Three numbers have been circulating since the start of September, and they are usually read separately. In one year, the share of companies abandoning the majority of their AI initiatives went from 17% to 42%. Over the same period, 46% of large companies have no framework at all for measuring the return on their AI investment. And in France, 61% of employees who use AI at work go through public tools, against 19% who use internal ones.
The three figures are reported by Geoffrey Harrazi, co-founder of Idun, in an opinion piece published by Maddyness on 1 September. They come from S&P Global (2025), Wavestone (2025) and KPMG (2026) respectively; we have not read them in the original reports, and we quote them as he quotes them. One more thing worth stating straight away, and his piece does not hide it: Idun sells precisely the usage-measurement layer those numbers argue for. That does not make them wrong - they come from three different firms, which is reassuring on convergence - but it invites reading the recommendations with that in mind.
What interests us here is not the bubble debate. It is what the three numbers say once you line them up: companies are cutting at scale, without knowing why, while usage walks out of the building.
The internal ChatGPT did not fail at its goal. It lost a race
The most instructive passage of the piece is not about technical failure. It describes a perfectly rational decision that turned against the people who made it.
In 2023 and 2024, almost every large company launched its own internal assistant. The goal was legitimate then and still is: stop sensitive data leaking into public tools. Many built what Harrazi describes as a plain wrapper around a language model, a chat disconnected from the information system, often delivered eighteen months late.
During those eighteen months, the public tools changed nature. They stopped answering and started searching the web and chaining actions on their own. The internal project, meanwhile, had been specified against the 2023 version.
We reinvented the wheel in a bunker, and we built it square.
The result is the 61% against 19%. Shadow AI did not come back because the company had banned something - that is the usual mechanism, and we described it elsewhere writing about shadow AI in the public sector. It came back even though the alternative existed, was funded, and worked.
That is a different and more uncomfortable case. The project did not miss its target: the target moved during construction, and nothing in the setup flagged it. Harrazi reports the answer he still gets in meetings - we are up to date, we even have Claude - when the model has changed and the product around it has not moved an inch.
The lesson is not that building was a mistake. It is that an internal product built against a fast-moving external benchmark needs a mechanism that says when the gap has opened. That mechanism existed nowhere, because nobody was looking at what employees actually used.
What you lose by cutting a project you never measured
The second case is a large luxury house, and it is told without flattery by someone who worked there. Three years of investment, plenty of agents in production, teams that genuinely levelled up: a real execution success.
Among them, an agent analysing customer reviews, pushed by the executive team. The idea was appealing: ask a question in plain language - what do customers think of product X in red? - and let the agent query a vast review database. Adoption was close to zero. Users understood neither the tool's scope nor what to do with its answers. It was eventually shut down for lack of use.
Harrazi puts it in one sentence: cutting blind throws the learning out with the project. The loss is not the agent. It is that nobody could say why it had not caught on. Scope misunderstood? Answers unusable? Wrong moment in the workflow? With fine-grained usage measurement, the project could have been rescoped instead of stopped. It was stopped, and the learning left with it.
The detail that is worth the whole story fits in one line: they had added thumbs and a comment box to collect feedback. Nobody ever fills those in.
That observation is worth keeping for any measurement scheme. Declarative feedback does not work. It does not work because it asks for effort from precisely the person who is dropping out - and a user who is dropping out does not fill in a form, they close the tab. What can usefully be measured is what people do, not what they say they do.
Successful adoption is not immunity
The third case is the most unsettling, because it looks like a win. An asset management firm whose internal assistant is used daily by 90% of staff, alongside automated workflows and self-service public tools. The adoption rate most IT departments dream about.
And yet, Harrazi writes, nobody there can say who does what with which tool, what it costs, or what would be worth industrialising.
This case settles a frequent confusion: adoption is not value. A 90% usage rate proves the tool is convenient, not that it produces anything measurable. It even licenses an expensive mistake - instrumenting nothing, because the dashboard is already green.
It is the same blind spot the Institut Montaigne described when it observed that the problem is not the technology but the organisation, and the same one Uber hit when its annual budget was consumed in four months. In all three, what is missing is not one more tool. It is an answer to who uses what, to do what, and what it changes.
Why usage measurement is always the first line cut
The three cases share one cause, and it is not cultural. Instrumenting the use of a tool returns nothing while everything is fine. It produces no feature, no demo, no announcement. It only becomes valuable at the moment a trade-off has to be made - which is exactly the moment when it is too late to install it retroactively.
An executive signs off on a customer-review agent. They do not sign off on the layer that will say, eight months later, whether that agent is useful to anyone. The only way we have seen that line survive is never to present it on its own: attach it to the first use case and deliver both together.
Harrazi's distinction between steering and controlling is right and worth repeating, because it decides how the thing is received internally. Observing real usage has nothing to do with monitoring employees. What you are looking for is teams, moments, drop-off points - not individuals. A scheme that confuses the two will be worked around, and it will deserve to be.
What Survol does with this, and what it does not
Let us state the limit first, because it is a clear one. Survol does not measure the use of an office assistant. That is not our subject, and the three cases above largely call for tooling other than ours.
The same blindness does exist on software delivery, though, and that is where we answer:
- Consumption attaches to a unit of work, not to an invoice. Tokens are spent on a feature and a version, not on a month. That is what makes the real cost of a delivery visible - including when it took four sessions instead of one.
- Abandonment is a recorded outcome, not a disappearance. A feature can come out of verification in the state abandoned: dated, with a reason, attached to what had been asked for. That is the direct answer to cutting blind - stopping remains a good decision, stopping with no written reason is a double loss.
- Status is derived from execution, not declared. What is in production is there because a version was promoted, not because someone ticked a box. A typed status is a claim; a derived status is a fact.
Which leaves the question neither a tool nor an opinion piece settles on a company's behalf: when you cut your next AI projects - and you will - will you be able to say which ones deserved rescoping instead? Companies cutting today while knowing why are choosing their trajectory. The others will have it chosen for them, a year later, under pressure from their CFO.
Sources
- Geoffrey Harrazi, “42% of companies are abandoning the majority of their AI projects. Good.”, Maddyness, 1 September 2026 (in French). The S&P Global (2025), Wavestone (2025) and KPMG (2026) figures quoted in this article are taken from that piece.
