The workplace emotion recognition ban and AI call analysis

The ban on emotion recognition in the workplace is one of the few provisions of the EU AI Act that already applies in full — since 2 February 2025, with no transition periods and no Digital Omnibus postponements. At the same time, the speech analytics market is still full of tools advertising “sentiment analysis” and “emotion detection”. For a call center manager it comes down to a practical question: which features are safe to switch on, and which expose the company to the highest penalty tier in the entire regulation — up to €35 million or 7% of worldwide turnover?

The good news: the line is sharper than the marketing noise suggests. Below we take it apart — what exactly Article 5(1)(f) prohibits, what the ban does not cover, where the risk of voice-tone analytics lies, and how to vet a vendor before you sign the contract.

What the AI Act’s emotion recognition ban actually means

Article 5(1)(f) of Regulation 2024/1689 prohibits placing on the market, putting into service and using AI systems that infer the emotions of a natural person in the workplace and in education institutions. The key to its scope sits in the definition in Article 3(39): an “emotion recognition system” is a system that identifies or infers people’s emotions or intentions on the basis of their biometric data — and biometric data includes, among other things, voice, facial expressions and keystroke dynamics.

Three things are worth stressing right away. First, the ban has applied since 2 February 2025 and — unlike the high-risk obligations, postponed to December 2027 — was not covered by any deferral. Second, it catches not only the tool’s provider but also the company that uses it: “using” is a deployer’s act, so a call center cannot hide behind its contract with the vendor. Third, a breach of Article 5 carries the highest penalty tier: up to €35 million or 7% of annual worldwide turnover.

The European Commission’s guidelines on prohibited practices, approved on 4 February 2025 (formally non-binding, but a clear signal of how the Commission intends to read the provision), leave the industry no room for doubt: the examples of prohibited practices explicitly include call centers tracking employees’ emotions — anger, for instance — using voice analysis and cameras. The legislator’s reasoning is set out in recital 44: inferring emotions rests on questionable scientific foundations, because emotional expression varies across cultures and situations, and the power imbalance in the employer–employee relationship makes such systems particularly intrusive.

The exceptions: medical and safety reasons — read narrowly

Article 5(1)(f) provides an exception for systems put in place for medical or safety reasons. The Commission’s guidelines require it to be interpreted narrowly: the use must respond to a specific need, be limited to what is strictly necessary — in duration, in the people covered and in scale — and come with real safeguards.

In practice: a system monitoring signs of fatigue in professional pilots or drivers does not fall within the ban’s scope at all — recital 18 of the AI Act expressly excludes physical states such as pain or fatigue from the notion of “emotions”, so such a system needs no exception in the first place. The medical and safety exception covers strictly medical uses, e.g. therapeutic systems supporting the treatment of emotional disorders. It does not, however, legalise monitoring the team’s “stress level” or burnout risk — even with the best intentions — because that is still inferring emotions from biometric data. Nor does “safety” cover the company’s own interests — emotion detection aimed at preventing theft or fraud does not fit within the exception. The conclusion for a contact center is short: no realistic call quality evaluation scenario will fit within these exceptions.

What the ban does not cover: evaluating call content is not employee emotion analysis

This is where the most common misunderstanding lies. The ban concerns inferring emotions from biometric data — it does not concern evaluating what was said. The Commission’s guidelines confirm this explicitly: inferring sentiment from text alone — for example from a transcript of what someone said — is not an emotion recognition system within the meaning of the regulation, because it does not process biometric data.

Substantive call evaluation is even further from the ban, because it does not infer emotions at all. Checking whether the agent disclosed that the call is recorded, explored the customer’s needs with open questions, or made a promise the company cannot keep — that is an assessment of observable behaviour. It is exactly what a coach does during manual call review, only on a sample; automation does it on 100% of calls.

The takeaway from the table: automated call quality evaluation is legal — provided the system evaluates content, not emotional states, and works on text, not on the acoustic features of the voice. We mapped out the full set of obligations that kick in for employee evaluation in December 2027 in our article on the EU AI Act in the call center.

The AI Act and emotions in the call center: a customer is not an employee — but that is not the end of it

The Article 5(1)(f) ban protects employees, not customers. Analyzing a caller’s emotions is therefore not a prohibited practice — but it is not a regulation-free zone either. An emotion recognition system used outside the workplace is a high-risk system (Annex III, point 1(c)), and from 2 August 2026 you must inform the people such a system is applied to (Art. 50(3)) — the standard “this call may be recorded” announcement does not satisfy that duty.

On top of that comes a purely practical problem: a phone call recording contains two voices. A tool that “analyzes customer emotions” from audio must credibly demonstrate that the agent’s channel never reaches the prosody analysis. And if the “call sentiment” score then lands on the agent’s scorecard, the construction becomes very hard to defend. Renaming the feature “call energy” or “engagement index” changes nothing — classification is decided by what the system does with the data, not by the label in the sales deck.

How CallSea solves this technically

At CallSea, compliance with Article 5(1)(f) is not a clause in the terms of service but a property of the architecture. The scoring models receive transcript text only — they have no access to the audio, so analysis of tone of voice, prosody or any acoustic features isn’t “switched off”; it is unfeasible. Audio recordings are passed to transcription and are not stored permanently.

The second safeguard operates at the configuration level: a validator makes it impossible to save scoring criteria that infer agents’ emotions or state of mind — including criteria that would try to draw emotional conclusions from the text itself. That is deliberately broader than the letter of the ban, because a criterion like “rate the agent’s enthusiasm” is not only legally risky but also impossible to score repeatably. Instead, you score behaviours visible in the transcript: did the agent answer every customer question, address the objection, propose a next step. We publish a full description of the data architecture and the system’s classification on our Trust & security page.

Checklist: 6 questions for your call analysis vendor

Before you sign a contract for a call analysis tool, ask the vendor these questions — and put the answers into the contract, not into meeting notes:

1. Do the scoring models process audio, or transcript text only? At which stage does the recording leave the data flow?
2. Can a criterion like “rate the agent’s attitude / enthusiasm” be configured? If not — what blocks it: a clause in the contract or a validator in the code?
3. Does the “sentiment analysis” feature work on prosody and tone of voice, or on text? Whose speech does it cover — the customer’s, the agent’s, or both?
4. Do customer sentiment results feed agents’ scores or rankings in any way? If they do, you are back to the Article 5 question.
5. How does the vendor classify the system under the AI Act, and will they confirm the processing architecture in writing — so it can go into your DPIA and stand up in an audit?
6. If you plan to analyze customer emotions: how will the vendor support the transparency duty under Art. 50(3), which applies from 2 August 2026?

There is one simple warning sign: a vendor who answers the first question with generalities or “that’s a technical detail” is asking you to carry the risk of the most expensively priced prohibition in the AI Act.