AI in the contact centre is no longer a pilot topic. Three dates matter – and the most important one is already behind us.
AI in the contact centre stopped being a pilot topic some time ago – which means the question of what rules apply is no longer theoretical either. The AI Act does not land all at once; it arrives in stages. Three moments matter for service operations, and the most important one has already passed.
Since February 2025: inferring employees’ emotions is prohibited
The AI Act’s prohibitions have applied since 2 February 2025. They expressly cover inferring emotions in the workplace. The only exceptions are medical and safety reasons.
In practice this means systems that score an agent’s mood, stress or motivation during a call are no longer a grey area. They are banned. That includes features sold as welfare measures – a dashboard meant to flag overload, for example.
Analysing the customer side is a separate legal question and is not prohibited outright, but it carries obligations of its own. If both run in the same system, you need to know precisely which analysis is aimed at whom.
Since August 2026: people must know they are talking to AI
Transparency obligations have applied since 2 August 2026. Anyone interacting with an AI system must be able to recognise it. For voicebots and chat that means disclosure belongs at the start of the contact, not in a footnote on the website.
In practice this is a design task more than a legal one. A line such as “You are speaking with an automated assistant; I can transfer you to a colleague at any time” satisfies the duty and answers the question the caller was going to ask anyway.
Moved to December 2027: the high-risk obligations
Requirements for high-risk systems under Annex III were originally due in August 2026 as well. The Digital Omnibus, in force since 27 July 2026, moved them to 2 December 2027, and to 2 August 2028 for AI embedded in products. The reason was that the harmonised standards were not ready in time.
More time is not the same as nothing to do. Documentation, human oversight and logging cannot be retrofitted shortly before a deadline – they either grow with the system or they do not exist.
What this means for service operations
The first step is not a legal question but an inventory: where does AI actually touch a customer conversation? Usually in more places than expected – routing, summarisation, suggested wording, quality scoring, knowledge search.
For each of those, three questions need answering. Who is affected, the customer or the employee? Is the output suggested to a person, or does it replace their decision? And is there a route back to a human that nobody has to hunt for?
That overview is worth having even with no deadline pressing. It tends to reveal first of all where AI is in use without anyone ever having measured the benefit.
Not purely a legal matter
The disclosure duty changes how conversations are designed; the emotion-recognition ban changes how quality is managed. Both end up in process documentation and coaching, not only in a file.
This article describes the position from an operational point of view and is not legal advice. The AI Act’s deadlines have already been adjusted once – before committing to larger investments, it is worth checking the current state.
