How to spot AI washing in a target company before you buy

AI washing in a target company takes two forms. The simple one is AI that does not exist, and a technical review can expose it. The hard one is AI that exists but does not do what is claimed. When a language model can be rented by the call, almost any software company can show something that answers like an AI, and the value then turns on a different question: how much of the work described as automated is still done by people, on whose technology, and at what cost per completed case.

Two of the three US enforcement cases described below concern AI that did not exist. The third, and the most instructive for a buyer, concerns AI that existed but did not do what was claimed. There the truth lay not in the presentation but in operations: in who actually took the order, and in what the automation rate counted.

Two measurements of AI washing, seven years apart

In 2019 MMC Ventures reviewed 2,830 companies in thirteen European countries, Sweden, Denmark, Finland and Norway among them, that popular tools classified as AI start-ups. The review went company by company, through activities, focus and funding. Evidence that AI was material to the value proposition was found at 1,580 of them. The review, in other words, tested whether there was any AI that mattered.

A study from the University of Hamburg, released as a preprint in October 2026 and accepted at the AIES 2026 conference, measures something else. The researchers coded the public marketing of 100 German AI start-ups, drawn from two publicly accessible directories of German start-ups, and compared it with what the products actually do (Pranav, Hartmann and Lauscher, 2026). 35 per cent had no detectable AI or were a wrapper around a rented language model with no contribution of their own. The coding separates having no AI from borrowing it without adding anything, but the figure reports the two together. The study's composite measure of AI washing placed 39 of the 100 companies in the top two levels of a five-point scale.

The figures cannot be compared with each other. The methods, populations and definitions differ. The Hamburg study covers German companies only, and MMC's sample spans thirteen countries, four of them Nordic. The Hamburg researchers' coding of the 100 companies was, moreover, based solely on public material.

Three cases where the claim and the operations parted ways

Presto Automation, which sold voice ordering for drive-thru restaurants and was listed on Nasdaq, is the case where AI existed but did not do what was claimed. According to the order the SEC issued in January 2025, every Presto Voice unit the company deployed from November 2021 to September 2022 ran on another supplier's speech recognition, without the company adequately disclosing this when it described the product. When its own technology was eventually rolled out, the company claimed it eliminated the need for human order takers. In fact, order takers it had hired abroad, primarily in the Philippines and India, handled the vast majority of orders on those units.

For the units running its own technology, the company stated that at least 95 per cent of orders were completed without intervention. According to the SEC, the figure referred to orders completed without intervention by restaurant staff, while the orders required significant human intervention off site. During the period when every unit ran on the supplier's technology, Presto reported rates of 95 to 99 per cent at its largest customer, without disclosing that the data came from the supplier and had not been independently verified by Presto. Presto consented to the order without admitting or denying the findings and paid no civil penalty, which the SEC attributed to its cooperation during the investigation and its remedial efforts.

An automation rate only tells you something once you know whom it leaves out.

Nate, a shopping app that its founder said completed purchases with its own AI, is the more serious case, and the criminal case against the founder is still undecided. According to the SEC's complaint of April 2025, the founder raised over $42 million on the claim, while during the funding rounds purchases were in practice placed by hand by contract workers, primarily in the Philippines. An employee is alleged to have described the automation rate in an internal message in June 2021 as "essentially zero", and engineers are alleged to have stood by behind the scenes when the app was demonstrated to investors. The founder has pleaded not guilty in the parallel criminal case, and the allegations have not yet been tested in court.

Delphia, an investment adviser, claimed that its AI used clients' data in its investment decisions. According to the SEC, no such data was used. In March 2024 the firm agreed to pay a civil penalty of $225,000, without admitting or denying the findings.

In the Presto and Nate cases, the claim sat in the material an investor reads, that is investor presentations and, for Presto, also press releases and filings with the SEC, while the evidence against it lay in operating data only the company could see. A review that stops at the management presentation therefore sees what the investors saw.

What the seller shows cannot be checked from outside

The Hamburg study shows why this is hard to see from outside. 43 per cent of the companies claimed AI capabilities without any publicly inspectable evidence. Of the 62 that gave performance figures, 49 disclosed neither source nor method, the same problem that makes vendors' benchmark figures hard to compare. 76 per cent named no limitations at all: no edge cases, no accuracy bounds and no need for human oversight. Where evidence existed, it was usually client logos, anonymised case studies and controlled demos, and the researchers' point is that the seller controls what each of them shows. A logo proves a customer relationship, not what the AI did for that customer.

A buyer in a deal has what the researchers lacked: access to the data room. That moves the review from what the company says to what it can show in operation, and four questions make the best use of that access.

The first concerns the definition. Every automation rate in the presentation should come with an exact definition: what counts as an intervention, and by whom. Presto's rate counted only interventions by restaurant staff.

The second concerns origin. Whose model or service runs in production today, and does the dependency appear in the supplier contracts and in what customers have been told? Building on a rented model is normal. Presenting it as your own is not.

The third concerns economics. The cost per completed case should be calculated with the people included, whether they sit in the company, at a subcontractor or in another time zone.

The fourth concerns time. Operating data over a period says more than a demo, because a demo runs on inputs the seller has chosen. How such a review is scoped to the time a deal allows is described on the page on AI in private equity.

The human in the loop is not the proof

There is an objection that weighs more than it first appears. Finding people behind the automation proves nothing in itself. Many AI services that work well are built so that people review the hard cases, because that is how quality is kept. Presto was not such a case: in the original version of its own technology the contracted agents entered every order, and an executive warned internally that the AI was doing none of the work. The difference lies in how large the human share is and whether it is disclosed, not in whether people are there.

People behind the automation are a fact to test, not a revelation.

As regards the people, the SEC found two faults at Presto: the claim that people were not needed, and the failure to disclose its reliance on them. Neither fault was that the people existed. The same applies to the rented technology: the problem was that the dependency was not adequately disclosed, not that the supplier was used. A review that flags every human step and every rented model as AI washing will condemn companies that are good deals and at the same time miss what actually decides the value, namely whether the economics hold once the cost of the people is counted, and whether their share falls at the pace the plan assumes.

The criterion: does the AI claim carry the price?

Strike the AI claim from the investment memo and rerun the deal. If the multiple, the growth assumptions or the value creation plan change, the claim carries the price. It should then be tested in operating data against the four questions, as rigorously as revenue. If nothing changes, the AI claim is marketing. It will still be a question at exit, because the next buyer will ask it, but it is not the question that deserves the diligence time.

That time then belongs to the opposite risk: that AI undermines what the target sells. In Alvarez & Marsal's survey of 83 senior private equity professionals in eleven European markets, Sweden, Denmark and Norway among them, 80 per cent said they factor AI into the investment thesis as a driver of value. 57 per cent named deals with a high risk of AI disruption as one of the situations where integrated due diligence adds most.

The same question therefore produces two different reviews. For a target whose price rests on AI, the question is whether the claim holds in operation. For a target whose price does not, the question is whether AI threatens what the company sells, and there AI washing is not the largest risk.