"Sentiment analysis" is a standard bullet on the feature list of most conversation-intelligence tools. It is also, depending on how it works, a practice the EU banned outright more than a year ago. The EU AI Act prohibits AI systems that infer the emotions of a person in the workplace, and unlike the transparency and high-risk rules that arrive in 2026 and 2027, this prohibition has been in force since 2 February 2025 with penalties live since August 2025. So the honest answer to "is sentiment analysis banned emotion recognition at work?" is: sometimes yes, and the deciding factor is not the label on the feature but the input it reads. This is where the line falls.
If a tool infers an employee's emotional state from the biometric characteristics of their voice (tone, pitch, prosody), that is emotion recognition in the workplace and it is prohibited under Article 5(1)(f) of the AI Act, in force since 2 February 2025. If a tool analyses what was said (topics, keywords, whether a deal sounds like it is progressing), that is ordinary content processing and generally sits outside the ban, though it still carries GDPR and, in Germany, works-council duties. The input decides it: how a voice sounds versus what the words say.
What Article 5(1)(f) actually bans
Article 5 of the AI Act is the list of prohibited practices, the reddest of the red lines. Point (f) prohibits putting into service or using "AI systems to infer emotions of a natural person in the areas of workplace and education institutions," with narrow exceptions where the system is used for medical or safety reasons (a fatigue-detection system for driver safety is the standard example). The bright line is the workplace: the same emotion-inference system that might be permitted in some consumer settings is banned when pointed at employees.
The reasoning is set out in Recital 44, and it is unusually blunt for a legal text. The EU legislator flags that these systems rest on a weak scientific basis: emotion inference has limited reliability, limited specificity, and does not generalise well across people and cultures, which makes it both unreliable and intrusive. The Act does not ban emotion recognition at work because it works too well. It bans it because it does not work well and gets used to make decisions about people anyway.
The hinge word is "biometric"
The definition is what makes this answerable rather than vague. Article 3(39) defines an "emotion recognition system" as an AI system "for the purpose of identifying or inferring emotions or intentions of natural persons on the basis of their biometric data." The load-bearing phrase is on the basis of their biometric data. Emotion recognition, in the Act's sense, is not any guess about how someone feels. It is specifically an inference drawn from biometric signals.
And a person's voice is biometric data. Voice recordings carry the physical and behavioural characteristics that let a system identify or characterise the speaker, which is what puts voice inside the biometric category under both the GDPR and the AI Act. So a feature that listens to how a sales rep's voice sounds on a call, its tone, pitch and cadence, and outputs "frustrated," "confident," or a sentiment score, is inferring an emotion from biometric data. In the workplace, that is the prohibited practice, not a grey area.
Where the line falls for call tools
The category markets "sentiment" as one word, but two very different mechanisms hide under it.
| What the feature reads | Example | Status at work |
|---|---|---|
| Voice biometrics (tone, pitch, prosody) to infer the speaker's emotion | "Rep sounded anxious in the last third of the call" | Prohibited (Art. 5(1)(f), since 2 Feb 2025) |
| Facial expression analysis to infer emotion (video meetings) | "Detected frustration from facial cues" | Prohibited (Art. 5(1)(f)) |
| Text of the transcript for topic, intent, deal signal | "Pricing was discussed; next steps agreed" | Generally outside the ban; GDPR still applies |
| Text sentiment of what was said, not tied to a person's biometrics | "The call was positive about the proposal" | Not biometric emotion recognition; boundary can blur if used to judge the individual |
The practical test for a buyer is one question to the vendor: does your sentiment feature read the sound of the voice or the content of the words? If the answer is that it models vocal tone or facial expression to score an employee's emotion, you are looking at a prohibited practice and no contract clause fixes it. If the answer is that it classifies the text of what was said, you are in ordinary-processing territory, with the usual GDPR obligations rather than an outright ban.
One honest caveat, because the honest answer has an edge case: text-based sentiment is not automatically safe. If a system uses the written content specifically to infer and record an individual employee's emotional state and feed employment decisions, regulators may still read that against the spirit of the prohibition, and it certainly triggers GDPR fairness and, in Germany, co-determination duties. The clean position is to analyse the deal and the conversation, not the person's feelings.
Consent does not save it
A common instinct is to reach for consent: put a line in the employment contract, have staff click agree, carry on. That does not work here. Article 5 practices are prohibited, not conditionally permitted, so there is no consent gateway. The only doors out are the medical and safety exceptions, and a sales-coaching use case is neither. This also aligns with a long-standing GDPR position: consent in the employer-employee relationship is rarely valid because the power imbalance makes it hard to call freely given. For a workplace emotion feature, consent is not a fix, it is a red flag that someone has misread the rule as a restriction rather than a ban.
This is live risk, not a future deadline
The reason this matters now rather than later: the prohibited-practices chapter of the AI Act applied from 2 February 2025, and enforcement with penalties has been live since 2 August 2025. Breaching a prohibited practice sits at the top penalty tier, up to 35 million euro or 7 percent of worldwide annual turnover, higher than the tier for transparency or high-risk breaches. So while much of the AI Act debate is about 2026 and 2027 dates, workplace emotion recognition is the one part that is already fully enforceable. If a tool in your stack scores employees' emotions from their voice, the exposure is present tense. We mapped the full set of what applies when in our explainer on what the AI Act actually requires on 2 August 2026 versus what moved to 2027.
What to do about it
- Ask the input question. For every call or meeting tool you use, get it in writing whether "sentiment," "emotion," "tone" or "mood" features derive from voice or facial biometrics, or from the text of what was said.
- Turn off voice or facial emotion scoring of employees. If a feature infers an employee's emotion from biometrics, disable it for workplace use. This is a prohibition, not a configuration preference.
- Do not rely on consent. Remove any assumption that an employment-contract clause or an opt-in makes emotion recognition lawful. It does not.
- Keep content analysis clean. Topic detection, action items and deal signals from the transcript are fine under the AI Act, but still document your GDPR basis and, in Germany, involve the works council before rollout.
- Write it into procurement. Add a line to your vendor questionnaire: "Does the system infer emotions of natural persons from biometric data?" A yes for a workplace tool is a disqualifier.
Where Numi sits, honestly
Numi does not infer employees' emotions from the sound of their voice or from facial expression. It transcribes meetings and analyses the content: topics, questions asked, commitments made, and the signals in what was actually said. That is a deliberate design choice, not a gap we plan to close. A workplace tool built on voice-based emotion inference would be a prohibited practice in the EU, and we were not going to build a core feature on a foundation the law bans. When we say a call was "positive," we mean the substance of what was discussed, drawn from the words, not a read on how anxious a rep sounded.
That is the same discipline we apply across the product: analyse the work, not the worker's physiology. It keeps our customers on the right side of Article 5, and it reflects what we think meeting intelligence is actually for, which is understanding the conversation, not surveilling the people in it.
This article is general information about the EU AI Act as of 31 July 2026, not legal advice. Whether a specific feature constitutes prohibited emotion recognition depends on its technical design and use; confirm against the final legal text and your own counsel before acting. Sources: EU AI Act (Regulation 2024/1689), Article 5(1)(f), Article 3(39) and Recital 44; the Act's application timeline (prohibited practices from 2 February 2025, penalties from 2 August 2025).