those figures on trust would have priced the wrong business.
Price what you have verified
Diligence rarely ends in a clean yes or no. More often it shows that part of the AI story is proven and part is promised, and the deal should be built around that split.
Buyers of small accountancy practices often defer part of the price and tie it to how many clients stay. The same logic suits an AI claim. Pay on completion for the revenue, margin and data rights you have verified, and tie the rest to an earnout or escrow measured on the AI revenue the business actually keeps.
The checklist
- Revenue. What share of revenue is at risk if the AI feature is switched off? Do the usage logs match the invoices? Are pilots kept out of recurring revenue?
- Data rights. Do customer contracts allow the data to be used for training? Is there a lawful basis for any personal data? Would a competitor struggle to get the same data?
- Cost to serve. What do model and cloud costs come to per customer? Which customers cost more than they pay? Which costs sit on credits that will end?
- Gross margin. Are model usage and human review in cost of sales? What is the margin by customer, not just blended?
- The books. Do the management accounts reconcile to the bank and to what was filed? Would any expected tax credit survive an HMRC enquiry?
None of these tests needs a data scientist. They need a finance team willing to treat AI as a claim about money and to check it like any other. If the AI is real, it is already in the numbers. The job is to go and look.
