People analytics: covert employee profiling with AI
AI that evaluates all team chats and calendars to assess employees is not a productivity feature. Why co-determination and data protection stop it.
In short
An AI that covertly evaluates all team chats, calendars and messages to assess employees is no harmless analytics feature. Under Section 87 of Germany’s Works Constitution Act it requires co-determination, is bound by strict employee data protection rules, and it must not infer emotions at work from biometric data such as voice or facial expression. Convenient does not mean permitted.
Auf Deutsch lesen: deutsche Fassung
Update of 20 August 2026. When I wrote this text, the Digital Omnibus had been adopted but not yet published. It is now law in force: Regulation (EU) 2026/1744 of 8 July 2026, in the Official Journal of 24 July 2026, in force since 27 July 2026. That changes nothing about the substance, but it does change how binding it is. I am leaving the text as it stands and putting the reference above it.
At trade fairs and in pitches I am shown the same tool again and again, only under changing names. It reads all team chats, evaluates calendars, measures who writes to whom and when, and turns that into a dashboard: who is at capacity, who seems unmotivated, who talks to whom, who is a resignation risk. It is sold as team health, as productivity analysis, often even as care. I call it by its name: this is covert employee monitoring with a coat of analytics paint. And the sentence that accompanies such tools is almost always the same: it is practical, it just runs in the background. That is exactly where the problem lies. Practical and permitted are two different things. This is my professional assessment as an auditor, not legal advice.
The foreman with the notebook behind every back
I come from the workshop and like to translate such things into the analogue world. Imagine a foreman who stands behind every workbench, writes down every conversation in the kitchen, notes who is on the telephone for how long, and cobbles a ranking together from it. Nobody in the company would tolerate that. You would immediately recognise it for what it is: seamless surveillance. As software in the background, nicely presented as a dashboard, exactly the same thing slips through because you cannot see the camera. The difference is the visibility, not the intrusion. The intrusion is the same, often even deeper, because the software forgets nothing and evaluates retrospectively.
What is really happening here: monitoring of behaviour and performance
A system that systematically evaluates chats, messages and calendars in order to make statements about individual employees is, in legal terms, not harmless statistics. It is processing of personal data to monitor behaviour and performance. In Germany several limits apply to this in parallel: employee data protection law, the GDPR and works constitution law. None of them is optional, and none can be overridden with the argument of convenience. Anyone who introduces such a tool must satisfy all three at the same time, not pick one.
On top of that, such a tool often comes into the company as an external cloud solution. That means: the communication of the entire workforce, in part with customer names, contract details and internal arrangements, travels for evaluation to a provider nobody has vetted. This adds a question of information security and of processing on behalf of a controller under Article 28 GDPR to employee data protection. A single tool thus opens up several construction sites at once, and none of them was planned beforehand.
The works council has a mandatory seat at the table
Where there is a works council, the most important hurdle is often the first one to be forgotten. Section 87(1) no. 6 of the German Works Constitution Act (Betriebsverfassungsgesetz, BetrVG) gives the works council an enforceable right of co-determination in the introduction and use of technical devices designed to monitor the behaviour or performance of employees. Case law interprets this broadly: it is sufficient that the device is objectively suitable for monitoring. It does not even have to have been built for that purpose.
A tool that evaluates all chats and calendars is the textbook case. Without the consent of the works council, as a rule in the form of a works agreement, such a system may not be introduced and may not be used. If it is switched on anyway, that is not just a formal error. The works council can stop the use, and the covertly obtained evaluations are often not admissible in a dispute. Anyone who installs the tool first and informs the works council later has the order the wrong way round.
Employee data protection: narrow limits instead of a free pass
Even without a works council such a system is not unrestricted. The processing of employee data needs a legal basis. In Germany that has so far been Section 26 of the Federal Data Protection Act (Bundesdatenschutzgesetz, BDSG), embedded in the GDPR. Importantly: in 2023 the European Court of Justice ruled on an identically worded provision of the state of Hesse that such general clauses do not meet the requirements of Article 88 GDPR (C-34/21). Since then a blanket reliance on Section 26 no longer holds up blindly. All the more do the basic principles of the GDPR count, and they stand in the way of an all-round evaluation.
- Purpose limitation (Article 5(1)(b)): data may only be used for the specified, explicit purpose. Evaluating chats that are intended for collaboration after the fact for performance assessment is a change of purpose that must be justified separately.
- Data minimisation (Article 5(1)(c)): only what is really necessary for the purpose is permitted. Reading every message without gaps is the opposite of minimal.
- Transparency (Article 5(1)(a) and Article 13): covert is out of the question. Employees must know what is being evaluated. Covert evaluation is conceivable only in narrowly limited exceptions where there is a specific suspicion, not as a permanent state.
- Proportionality: a permanent total evaluation of all communication is, as a rule, disproportionate in relation to the employees’ right of personality.
And one more thing is often overlooked: systematic, extensive monitoring of employees triggers a data protection impact assessment under Article 35 GDPR. The German supervisory authorities list exactly such processing on their list of cases that must be assessed. Anyone who has not carried out this assessment has already breached an obligation before the first report has even been generated.
