A button appears in the messaging app. It offers to summarise a meeting, draft a reply or find information. Then comes the message from management: artificial intelligence will “free up time”. All well and good. But can you share a client file with it? Who can see the conversations? Will employees be expected to produce more? As work resumes in September 2026, this is the real challenge for internal communications: replacing spectacular promises with credible guidance that also addresses limitations and trade-offs.
The facts cited here are documented through August 2025; the outlook for September 2026 is analysis, not an assessment of developments that have already taken place. One trend was nevertheless already under way: with Microsoft 365 Copilot, Gemini for Google Workspace and ChatGPT’s business offerings, generative AI assistants were becoming part of everyday tools. Their presence lent urgency to a long-standing question: how do you explain a change when its effects remain uncertain?
The wrong starting point: promising time savings without discussing the work
The productivity pitch seems simple: fewer repetitive tasks, more added value. It begins to falter when you look at an actual working day. An automated summary can speed up reading but omit a crucial caveat. A first draft can get the writing process moving but require lengthy checks. The gains depend on the role, the document, the user’s experience and the expected standard of quality.
Research published before August 2025 already called for this nuance. A field study in customer support, conducted by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, showed uneven benefits, particularly pronounced among less experienced workers. Other research on consulting tasks highlighted an uneven boundary: assistance helped with some exercises but could worsen results when participants trusted it beyond its capabilities.
The editorial implication is straightforward: do not promise universal gains. Instead, describe a specific task, the conditions for success and the checking required. “Prepare a summary for review” describes a use case. “Reinvent our efficiency” explains neither what to do nor when to stop.
Publish rules people can use tomorrow morning
A twenty-page policy is not enough if nobody knows what to do with a sensitive attachment. Internal communications teams must produce a short document, accessible from within the tool and organised around workplace situations. The goal is not to simplify everything: it is to make the right decision possible when it needs to be made.
- What is permitted: approved tools, required business accounts and covered tasks.
- What requires precautions: personal data, internal documents and content intended for a client.
- What is prohibited: information that must never be shared in certain environments and decisions that must not be delegated.
- Who to contact: a designated contact for questions, incidents and exceptions.
These rules need examples. Can a salesperson reword an anonymised proposal? Can an HR manager summarise one-to-one interviews? Can a developer submit proprietary code? The answer depends on the tool, the contract, the data and the purpose. An approved example is better than blanket permission to “experiment with good judgement”.
Confidentiality: avoid reassuring but incomplete answers
“Your data is protected” does not answer the questions that matter. Employees need to understand whether their inputs are used to train models, how long they are retained, which administrators can access logs and which service providers are involved. A distinction must also be drawn between a consumer service and a business offering: their settings and contractual commitments may differ.
Another pitfall is assuming that an assistant connected to internal files only saves time. It can also make documents with overly broad access permissions much easier to find. Communications will not fix poor access management. They must explain this risk and point to a reporting procedure, without asking employees to become cybersecurity experts.
Transparency also applies to usage monitoring. Does the company measure adoption overall or track individual practices? Can prompts be examined following an incident? Will they be used to assess performance? These purposes must be explicit and governed by a clear legal framework. Ambiguity fuels fears of a new surveillance tool.
Teach people to verify, not just to prompt effectively
A successful demonstration sparks interest. It does not provide training. Learning to write a prompt is useful but insufficient: people must know how to spot a fabricated claim, trace a source, identify an omission and decide that a task should remain outside the tool. Training must also demonstrate failures, using realistic examples without sensitive data.
The European framework supports this requirement. Article 4 of the AI Act, applicable since 2 February 2025, requires providers and deployers to take measures to ensure a sufficient level of AI literacy among the people concerned, taking account of their knowledge and the context of use. It is not simply a requirement to impose a standard module or a single certification.
In practice, a role-specific workshop can have participants compare a human response with a generated draft, then examine the corrections required. The right metric is not just training attendance: it is the ability to recognise an error and apply the relevant rule. That requires time allocated within the working day, not a video to watch between urgent tasks.
Explain what any gains will change
The most sensitive question often comes after the presentation: what will the company do with the time saved? Improve quality, absorb a growing workload, shorten turnaround times, reorganise teams? Internal communications cannot, on their own, guarantee that there will be no impact on jobs. They can, however, insist that known decisions be distinguished from assumptions.
Managers need a shared framework here: the objectives of the rollout, the limits of the commitments they can make and how to escalate difficulties. Employee representatives must have a role in line with applicable obligations and the changes being considered. Presenting every reservation as resistance to progress would shut down precisely the dialogue the organisation needs.
For rollouts in 2026, a credible approach would be to publish regular updates on the uses retained, those abandoned and the problems encountered. Measuring quality, corrections, workload and teams’ experience would help avoid confusing logins with benefits. Acknowledging that a trial is inconclusive can do more to build trust than another enthusiastic testimonial.
What next? The most plausible scenario is that assistants will become ever more integrated into everyday tools, making them less conspicuous as a novelty. Internal communications will need to keep pace without becoming invisible themselves: updated rules, context-specific training and frank answers about how work is changing. Maturity will not be measured by the number of employees convinced that AI is magic, but by the number who know when to use it, when to check it and when to do without it.