AI assistance in buying a company

The question of whether to use AI in an acquisition has already been answered by the platforms. Flippa markets a system it calls BrokerAI that builds the information memorandum, scores inbound buyers, manages the data room and matches across a buyer pool it states is over two million, alongside document builders for letters of intent and asset purchase agreements with electronic signature. Acquire.com sells buyer membership from $390 with integrated legal, diligence and closing tools described as AI-powered.

So the useful question is narrower. Which parts of diligence can a model actually hold, which parts does it fail at in ways that are measurable, and which parts have to end with a human name on them. All three have real answers.

The tasks AI is genuinely strong at share one property: the answer is in the document in front of it. Nothing has to be recalled.

• Extraction. Pulling every change of control clause, notice period, exclusivity term and termination right out of a contract bundle and putting them in one table with the page reference
• Reconciliation. Flagging where the management accounts and the filed accounts disagree, and where a schedule does not sum to the balance sheet
• Coverage. Reading all four hundred documents instead of the forty a human had the hours for, which changes what gets found rather than how fast
• First pass summarisation. The public Vectara hallucination leaderboard puts error rates on tight summarisation as low as 1.8 per cent, precisely because the source text is supplied rather than remembered
• Drafting the information request list, which is a known and largely repeatable shape
• Turning a finding into a question you can actually put to the seller, which is a translation task rather than a judgement

Coverage is the one that changes outcomes most and gets discussed least. Roughly half of business sales collapse in diligence, and late disclosure is one of the named causes. Reading everything rather than sampling is exactly how late disclosure becomes early disclosure.

The failures are not folklore. They are benchmarked, and the numbers are worse than the marketing suggests. Researchers at Stanford RegLab and the Stanford Institute for Human-Centered AI tested the purpose-built legal research tools sold to law firms, in a study titled Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. These are retrieval-backed products marketed on their accuracy, not general chatbots.

Lexis+ AI and Ask Practical Law AI

Incorrect more than 17 per cent of the time

Westlaw AI-Assisted Research

Incorrect more than 34 per cent of the time

General purpose models, earlier Stanford study

Between 58 and 82 per cent on legal queries

Tight summarisation, Vectara leaderboard

As low as 1.8 per cent

The spread in that table is the actual lesson. The same underlying technology is nearly reliable when the answer is in front of it and badly unreliable when the answer has to be recalled and located. Diligence contains both kinds of task, and they look identical from the outside.

Stanford also names a failure mode that is more dangerous than invention, which they call misgrounding. The tool describes the law correctly but cites a source that does not support the claim. The citation is real, so a spot check finds a genuine document and moves on. Anyone verifying by confirming that the reference exists rather than by reading what it says will pass a misgrounded answer straight through.

There is a second failure mode that is almost purpose built to hurt an acquirer. Models tend to agree with the premise they are handed. The 2026 AI Index from the Stanford Institute for Human-Centered AI reports accuracy collapsing when a false statement is put to a model as established fact: GPT-4o falling from 98.2 per cent to 64.4 per cent, and DeepSeek R1 from over 90 per cent to 14.4 per cent. The same reports note that models handle a false statement much better when it is framed as somebody else believing it, rather than as the user asserting it.

Now consider how a buyer actually prompts during diligence. You arrive with a thesis. You believe the customer concentration is fine, or that the lease is assignable, or that the earnout structure is standard. You type that belief into the question. That is a false premise waiting to be confirmed by a system that is optimised to be helpful.

The consequences are no longer theoretical. The AI Hallucination Cases database maintained by Damien Charlotin tracked more than 1,700 judicial decisions worldwide involving AI-generated fabrications as of July 2026. Those are the ones that reached a courtroom and were caught.

The third failure is staleness, and it is worth showing with a live UK example rather than describing abstractly.

GOV.UK states the current position on Business Asset Disposal Relief without ambiguity: 18 per cent on qualifying gains disposed of from 6 April 2026, 14 per cent between 6 April 2025 and 5 April 2026, and 10 per cent on or before 5 April 2025. Meanwhile a widely read UK guide to selling a business, published on the Xero site in April 2026 and written by a qualified accountant, still states that qualifying gains are taxed at 10 per cent up to a lifetime limit of £1 million.

That guide is not unusual and it is not dishonest. It is one of thousands of pages describing a rate that has moved twice in two years, all of which are training data. Ask a model about the relief and it has read a great deal of writing that says 10 per cent, and comparatively little that says 18. A rate, a threshold or a filing deadline is the single worst class of fact to accept from a summary, because it is exactly the class most likely to have quietly changed.

The practical rule that follows is narrow and easy to hold. Use the model to tell you which register, statute or guidance page to read. Take the number itself from that page.

Some parts of an acquisition cannot be delegated to a system, not because the system is weak but because the obligation attaches to a person.

Identity is the clearest. Since 18 November 2025 Companies House has required directors and people with significant control to verify their identity, either directly through GOV.UK One Login or through an Authorised Corporate Service Provider, and an Authorised Corporate Service Provider has to be a firm carrying out anti-money laundering supervised activity, such as a solicitor, an accountant or a company formation agent. Supervision attaches to a supervised person. There is no version of that check a model performs.

The agreement is the second. English law runs on caveat emptor, and as LegalVision UK puts it, if defects come to light after completion the buyer generally has no recourse against the seller unless the purchase agreement says otherwise. Everything that protects you sits in negotiated warranties, indemnities and disclosure. A model can draft a clause and can compare it to a thousand others. It cannot stand behind it, and it carries no insurance.

Materiality is the third and the least discussed. A model can tell you reliably that a change of control clause exists in the largest supplier contract. Whether that clause is a negotiating point, a price adjustment or a reason to walk depends on your funding, your appetite, your alternative deal and how badly you want this one. That is not a retrieval problem. Nobody outside your head holds the inputs.

Then there are the people. The Transfer of Undertakings regulations move employees, their terms and their continuity of employment to a buyer automatically, and GOV.UK notes that the size of the business makes no difference to that. What the regulations cannot tell you is whether the two people who actually make the thing work intend to stay once you own it. That answer is not in the data room in any form, and it is frequently the largest single variable in the price.

Finally the number itself. Empire Flippers publishes achieved multiples that run from 14.4 times monthly profit for distressed businesses to 36.5 times for premium businesses above a million dollars. Where a specific business sits inside that spread is an argument you make, not a value you look up, and the person who makes it is the person who has to live with it.

• Give it the documents, not the question. Retrieval over your own data room beats recall from training on every measure that has been benchmarked
• Make it cite the page, then read the page. Confirming that a citation exists is not verification, because misgrounding produces real citations that do not support the claim
• Never state your thesis in the prompt. Ask what the contract says about assignment, not whether the contract permits assignment as you assume it does
• Take every date, rate, threshold and deadline from the primary source, and treat any of them appearing in a summary as a prompt to go and check
• Keep a written record of what was machine reviewed and what a human read. Your adviser and your insurer will both ask, and the answer is much easier to give if it was recorded at the time
• Spend the professional fees on the three things that bind you: identity verification, the purchase agreement, and the price

Used this way the gain is real and large. Reading everything instead of sampling is a genuine change in what gets found, and it arrives early enough to be disclosed rather than discovered. That is worth more than the hours saved, because late discovery is one of the recognised reasons deals collapse after the price is agreed.

The summary is short. AI has pushed the cost of reading toward zero. It has not moved the cost of being wrong, and in an acquisition the second number is the one that decides the outcome.