A new feature promises to measure how engaged your team looks during video calls. Another offers a confidence score for candidates based on their facial expressions. Before trying either on staff, ask what the software is actually measuring and why.
Some workplace uses of AI are prohibited under the EU AI Act. They don't become acceptable because the feature comes with software you've already bought or because a manager plans to check its scores afterwards.
Understand the particular prohibition
For activities within the Act's scope, Article 5(1)(f) prohibits workplace and education uses that infer people's emotions, with an exception for medical or safety reasons. A general staff consent form doesn't create another exception.
The relevant definition in Article 3 concerns emotions or intentions inferred from biometric data. That makes the inputs and method important, rather than the supplier's marketing label.
The prohibitions have applied since 2 February 2025. The Commission's July 2026 Omnibus announcement explains later dates for high-risk requirements. Those extensions aren't permission to use a prohibited feature while waiting.
If you're a UK business, start by checking whether the Act applies to your activities. A UK address alone doesn't settle the scope question.
Ask what goes into the score
Take a screenshot of the feature description and ask the supplier to explain its inputs in plain language. Does it analyse facial movements, vocal characteristics, typed words or a person's answers to questions? What does the output claim to show?
The Commission's prohibited-practices guidelines distinguish biometric emotion inference from text sentiment analysis that doesn't use biometric data. They also explain that workplace coverage includes recruitment and remote working.
So reading the words in a written comment isn't automatically the same practice as inferring an employee's emotional state from their voice. Equally, a label such as “communication coaching” doesn't explain which of those the product does.
Ask for the written assessment for the exact feature and intended use. Keep an unclear feature disabled while you get the answer. A general statement that the whole product complies with the AI Act won't resolve a question about one setting.
Separate transcription from assessment of people
Consider a fictional business buying a meeting assistant. It wants draft action points, but the package also includes a manager dashboard showing each attendee's supposed enthusiasm.
Write down the difference between those purposes before anyone switches the dashboard on. The first asks what was said. The second claims something about the person saying it.
Ask whether the additional analysis can be disabled for the whole organisation. Check what ordinary users and administrators can enable separately. If the supplier can't separate the features, choose a service whose limits match the approved purpose.
Don't run a trial on real staff to find out whether the scores seem convincing. Request documentation and a supplier demonstration using fictional material first. A trial can still involve the very use you're trying to assess.
Be cautious about medical and safety claims
The exception in Article 5 is specific. Calling a product a wellbeing service doesn't establish that a particular workplace use falls within it.
If a supplier relies on the exception, ask it to identify the intended purpose, the applicable reasoning and the people qualified to assess that purpose. Get appropriate advice before using it. Don't stretch an exceptional use into ordinary staff performance monitoring.
Even where this prohibition doesn't apply, that isn't a clean bill of health. Other AI Act provisions, data protection rules and employment obligations may still matter.
For UK worker information, the ICO's monitoring guidance addresses lawfulness, necessity, proportionality and telling workers what happens. Assess those issues independently of what the AI supplier says.
Look through the tools you already own
You don't need to buy a product called an emotion detector to encounter these features. Start with your recruitment platform, call software, meeting tools and staff analytics settings.
Our suggested review is to record five things for any feature that appears to assess people:
- The feature name and whether it is enabled.
- The information it collects or analyses.
- The claim its score makes about a person.
- Who sees the result and what they might do with it.
- Who can stop it and obtain a written supplier explanation.
Follow the result into other systems. A dashboard may export a score into a recruitment record or a manager's report. Turning off a visible chart won't help if the underlying assessment still runs and gets used elsewhere.
Deal with a feature that's already running
If you identify a potentially prohibited use, pause that feature and keep its output out of decisions while you obtain advice. Note when it was enabled, who could see the results and where the information went.
Don't rush to erase everything. Ask the appropriate privacy or legal adviser what records need preserving to investigate and what should be removed under the relevant retention rules.
Review any decisions that relied on the output. For an imagined applicant shortlist, that might mean returning to the original applications and the genuine selection criteria instead of the generated scores. Decide who needs to be told and how an affected person can raise a concern.
Update your AI inventory with the decision and the feature's permitted limits. The next person renewing the licence should be able to see why a tempting toggle stays off.