Detecting emotions at work with AI? That is already prohibited
Many of these tools advertise sentiment analysis. They want to read from chats whether the mood in the team is tipping, who is frustrated, who is about to resign. This is exactly where the EU draws a hard line. In Article 5 the EU AI Act prohibits the use of AI systems that infer the emotions of people in the workplace and in educational institutions. There are exceptions only for medical or safety reasons. This prohibition has already applied since 2 February 2025. According to the definition in Article 3(39), what is meant are systems that infer emotions from biometric data, for example from voice, facial expression or typing behaviour. If a tool evaluates voice messages or typing behaviour, that is a prohibited practice. According to the Commission’s guidelines on prohibited practices, pure text analysis of chats does not fall under this prohibition. That is still a long way from making it permitted: data protection, co-determination and the high-risk classification under Annex III remain.
A tool that reads all chats in the background and measures people does not become permissible because it is convenient and nobody is looking. Convenient does not mean permitted.
And even an evaluation that assesses performance and behaviour rather than emotions is not classed as harmless in the EU AI Act. AI for monitoring and evaluating employees is one of the high-risk applications under Annex III. According to the amended timetable of the Digital Omnibus, the full high-risk obligations for it apply from 2 December 2027. That is no reason to sit back. Data protection and co-determination apply today, and the prohibition of emotion recognition applies today. The high-risk framework comes on top.
Even if it were permitted: the figures are of little use
There is a second reason why I view such tools critically, and it has nothing to do with legal provisions but with craftsmanship. The metrics that come out of them rarely measure what they claim to measure. Someone who writes a lot is not productive, they write a lot. A full calendar is not proof of added value, but often proof of too many meetings. The quiet colleague who works with concentration and chats little quickly ends up in the dashboard as a supposed underperformer, while the one who sends messages all day counts as committed. That is not analysis, it is confusing activity with performance.
Added to this is an effect everyone knows from practice as soon as people know they are being measured: they adapt their behaviour to the measurement, not to the work. People write in order to show activity. Honest queries are not made, because nobody wants to appear to be the one who does not know something. Criticism in the chat falls silent. A tool that monitors collaboration damages exactly the open communication it believes it is measuring. In the end there is an expensive dashboard that presents a distorted picture nicely. Anyone who derives personnel decisions from it is deciding on the basis of noise.
And even clean figures do not solve a leadership problem. If a team is overloaded, a good manager sees that in conversation, not in the message counter. If someone is dissatisfied, an honest one-to-one conversation helps more than an alert in the system. Technology does not replace leadership here. At best it conceals that leadership is missing.
What ISO/IEC 42001 requires of such a system
Anyone who wants to use AI responsibly finds the right framework in ISO/IEC 42001, the standard for AI management systems. For AI systems with effects on people it requires a documented impact assessment: what effects does the system have on those affected, here the employees? It demands a clearly defined, legitimate purpose instead of a vague let us see what the data yields. And it requires objectives for responsible use, among which it counts human oversight, as well as information for those affected.
A people analytics tool that covertly reads everything fails on each of these points: no clean definition of purpose, no real transparency, no impact assessment worthy of the name. As an auditor I do not need ten minutes to see that. I ask about the legal basis, the works agreement, the data protection impact assessment and the information given to employees. If that is missing, the finding is unambiguous. And in practice it is almost always missing.
My clear position: employee analytics can be useful if it is open, purpose-bound, agreed with the works council and limited to what is necessary. Aggregated workload of a team without looking at the individual person can help to identify bottlenecks. Covert and comprehensive, by contrast, it is a breach of trust with legal consequences. Anyone who really wants to keep their team healthy starts with transparency. That is not a weakness. It is the difference between a tool and a barn door left wide open.
Are you using AI in your company or planning to, and unsure where the limits are? Check in a few minutes whether your use of AI is properly set up before a tool touches your employees’ data.
To the AI readiness checkPrimary sources
Frequently asked questions
Is people analytics prohibited in general?+
No. What is prohibited is covert, comprehensive evaluation without a legal basis. Communicated openly, purpose-bound, limited to what is necessary and, where there is a works council, secured by a works agreement, team analytics can be permissible. Aggregated metrics without reference to individuals are considerably less critical than profiles of individual employees.
Does the use need the consent of the works council?+
Yes, if there is a works council. Under Section 87(1) no. 6 of the German Works Constitution Act it has an enforceable right of co-determination for technical devices that are objectively suitable for monitoring behaviour or performance. A tool that evaluates chats and calendars falls under this. Without a works agreement it may not be introduced and may not be used.
May an AI infer the mood of employees from chats?+
Not without further ado. Since 2 February 2025, Article 5 of the EU AI Act has prohibited AI systems that infer emotions in the workplace from biometric data, for example from voice, facial expression or typing behaviour, except for medical or safety reasons. According to the Commission’s guidelines, pure text analysis of chats does not fall under this prohibition. It remains bound by data protection and co-determination and can count as a high-risk system under Annex III.
When is a data protection impact assessment mandatory?+
For systematic and extensive monitoring of employees, Article 35 GDPR requires a data protection impact assessment before deployment. The German supervisory authorities expressly include such processing on their list. If the impact assessment is missing, an obligation has already been breached before the system generates its first evaluations.
What does an auditor check first with such a tool?+
Legal basis, works agreement, data protection impact assessment and the information given to employees. In addition, purpose limitation and data minimisation under ISO/IEC 42001 and the GDPR. If one of these elements is missing, the finding is clear. Covert all-round evaluation fails on practically every one of these points.
Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)
Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.
Questions about your own case?
In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.