Friction or evaporation: two regimes of AI transformation in SMEs
For several months I have been working with executives on the AI transformation of their companies. With several of them, I see with pleasure that they have been pushing the topic internally for quite some time. They regularly ask their teams where they stand on their AI projects.
They invest in tooling so they can run solutions without external dependencies. They reinvest as use cases broaden, and they try to bring in outside expertise to give the movement some structure.
And still, they tell me: "we feel like we are not moving forward."
When I look at what they do, it is exactly the list of what the transformation literature describes as good practice: sponsorship from the top, regular governance, hardware investment, outside support. They do everything. And still, it grinds.
So they ask me what they are missing.
For me, the explanation comes down to this: a wrong estimate of the normal cost of transformation at this stage. The friction they observe is not the signal that the effort is failing. It is the entry price when you carry an AI transformation from the top of an SME.
This dissonance between the effort spent and the felt result is documented.
In its report The state of AI 2025: How organizations are rewiring to capture value, McKinsey observes that high performers on generative AI are 3.6 times more likely to pursue transformational change than other organizations. The firm also identifies direct CEO oversight of AI governance as one of the elements most strongly correlated with actual EBIT impact.
In its annual report Gen AI Fast-Tracks Into the Enterprise (October 2025), Wharton observes that executive leadership on generative AI has surged by 16 points in one year and that the Chief AI Officer role is now present in 60% of the enterprises surveyed. Strategy and AI accountability are moving into the C-Suite.
Bpifrance Le Lab, on the French SME landscape, measures that in 73% of cases the AI impulse comes directly from the executive.
These three sources converge on one observation: the main condition for an AI transformation that actually lands is that it be carried directly by the top, week after week.
But what these reports do not say is that what these executives are going through is painful. It is the expected pattern of the phase they are crossing, not a sign that the effort is failing to take.
There is a moment, in this kind of transformation, where the fruits arrive more slowly than the effort. That moment makes you believe the work is stalling. In reality it signals that the transformation is paying its entry ticket.
Because it is hard. For real. It is not that AI is useless, incompetent or overhyped. What we do with generative AI, whatever the detractors say, is just incredible. But these executives find themselves caught in a vice. On one side, dishonest announcements selling a dream. On the other, teams who are either underwater or reluctant to change, especially when change is perceived as dangerous or time-consuming.
I have another point of comparison. A shorter engagement, in a nice SME, on a piece of technical analysis. The deliverable was documented, with a series of observations to take on board, several of which were blocking for what came next. I had taken care to reproduce everything and to attach what was needed for the team to verify it themselves.
The report was not read. Full stop.
In the weeks that followed, I noticed that the observations I had raised had remained unanswered. The technical choices in progress had not been revisited. The top had preferred to push on with its own track rather than integrate what had been handed over.
It is not that the analysis was wrong or that the points raised were not real. It is that no one at the top looked. The work delivered did not find a reader in the decision chain, and it dissolved.
What this second case shows is not friction comparable to the first case, only more intense. It is a totally different regime. In the first case, those executives push, they read what gets reported back, they integrate the feedback, they pay to understand, and the work moves forward slowly. In the second, the top did not look at what had been handed over. We are no longer talking about friction. We are talking about evaporation of the work.
When I compare the two situations, I cannot explain the difference in progress by the variables we usually point to. Size is not the factor. Budget is not the factor, neither of them had unlimited means. Technical maturity is not the factor either, the team in the second case had at least equivalent skills.
The variable that separates the two regimes is who at the top reads what gets delivered, and who at the top carries the topic week after week.
In the first case, those executives read. They take on board the technical trade-offs even when those exceed their comfort zone. They validate budget decisions on that basis. They publicly carry the topic internally, week after week. The result is slow, but it is real.
In the second case, no one at the top reads. The top keeps going on its own track without integrating what gets reported back. The work that was commissioned disappears into the void.
What sits between the two is not a gradient. It is a binary switch. An AI transformation whose top reads moves forward slowly. An AI transformation whose top does not read does not move forward at all.
If you carry your SME's AI transformation, if you read what is reported back to you week after week and you still feel that it grinds, then you are paying the right cost.
The pattern you observe is the expected pattern of the phase you are crossing. Friction is not the symptom of a coming failure. It is the entry price of the transformation you are leading.
Keep going. The condition for the switch is not a new technical skill in your team. It is what you are already doing, held over time.
If you run an SME and you have not read the latest technical report that was delivered to you on AI, open it. The regime you are about to fall into depends on that reading.
PS: in parallel I keep developing herbert.rs, my open source inference engine. A new version is in the final stretch. What I see on the ground in SME AI transformations regularly feeds what I code, because questions of technical architecture and questions of executive sponsorship do not show up separately in practice.