A targeted postponement does not mean a general pause
The timeline of the Council of the European Union indicates that the majority of the rules became applicable on August 2, 2026. It distinguishes the new deadlines concerning high-risk systems: December 2, 2027 for the autonomous systems concerned and August 2, 2028 for those integrated into products. These dates therefore do not exempt us from identifying the rules already applicable to a given situation. [1]
The definitive green light from the Council on June 29, 2026 concerns a precise set of simplifications and modifications. As of October 7, 2026, communicating from an old, undated calendar exposes you to contradictions. A sales presentation, tender response and internal policy must refer to the same verified version. [1]
The reputation issue appears when the company announces too quickly “the AI Act is postponed” or “we are already compliant”. The first formula can make one believe in an abandonment of protections; the second promises comprehensive assurance without explaining the systems covered. The slogan must be replaced by a useful description of the use, the role of the company and the controls carried out.
Based on the uses and role of the company
The term “artificial intelligence” brings together very different situations. A company that uses a writing assistant, markets a recruiting tool, and integrates AI into a product cannot treat them as a single line in a presentation. The inventory must identify each system, its intended use, the users involved and the possible consequences of an error.
The legal qualification of the role and the risk falls to the competent teams. The communication work then begins: translating this analysis without distorting it. It is necessary to specify which entity is speaking, for which version of the product and in which market. A French subsidiary must not mechanically take over a global claim if the uses or deployment conditions differ.
For each priority system, request an operational owner, a legal referent and a person responsible for public formulation. This distribution prevents communication from being the only one to validate a technical assertion. It also allows you to quickly correct a message when a supplier modifies its model or a new usage appears.
Keep three separate records
We recommend distinguishing between applicable obligations, future maturities and voluntary commitments. These categories can coexist in the same program, but they do not give rise to the same formulations or the same supporting documents.
| Register | Control question | Formulation to construct |
|---|---|---|
| Applicable obligations | What rule concerns this usage today? | Describe the provisions and controls actually concerned |
| Preparing for a deadline | What work remains to be done and by what date? | Present a program, its manager and its milestones |
| Voluntary commitment | What level of protection have we chosen? | Explain the commitment, its scope and its monitoring |
| Uncertainty to be resolved | What qualification or data is still missing? | Say what is under review and when an answer will be given |
A voluntary measure can be useful without becoming a general guarantee of compliance. Conversely, an obligation fulfilled on one use does not prove that the entire group meets all the requirements. The table must retain the review date and the link to the technical, legal or organizational elements that support the response.
Prepare evidence adapted to the interlocutors
A customer seeks to understand what he is buying and what he remains responsible for. An employee wants to know what tasks are assigned and who to turn to. A journalist questions the risks, incidents and the consistency of commitments. An authority expects precise information within the framework of its competences. The factual basis must remain common, even if the depth and channel vary.
The communication file can include a usage sheet, the list of checks carried out, the incident reporting circuit, the known limits and the schedule of the remaining steps. A test result should not be presented as valid in all contexts. Human supervision must be described concretely: who can interrupt the system, with what information and within what operational time frame?
The aim is to enable verification. If an indicator is published, explain what it measures, on what sample and on what date. If a commitment concerns training, distinguish between those registered, trained and actually evaluated. These details are more useful than an accumulation of internal labels.
Fictional example: a recruitment tool and a revised calendar
Let’s imagine a company preparing to deploy a recruitment support tool. After a regulatory change, its management plans to suspend bias testing and user training. The question of communication is then inseparable from the decision: what risks does the technical postponement allow to remain, and what measures does the company consider necessary regardless of the deadline?
A prepared speech explains that the regulatory schedule has been re-examined for this use, then presents the work maintained and the conditions for authorization of the deployment. It indicates the pilot's limits and the device for recovering from a difficulty. It does not infer from a change of date that the system is secure, nor from a successful test that it is legally compliant with all the regulations.
This example is fictitious. The classification of a real system and the corresponding requirements must be established with the relevant specialists. The proposed method aims to prevent the public message from being broader than the analysis on which it is based.
Organize a word that remains accurate after publication
The first useful deliverable is a register of authorized formulations: sentence, usage covered, proof, validator, date of review. It must supply the site, commercial responses, management speeches and internal documents. A significant change triggers a review of the materials, not just a new legal notice.
For a first exchange, bring together the inventory of priority uses, the messages already broadcast, the available qualification and recurring questions from customers or teams. LLM advisory support helps to connect uses, governance and explanation. Public affairs makes it possible to monitor the institutional context and formulate a substantiated contribution. Credibility depends on a simple discipline: each assertion must be able to find its scope and its proof. [1] [2]
Sources and benchmarks
Sources consulted on October 7, 2026. Fictional examples are noted in the text; the analyzes and methods proposed are those of Belief System.