What AI capability should we look for when acquiring a company?

Most of what is sold as AI capability is a wrapper around a model the seller rents from somebody else, and the model will have changed twice before the earn-out ends. That is the first thing to know when acquiring a company for its AI, and the second is that it does not make the acquisition wrong. It makes the model the least durable thing you are buying. I have led the acquisition of an AI company and directed it afterwards. It had three and a half people and a product that schools bought one at a time from a person they trusted. What we bought was not the model. This page is about what survives diligence, and what diligence has to start asking.

What is not there#

Gartner estimates that of the thousands of vendors claiming agentic AI products in 2025, about 130 were real, and calls the rest agent washing: existing chatbots and automation renamed. MIT NANDA's 2025 report, small as its sample is, found that tools bought from specialist vendors reached deployment about twice as often as internal builds, which is a point in favour of buying capability and a warning about what most of it consists of. Diligence that asks 'what is the model' gets an answer that will be obsolete by completion. The questions that hold are about everything around it.

Four things that survive#

The data, and the right to use it. What the company's system learned from, whether the people whose data it was consented to that use, and whether the consent transfers on a change of control. A model can be rebuilt. A dataset gathered with permission that competitors do not have cannot, and one gathered without permission is a liability with a valuation attached.

The allocation, and who checks it. Which decisions the company's system makes on its own, which it recommends, and who reviews them. Ask for the disagreement rate: how often a human reviewing the output reaches a different answer. A company that cannot say has a system nobody is checking, and after completion the acquirer is accountable for every decision it makes. If the company operates in recruitment, credit, education or any other Annex III domain of the EU AI Act, that accountability is now a legal one under Article 14.

The people who can still tell when it is wrong. A small AI company's capability is usually two or three people who understand the domain well enough to know when the output is nonsense. They are the asset, they are the first to leave, and no code escrow replaces them. Ask what the company can still do with the model switched off. If the answer is nothing, the company is a subscription.

What the customers actually bought. In the company I bought, the customers had bought a person, a workflow and a promise of support, and the AI was the reason the person could keep the promise. The retention after acquisition depended on keeping the person, not the model. Ask the customers, not the deck, what they would lose if the product went away.

How the approach to M&A changes#

Diligence acquires a capability question it never used to ask: what does this company know how to do without its model, and what happens to it when the model changes underneath it. That is capability debt read as a balance-sheet item, and it applies to the acquirer as much as the target, because integration tends to switch off the acquirer's own practice in whatever the target automates. And integration acquires a rules question: which of the target's automated decisions the acquirer is now accountable for, and whether anyone in the acquirer knows how to check them. A target with a clean allocation and a measured disagreement rate is worth more than one with a better demo, and almost no valuation model prices that.

What nobody has measured#

There is no study of AI acquisitions by what was bought and how they performed, and the experience on this page is one acquisition, described by the person who led it. Gartner's vendor estimate has no published method. The MIT figure rests on 52 interviews. The four things listed as durable are an argument from what has changed hands and what has not, in one case and in the public reversals documented elsewhere on this site, and should be read as a diligence checklist rather than a finding.

Key sources

• Gartner (2025). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Graded entry.
• Challapally, A. et al. (2025). The GenAI Divide. MIT NANDA. Graded entry.
• Regulation (EU) 2024/1689, Annex III. Graded entry.
• Boston Consulting Group (2026). When Everyone Uses AI, Companies Risk Losing Critical Skills. Graded entry.

Related SuperSkills research#

On what happens when the model changes underneath a workflow, preserving capability across vendors. On the capability question as an audit, what is a capability audit. On the accountability the acquirer inherits, meaningful human oversight. On who should hold these decisions, AI leadership.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. He has run, grown, bought and advised businesses with AI in them. Findings are attributed to the studies and statements that produced them and kept separate from the interpretation. This is a living reference, reviewed and updated as significant new evidence appears.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.