What an AI you can prove is worth

Rules are not a cost. They are what puts a price on your AI.

Ever since the AI Act came into being, I have heard the same line in executive committees, from Paris to Dubai: "America innovates, China copies, Europe regulates." It raises a smile, it is often right on one point, and I do not dismiss it. A rule has a cost. The postponement the EU voted this summer, which pushes back the obligations on high-risk systems by sixteen months, shows that lawmakers themselves have heard the argument.

But I believe this debate is reading the wrong document. Read in the income statement, regulation looks like an expense, and that is where everyone files it. I have spent twenty-five years buying and selling companies, and I have learned that a buyer does not pay for expenses avoided: he pays for what the company will earn tomorrow, and he discounts everything he cannot verify. Seen from there, the question is no longer "how much does compliance cost?". It becomes: "what is an AI worth when you can prove what it does, and what does one cost when you can prove nothing?"

Your AI is already an asset

According to Ocean Tomo, intangible assets accounted for about 92% of the market value of S&P 500 companies at the end of 2025. Brands, patents, software, data, know-how: most of what a buyer pays for today cannot be touched.

AI is joining this value that the balance sheet does not show. An agent that handles customer requests, a model that sets prices, an assistant that drafts sales replies: these are assets that generate revenue or savings. And like any asset, their value has always been calculated the same way. What it earns. Minus what it can cost. All of it adjusted for the uncertainty on those two figures.

Everything turns on that last term.

Unmonitored AI is a hidden liability

Two true, documented stories, which I tell every executive who assures me their AI "runs fine".

In November 2022, an Air Canada customer asks the website chatbot how to obtain the reduced fare offered for bereavement. The chatbot invents a procedure: buy now, get refunded later. The airline's actual policy said the opposite. Before the British Columbia tribunal, Air Canada argued, in effect, that the chatbot was responsible for its own actions. The tribunal called the argument "remarkable" and stated the obvious: a company answers for all the information on its website, whether it comes from a static page or a chatbot. On 14 February 2024, the airline was ordered to pay 812 Canadian dollars. The amount is trivial. The principle is not: every sentence your AI speaks commits you.

Now change the scale. In 2021, Zillow, the American real estate listings giant, buys homes at scale on the strength of an algorithm that estimates their resale price. The market moves faster than the model. On 2 November 2021, the company announces a write-down of about 304 million dollars on its inventory, expects a further 240 to 265 million in losses the following quarter, shuts the business down and lets go of about a quarter of its staff. Its CEO sums it up: the unpredictability in forecasting home prices "far exceeds what we anticipated". The model was not bad on day one. It drifted, and nobody had the instrument to see it in time or the hand to stop it.

In both cases, the liability existed before it cost anything. It sat on no balance sheet. It was waiting.

The AI Act now gives it an order of magnitude. Fines can reach 35 million euros or 7% of worldwide turnover for prohibited practices, and 15 million or 3% for other obligations. A buyer, a lender or an insurer who reads those figures does not wonder whether your AI is brilliant. They wonder whether you can prove it is under control.

What rules did to real estate

To understand what is happening, look at what happened to housing in France.

For decades, an energy sieve sold for roughly the same price as any house of the same size on the same street. The defect existed, heating cost a fortune, but it was invisible at signing. Then the energy performance diagnosis (the French DPE) became legally binding in July 2021, and the law set a calendar known to all: homes rated G can no longer be let under a new lease since 1 January 2025, F will follow in 2028, E in 2034.

The market did the rest. According to French notaries, a 90 to 110 m² house rated A or B sold in 2021 for one and a half times the price of a house of the same size rated F or G. By 2024, the ratio had reached 1.9.

The rule did not create the defect. It made it visible, comparable, and therefore priceable. And it rewarded those who had diagnosed and renovated ahead of the others: their property did not gain value because it was better, but because they could prove it.

Finance has known this mechanism for a long time. A study of private US companies, which nothing obliges to have their accounts audited, shows that those who do so anyway borrow more cheaply: banks give more weight to a verified figure when setting their rate. An audit does not create profitability. It makes it believable, and credit is paid for in belief.

Your AI is entering exactly this movement. Tomorrow, an AI that is documented, monitored, whose incidents are traced and corrected, will be worth more than an identical AI nobody knows anything about. Not because it works better. Because it can be shown.

Stability is a competitive advantage

There is a second argument, one that finance executives grasp immediately: what destroys the value of an asset is not the rule, it is uncertainty about the rule.

Let us look at the facts without taking sides. Europe has a text and a calendar published in its Official Journal, and when it decided to postpone it, it did so through a regulation that sets precise new dates: 2 December 2027 for the high-risk systems of Annex III, 2 August 2028 for those embedded in regulated products. You can dispute the content. You can plan.

In the United States, Colorado passed the first comprehensive AI law in 2024, due to apply in February 2026. It was pushed back to June 2026, then repealed in May 2026 and replaced by a different text, applicable in January 2027. Meanwhile, a presidential executive order of 11 December 2025 set up a task force to challenge state AI laws in court. I am not saying which of these two models is better for innovation. I am saying which one is easier to put into a valuation spreadsheet.

A buyer, a fund or an insurer always prefers a known rule to an unknown one, even a more demanding one. And the AI Act applies to any company whose AI has effects in the Union, wherever it is based. A group that can already prove what its AI does, under the most readable rule on the market, holds an asset its competitors will have to rebuild in a hurry.

It is a first-mover rent. Like the houses renovated before 2025.

What the executives who get it are doing

Those who are pulling ahead do not talk about compliance. They talk about assets. In practice, they do four things.

They keep the register of their AI systems like a fixed-asset register: every system inventoried, with its named owner, its use, its risk class and what it earns. An AI missing from the register is an asset that can neither be valued nor defended.

They produce evidence as they go, like bookkeeping, not the night before an inspection. Evidence dated on the day of the event is worth something. Evidence rebuilt three months later convinces no one.

They report to the board, with a few indicators that stay stable from one quarter to the next: the reliability of answers, real costs, the state of the evidence, actual usage, and the freshness of the knowledge that feeds the systems. What gets presented to the board ends up being managed.

And they prepare for the day someone asks them to show it. That day always comes: a funding round, a sale, a tender, an audit, an incident. On that day, nobody will ask whether your AI performs well. They will ask you to prove it.

The danger, the advice

The danger	The advice
Your AI systems appear in no inventory: they create value nobody can price and risks nobody can see.	Keep a register of your AI systems like a fixed-asset register, with a named owner for each system.
Evidence is rebuilt at the time of an inspection or a due diligence.	Produce dated evidence as you go, from the day a system goes live.
The board discovers the company's AI systems when an incident happens.	Present a few stable indicators every quarter: reliability, costs, evidence, usage, reference sources.

Rules set the currency of proof

We can keep debating whether Europe regulates too much. It is a legitimate debate, and it will last. But while it lasts, something more concrete is taking hold: the rule defines what counts as evidence. It sets the currency in which trust will be paid, to customers, regulators, buyers and lenders.

The companies that win will not be those with the fewest rules. They will be those able to pay first in that currency.

So, to every executive reading this, one question only, the one a buyer will ask you sooner or later: if you had to sell your company tomorrow, could you prove what your AI does?

Pascal Mennesson, founder of Maltem Consulting Group, founder of 56North. If you are an executive and would like to discuss this, request a first conversation.