An email announces the arrival of an artificial intelligence assistant. A video promises to free up time. Then everyone returns to their screens with the same questions: can I share this client file with it? Will my manager know I am using it? And if the tool does part of my work, what will happen to my job? The challenge is not to make employees enthusiastic, but to give them the means to understand, choose and discuss. Looking ahead to September 2026, this distinction is likely to matter more than the quality of a launch campaign.
The misunderstanding begins with the promise
Since generative assistants began spreading in late 2022, demonstrations have often followed the same script: text produced in seconds, a draft presentation, an instant meeting summary. The effect is spectacular. But the demonstration can easily gloss over the context: preparing the right information, checking the response, correcting errors and taking responsibility for the result.
Within a company, this gap between promise and practice becomes a communication problem. Presenting AI as a tool that “takes care of repetitive tasks” sounds reassuring. But those tasks still need to be identified, their role in learning understood, and what will fill the time they occupied clarified. For someone starting out, writing a first meeting summary may be precisely how they learn the job.
Credible communication therefore begins by distinguishing three things: what the tool can do, what the company allows and what it plans to transform. Mixing the three fuels misunderstandings. A technical capability is neither permission to use it nor an organizational decision.
Teach judgment, not just prompt writing
Training cannot be reduced to a collection of formulas for getting better answers. Generative models can produce plausible but false content, omit essential information or reproduce biases. Their fluent writing can sometimes make these flaws difficult to spot, especially in an area the user knows little about.
The right starting point is a workplace scenario. A sales team can compare a manually written email with a generated draft, then identify exaggerated promises. A human resources department can examine why summarizing a job application requires caution. Internal communications teams can test a draft of a sensitive announcement and identify what the machine fails to understand about employee relations.
These exercises must also teach people when not to use AI. For some tasks, the tool offers no benefit once verification is taken into account. For others, the data is too sensitive or the consequences of an error too serious. Knowing when to say “not here” is a skill, not resistance to change.
A requirement that goes beyond training
In the European Union, the AI Act, which entered into force in August 2024, established a framework for this issue. The provisions of Article 4 on AI literacy have applied since February 2, 2025. They require providers and deployers to take measures to ensure a sufficient level of AI literacy among the people concerned, taking into account their knowledge, experience and the context of use. This is not a blanket certification requirement for all employees.
For companies, the practical implication is clear: a general presentation is no substitute for tailored support. A manager who approves an output, an employee who uses an assistant and a team that configures a system do not have the same needs.
Rules that help when doubt arises
Policies fail when they remain abstract. “Protect confidential information” offers little help to someone hesitating over a partially anonymized contract. A useful rule connects a category of data, a specific tool and a permitted action. It also explains whom to ask and what to do after a mistake.
A clear foundation can be built around a few guidelines:
- Approved tools: which ones to use, with which account and for which activities.
- Data: what can be shared, what must be removed and what remains prohibited.
- Validation: who checks facts, sources and sensitive outputs.
- Transparency: when to disclose an AI contribution and how to report an incident.
These guidelines must reflect the contracts and settings actually in place. Claiming that a business tool never reuses data would be unwise without checking the applicable terms. Conversely, banning all uses indiscriminately can encourage covert use of personal services, outside the organization’s safeguards.
Internal communications teams benefit here from working with IT, security, human resources and legal departments. Their role is not to disguise complexity, but to make it manageable: an up-to-date reference page, concrete examples and a clearly identified channel for answers.
Discuss jobs without promising the impossible
Questions about jobs do not disappear behind questions about the tool. Research published notably by the International Labour Organization in 2023 highlighted that exposure to generative AI could lead to the transformation of tasks rather than the full automation of jobs. But a broad trend is not an individual guarantee. Everything depends on decisions about investment, organization and how gains are distributed.
Promising that “no one will be replaced” without a firm commitment undermines trust. A more honest approach spells out the decisions made, the scenarios under consideration and the unknowns. If the objective includes cost reductions or restructuring, the language of personal fulfillment must not serve as a smokescreen.
This discussion requires involving employee representatives in accordance with applicable obligations, but also those who understand the work as it is actually done. Who will take over complex cases? How will junior employees learn? Will the time saved become time for better-quality work, or lead to higher targets? These questions belong to management and dialogue with employees and their representatives, not a promotional FAQ.
Measure more than logins
An adoption dashboard can show many activated accounts and very little value created. Logins say nothing about the quality of the final result, the verification workload or employees’ sense of control. They may even reward unnecessary use if teams come to see using the tool as proof of being up to date.
It is better to examine a few activities: is the result better? Does the time required decrease once corrections are factored in? Are errors detected? Do employees know whom to turn to? These assessments should focus on uses and the conditions surrounding them, rather than becoming disguised individual surveillance.
What next? Looking ahead to September 2026, the likely spread of assistants integrated into software could make AI less visible without making it any less transformative. The priority would then be to establish ongoing learning: revising rules, training new hires, discussing changes to tasks and publishing lessons from trials. Trust will not come from mandatory enthusiasm, but from the ability to understand, challenge and shape the transformation.