Leadership and AI: why both fail when you run them separately

There are today two perfectly watertight markets.

On one side, executive coaches. They work on stance, on decision-making, on the relationship with the leadership team. They do not touch information systems, which is consistent with their trade.

On the other, artificial intelligence providers. They map, deploy, integrate. They do not work on the leader, which is equally consistent.

Between the two sits a mid-market chief executive who funds both, and who finds after a year that neither has produced what was hoped for.

This is not a question of the quality of the people involved. It is a question of scope.

The figure that should close the technical debate

When an AI project fails in a company, where does the cause sit?

The Boston Consulting Group put the question to 1,000 executives in 59 countries and more than twenty sectors. Its conclusion, published on 24 October 2024, is a split: 70 % of the difficulty comes from people and processes, 20 % from technology, 10 % from algorithms. It recommends allocating resources in the same proportion, which nobody does.

The same survey notes that 74 % of companies have yet to demonstrate tangible value from their use of AI, and that only 26 % have moved beyond proof of concept.

This is not a model problem. It is not a vendor problem. It is an organisational problem, and therefore, in a mid-sized company, a leadership problem.

Abandonments are not slowing, they are accelerating

One figure has been circulating since the summer of 2025: “95 % of AI projects fail”. It comes from an MIT Project NANDA report and is almost always misquoted. The report describes a funnel in which roughly 5 % of organisations reach a deployment with measured value, which is not the same as 95 % of projects failing. Its authors themselves describe their results as “directionally correct”, and the study was not peer reviewed.

There is a soberer and more solid measure. In its Voice of the Enterprise survey of March 2025, covering more than a thousand respondents in North America and Europe, S&P Global Market Intelligence found that 42 % of companies abandoned most of their AI initiatives in 2025, against 17 % the previous year. On average, 46 % of proofs of concept never reach production.

The abandonment rate more than doubled in a year, at precisely the moment the tools were getting better. That detail alone disposes of the technical hypothesis.

Plenty of use, little effect

The most recent picture comes from McKinsey, in August 2026, across 1,719 respondents in 97 countries.

Close to nine organisations in ten report regular use of AI in at least one function. 44 % report enterprise-wide deployment. And only 37 % see a contribution to operating results, a figure almost unchanged on the previous year. Companies achieving significant impact, beyond a 5 % effect on results, are only 6 %.

What McKinsey observes in that 6 % is what matters here. They pursue growth and not only efficiency. They redesign processes instead of inserting AI into the existing ones. And their leaders are visibly engaged on the subject.

An earlier survey from the same firm, in March 2025, is blunter still: among twenty-five organisational attributes tested, oversight of AI governance by the chief executive personally is among the elements most correlated with reported impact on results. That is the case in 28 % of the organisations surveyed.

Correlation, not causation: that needs saying. A leader who personally oversees the subject is probably also the one who understood it, arbitrated it and prioritised it. But the order of magnitude is there, and it points neither at the vendor nor at the IT department.

In France, the discriminating factor is the leader

The French data closes the argument.

Insee, in a survey of around 11,000 companies published on 21 July 2026, measures that 18 % of companies based in France report using at least one AI technology in 2025, a rate multiplied by three in two years. The breakdown by size is clear: 15 % for companies of 10 to 49 employees, 31 % from 50 to 249, 58 % above 250. The main brake cited by non-users is not cost, it is lack of perceived usefulness, at 71 %.

It is the Bpifrance Le Lab study of June 2025, covering 1,209 leaders of small and mid-sized companies, that draws the most useful conclusion. It establishes that adoption depends less on sector and size than on the leader’s profile, and distinguishes four archetypes: sceptics 27 %, blocked 26 %, experimenters 28 %, innovators 19 %. Among the brakes cited, two are strictly organisational: difficulty identifying use cases, for 23 %, and employee resistance, for 22 %.

In other words: in a French mid-sized company, the variable that determines whether an AI project succeeds sits in the leader’s office, not in the infrastructure.

What happens when you automate around a saturated leader

So much for the data. Here is the mechanism.

A company deploys a tool that accelerates the production of information: summaries, proposals, analyses, dashboards. That output arrives somewhere. In a mid-sized company, nearly all of it arrives in the same place.

The leader who used to receive ten notes a week now receives twenty-five, better written, harder to sort because a mediocre note is now well drafted. The rate of decision has not moved. It has in fact fallen, since there is now more material to assess.

AI did not create the bottleneck. It made it visible, by raising the pressure upstream. That is exactly what McKinsey’s observation about redesigned processes, as opposed to AI inserted into existing ones, means: putting an accelerator into a chain whose blocking point is downstream simply raises the pressure on that point.

And the reverse is just as true. A leader who works on their decision load without touching the information flows of the company gets a real result, but one that does not hold: they decide better inside a system that keeps producing, at the same rate, subjects that climb to the same place. Six months later, the load is back.

A three-stage order of battle

The sequence that works is always the same, and it is counter-intuitive because it puts the tool last.

One: the decision. Identify the two or three decisions the rest of the company depends on, and for each, who actually takes it, on what information, at what frequency. That is half a day’s work and it needs no software.

Two: the flow. Look at what feeds those decisions. Where the information comes from, how long it takes to arrive, how many people handle it before it is usable. That is where the use cases that matter sit, and they almost never resemble the ones in a provider’s catalogue.

Three: the tool. And only then. On a narrow perimeter, with a measure defined before the start, and a named person responsible.

That order explains why so many technically successful projects have no measurable effect: they started at step three.

What to demand of anyone claiming to cover both

The crossover between leadership work and applied AI has become a sales argument. It deserves checking, and that is done in three questions.

What deliverables do you produce in each register? Anyone genuinely covering both produces deliverables of both kinds: framing notes on the leader’s decisions, and a data map, costed use cases, a deployment schedule. If there are deliverables on only one side, the other is an argument, not a service.

What happens if you conclude the AI project is not worth it? That is the sorting question. A provider whose revenue depends on deployment will never conclude that you should not deploy. Someone whose engagement is yearly and whole can say it, and should be able to cite a case where they did.

Who arbitrates between the two when they conflict? They always conflict, usually over the leadership team’s time. If nobody arbitrates, both advance in parallel, get in each other’s way, and end up where the 42 % of abandoned initiatives of 2025 ended up.

It is not AI that transforms a mid-sized company. It is a leader whose decision capacity has been cleared, in a company whose flows have been redrawn, who then equipped themselves accordingly. In that order.