AI is essential everywhere, but without vision or governance, companies repeat the mistakes of “digital transformation”: patchwork of tools, shadow AI and risks rather than a controlled architecture.
In two years, AI has been installed everywhere in the company. In office suites, CRMs, business tools, customer relations platforms. Each structure added its “AI” layer, each department tested its uses, each employee found its shortcuts. The move was logical. It is now becoming more difficult to fly.
The topic is no longer whether to adopt AI. This stage is behind us. The subject becomes more demanding: how to transform a diffuse, heterogeneous and sometimes opportunistic presence into a coherent, controlled and value-creating work architecture?
From tool catalog to real work
Because the blind spot is there. By relying on tools, the company ends up losing sight of the essential: the real work. Who produces the information? Who reformulates it? Who validates it? Who operates it? Where are the frictions, the waste of time, the duplication, the blind spots of responsibility? As long as AI remains an overlay added to existing software, it sometimes improves isolated gestures. It rarely transforms the organization.
The first risk is that of patchwork. A summary function in messaging. An editorial assistant in the office suite. A co-pilot in the CRM. An analysis engine in a business tool. Taken separately, each of these additions can make sense. Together, they often create a landscape that is difficult to read: functionalities that overlap, uses that vary depending on the teams, costs that are dispersed, security rules that are not always aligned.
Added to this is a third, more discreet bias: some of the current uses simply do not contribute much to real work. We generate reports that no one rereads, summaries of emails that are already short, automatic presentations that are more of a technological demonstration than a business need. As long as AI remains focused on content that is easy to produce but not strategic, it consumes time, budgets and attention, without really reducing the workload that matters to teams.
Shadow AI: a signal more than a fault
The second risk is quieter, but more structuring: that of ungoverned individual uses. When internal tools are absent, limited or considered too restrictive, employees work around them. They open a personal account on a consumer assistant, test a presentation generator, a transcription tool, an analysis engine or a code assistant. This “shadow AI” phenomenon is now very real. With it, a simple and major risk appears: the exposure of sensitive data outside the company’s control perimeter.
You have to look at it without naivety. If employees circumvent, it is generally neither casually nor maliciously. This is because they are looking to save time, to simplify a task, to lighten a mental load. In this way, “shadow AI” is less of a disciplinary problem than an organizational signal. It often says two things at the same time: that a need for productivity is real, and that the company has not yet provided a sufficiently simple, useful and readable working framework.
Change level: from tool to workflow
It is precisely for this reason that we must change the level of analysis. A company does not transform because it has activated a lot of AI tools. It transforms when it decides how AI fits into its workflows. The right starting point is therefore not the catalog of available solutions, but the mapping of processes: where information enters, where it is enriched, where it is validated, where it gets blocked, where it loses value.
From there, AI returns to what it should be: an information processing layer serving a specific workflow. Some steps can be prepared more quickly. Some syntheses can be made more reliable. Some recommendations can be made earlier. Certain interfaces between teams can be streamlined. But this assumes a simple condition: start from the real process, not from the technological demonstration.
Architecting, in this context, does not mean immediately launching a major theoretical program or yet another abstract platform. This means first of all putting things in order. What AI tools or modules are already in use? By whom? For what tasks? With what data? With what level of human supervision? With what visible and invisible costs? And above all: what parallel practices are already developing outside the official framework?
This exercise is much more strategic than it seems. It allows us to move away from fascination and into readability. We then see the redundancies, the blind spots, the scattered expenses, the really useful use cases and the insufficiently covered risk areas. We can then select a few priority processes, not the most spectacular, but those where AI can produce a tangible effect: a lot of volume, a lot of cognitive load, a lot of back and forth, a lot of wasted time.
A framework that must come from management
But this framework cannot be based solely on project teams or on a few passionate profiles. It must be supported by management. Without an explicit vision of what the company really wants to change in its way of working, AI will reproduce the same shortcomings as other waves of transformation: stacked projects, competing initiatives, broken promises and team fatigue.
It is also at this moment that a more decisive question appears: who does what? As long as this new division of labor is not made explicit, AI remains either a vague promise or a diffuse concern. But value is at stake in very concrete trade-offs. What remains entirely within human control? What can be prepared, synthesized, structured by a model? What could, tomorrow, be entrusted to more autonomous agents, provided they are supervised and limited?
Do not repeat the mistakes of “digital transformation”
Here we find a known pattern. During digital transformation, many companies multiplied projects, tools and redesigns, without always clarifying what they really wanted to transform in their model and in their businesses. AI today follows a similar trajectory: many initiatives, little prioritization, even fewer renunciations. The difference is that the cycle is faster, the expectations higher and the risks, particularly on data, more sensitive. Repeating the same mistakes this time would have a higher cost.
It is on this condition that AI can become an organizational lever rather than a factor of dispersion. Not because it replaces work, but because it more clearly redistributes certain tasks, certain controls, certain preparations. In other words: the issue is not to have more AI in the company. The challenge is to have more control over the way in which it already works with us, sometimes visibly, sometimes in the shadows.
The next few months will not decide which companies have tested the most tools. They will decide between those who have been able to connect three levels that are still too often separated: the real uses of employees, the business processes where value can be measured and the common base of governance, security and responsibility. It is this passage which allows you to escape from the pile-up.
AI has already entered the enterprise. The real question is now simpler, and more demanding: will it remain an addition of functionalities, workarounds and risks or will it finally become an accepted work architecture?
