Emotion inference and the AI Act: the line is clear, and it is easy to hold

The European artificial intelligence regulation drew a clear line between two things that used to be conflated: measuring what was said, and guessing what someone felt while saying it. The first is allowed, and it is the one that makes a rep better. The second is prohibited. Once the distinction is set out, the subject closes.

What article 5 says

The regulation sets out a short list of prohibited practices, in its article 5. Point f covers AI systems used to infer the emotions of a natural person in the workplace or in an education institution.

The definition, in article 3 point 39, is that of a system intended to identify or infer the emotions or intentions of natural persons on the basis of their biometric data. Voice is one of those. This is the point several conversation intelligence vendors have not yet taken on board: analysing a rep's tone, energy or emotional state from their voice, in the course of their work, falls within the scope of the prohibition.

What stays entirely out of scope

The prohibition is precise, and its exact reach is narrower than it is often read to be:

• Physical states are not emotions. Pain and fatigue are explicitly distinguished from emotions in the sense of the text.
• Immediately apparent expressions are not enough. Noting a smile is not inferring an emotional state.
• Inference from written text falls outside the scope. The prohibition turns on the notion of biometric data; text is not biometric data.
• A medical and safety exception exists. It is read strictly and does not cover general wellbeing monitoring.

Put differently: counting open questions, measuring talk time, spotting that a price objection was not handled, checking that a mandatory point of the script was said, all of that stays lawful. These are facts of the conversation, not inner states inferred from a voice. And they are precisely the indicators a rep improves on, because you can show them the evidence.

The timetable, and why it catches people out

The regulation entered into force on 1 August 2024, but its provisions apply in waves. The prohibited practices in article 5 have applied since 2 February 2025, ahead of most of the rest of the text. Many buyers are still working to a 2026 or 2027 timetable; on this particular point the deadline has already passed.

The penalty regime in article 99 places breaches of article 5 in the highest band, up to 35 million euros or 7% of worldwide annual turnover. That is a reason to check your tool, not a reason to give up on conversation analysis: the lawful version exists and it is more useful.

The three questions to ask a vendor

A sales director or a compliance lead assessing a meeting analysis tool can settle it in three questions:

• Does your product infer an emotional state, a sentiment or an intention from the voice? If so, on what legal basis in the European Union?
• Are your scores calculated on observable facts from the transcript, or on an inference of tone?
• Can you contractually switch off any emotion inference for our European users, and document it?

A vendor who answers the first question with a marketing line about the rep's empathy or energy has just answered yes. A vendor who has dealt with the subject answers in one line and sends you the documentation.

The design choice behind MIHOS

MIHOS does not infer emotions. The scores cover what was said and the structure of the exchange: coverage of the playbook steps, open questions, talk time, objection handling, commitments made. Every number points back to a timestamped passage in the transcript, so it can be challenged and explained. This is not only a regulatory precaution: a rep does not accept a score you cannot show them the evidence for.