AI Act:
the day an employee demands answers.

A word from Jérôme

JDENIS Consulting · Toulouse

My perspective

I do not believe fear of a fine is enough to produce serious governance. It can even encourage superficial compliance: a charter signed, training completed, then decisions made exactly as before. What seems decisive to me is the ability to reconstruct the reasoning that led to a person being treated in a particular way.

This requirement also protects the employer. A manager should be able to explain what they accepted from an automated analysis, what they rejected and why. That requires time, skills and sometimes the ability to challenge a productivity target. Without those conditions, “human validation” risks becoming one more signature. Regulation alone will not resolve this organisational contradiction. It will, however, make it harder to leave it invisible until the first dispute. I would focus the company’s preparation here, on the actual work of its managers.

My perspective

I would be cautious about reducing skills management to how often someone uses an assistant. Usage frequency tells us about a practice, not directly about the quality of work. An employee may rarely use the tool because they handle the most difficult situations; another may produce more while passing the checking work to colleagues.

The studies presented here invite us to observe work in greater detail: which errors are detected, which lines of reasoning become explicit, which learning opportunities disappear? I would also preserve space for beginners and experts to compare methods without turning every trial into a performance measure. The aim is not to preserve every task unchanged. It is to understand how people develop professional competence when some steps are delegated. A company that measures only speed may discover too late that it has weakened the skills needed for oversight.

My perspective

The Japanese findings interest me because they also report difficulties. An average of self-reported time savings cannot fully describe what happens in a department. Time freed up may improve an answer, absorb additional demand or simply shift the workload to the person who knows the tool best. These effects do not have the same value for workers.

I would therefore follow a complete activity, including rework and handovers. In customer service, for example, a faster first response is progress only if it prevents a further complaint or helps resolve the problem more effectively. This reasoning requires us to look beyond the production of text. It also means asking the people who receive the automated work for their views. Their experience often tells us more than a query counter, even when it contradicts the project’s initial story.

My perspective

In a small business, a reasonable approach must remain workable on a busy Monday. I would start with a frequent, limited and reversible task whose result can be checked by the person who already performs it. An experiment becomes more informative when you know what you are prepared to stop if it disappoints.

I also think the additional work needs counting. Cleaning documents, correcting responses and managing a subscription are not free activities. A business owner can estimate that cost simply, across a few cases, without building an elaborate measurement system. A good first experience produces two results: a useful service and a clearer understanding of its limits. If it produces only impressive text, without reducing rework or improving customer relationships, I would not yet consider the investment justified.

My perspective

I would not choose an architecture solely on the basis of a model ranking. The vocabulary of the occupation, the documents available and access rights all influence the result substantially. Progress in smaller models and local execution opens up interesting options, but also transfers part of the operational work to the company.

The next question is who will handle updates, regression testing and incident diagnosis. These are lasting responsibilities, rarely visible in a demonstration. Conversely, outsourcing operations to a remote service can simplify daily work while increasing contractual dependency. I prefer to compare these situations around a specific use, with criteria for quality and the ability to switch away. The discussion becomes less spectacular, but it helps identify a solution the team can actually maintain.

My perspective

A recurring difficulty in economic evaluation concerns the unit of measurement. A subscription is easy to count; a better-informed decision is less so. We may then take the available numbers, such as queries or documents produced, and treat them as evidence of value. This substitution deserves discussion with the teams involved.

I would suggest a more modest calculation: how much does a correctly completed case cost, including checks and rework? Comparing a few situations before and after AI is introduced will not establish a universal truth, but it can inform a local decision. We also need room for benefits that do not immediately translate into time: better understanding, a more accessible answer, or fewer costly errors. Those benefits should be described and observed. Announcing them before measuring them would expose the project to the same weakness as the productivity promises it claims to replace.

My perspective

In field work, the gap between a suggestion and an action deserves particular attention. A documentary error can be corrected; a physical intervention operates on a different timescale and may involve safety. I think it is reasonable to begin by improving access to information and preparation for operations before extending a system’s autonomy.

This caution does not rule out innovation. It helps establish what has actually been demonstrated. A simulation, a sales presentation and an on-site trial do not provide the same level of evidence. I would also ask operators which incidents the demonstration protocol leaves out: damaged equipment, contradictory instructions, unfamiliar surroundings. Their experience may improve the specification more usefully than another performance score. Finally, the authority to stop a system must be usable, understood and exercised in a trial before it is cited as a safeguard.

My perspective

Ethics becomes concrete when we ask what a person can do about a decision they do not understand. Can they access the information concerning them? Will an error be corrected before it has an effect? Will their disagreement be examined by someone with the necessary time and authority? These questions carry more weight than a general promise of responsible AI.

I do not, however, think that every collection of data is illegitimate or every automation unfair. An organisation may need to verify access, prevent fraud or distribute a workload. The difficulty lies in justifying the means chosen, their proportionality and their secondary uses. A metric collected to improve a service can become something quite different if it is later used to penalise a person. That change of purpose, often gradual and barely visible, is what I would make systematically open to challenge within the project’s governance.

My perspective

Postponing a deadline may have a practical justification: time to clarify standards, organise oversight or adapt products. It may also be read as a signal to wait. In my view, the useful distinction concerns what a company can already clarify without waiting for another guideline: its uses, its data and the people making decisions.

I would also pay attention to how costs are distributed. Documentation requirements without support can weigh heavily on a small business, while poorly designed easing of the rules can leave affected people without an effective remedy. These tensions deserve precise discussion, rather than an abstract choice between innovation and regulation. A useful criterion is whether the arrangements allow meaningful experimentation while making it possible to account for a contested decision. That requires resources for implementation, not just a timetable.

My perspective

The ease of producing coherent text does not remove the need for editorial judgement. It makes that judgement more important, because convincing wording can be obtained before the facts are established. For company communications, I would separate checking claims from editing style: these are different tasks, with different criteria.

The fictional story in this edition follows that reasoning. It makes organisational decisions and their possible consequences visible; it should not borrow the authority of a court case that never happened. The same requirement applies to images, testimonials and client feedback. I prefer content that honestly explains what it shows to a more striking narrative whose status remains ambiguous. That transparency does not diminish the story’s interest. It lets the reader discuss the reasoning rather than having to guess what is true.