Skip to content
All articles
Law & regulation 8 min read· by Lars Zimmermann

AI coach rates employees: training or performance monitoring?

An AI practice coach that grades sales calls evaluates work performance. Why the works council has a say today and what applies from December 2027.

In short

An AI coach that grades employees in practice calls against fixed criteria evaluates work performance. In Germany it is already subject to co-determination under Section 87(1) no. 6 BetrVG and may fall under the EU AI Act’s high-risk annex from 2 December 2027. Four rules help: ratings stay with the employee, nothing enters HR systems, written purpose, works council first.

Auf Deutsch lesen: deutsche Fassung

An AI service provider sells a practice coach for sales. The employee conducts a sales call with a simulated customer. If he makes a mistake, the AI interrupts him and explains what would have been better. At the end there is a rating against fixed criteria, such as need established, objection handled, close attempted. Points per criterion, plus an overall grade.

I like the idea. Practising without a real customer is cheap, and nobody embarrasses themselves in front of colleagues. I played football up to the Oberliga Bremen, a German regional league. Training only works as long as you are allowed to make mistakes there. That is exactly where the coach has a problem. It looks like training. Legally it is a performance evaluation, and different rules apply to that.

The same vendor says publicly that human in the loop is outdated. One should insert an agent that checks whether the human’s work is correct. That turns oversight upside down. What this means for the coach comes further below. First the question every managing director should ask first: what does the thing actually do with the data?

What the coach does with every call

Technically, such a coach works in essentially the same way everywhere. It records the voice, transcribes the call, compares it with the criteria and awards points. The points are attached to a user account. The account is attached to a name. After three months there is a series of grades for every employee, neatly sorted by date.

That makes the coach a case of profiling. Article 4(4) GDPR understands this to mean any automated processing that evaluates personal aspects relating to a person. Performance at work and behaviour are expressly listed there as examples. The EU AI Act adopts exactly this term (Article 3(52) of the EU AI Act). Whether the customer is real or simulated makes no difference. It is the person speaking who is evaluated.

It gets delicate in the admin settings. If there is a team view for managers there, it shows who has practised how often and who is weak on which criterion. Check whether this view can be switched off. Anyone who switches it on has turned a training tool into a performance register, whether they intended to or not.

One function should be switched off immediately. If the coach offers to read nervousness or lack of enthusiasm from the tone of voice, that is no longer a question of co-determination. Since 2 February 2025, Article 5(1)(f) of the EU AI Act has prohibited AI systems that infer the emotions of a person in the workplace. Exceptions exist only for medical or safety reasons. Sales training is neither. More on this prohibition is in my article on covert employee profiling.

The works council already has a say today

Section 87(1) no. 6 of the German Works Constitution Act (BetrVG) gives the works council a right of co-determination over technical devices that are designed to monitor the behaviour or performance of employees. The word “designed” invites a false conclusion. Many read it like this: our coach is designed for practising, so it is not covered. The Federal Labour Court has interpreted the provision differently for decades. It is enough that the device is objectively suitable for monitoring performance or behaviour. What the employer intends to do with it is irrelevant to this question.

A coach that stores points per person is the textbook case. It may not be introduced without the consent of the works council or a ruling by the conciliation committee. In practice that means: a works agreement before the first employee practises. The pilot with five volunteers from sales also runs with real employees and real grades. I would not start it without the works council.

With or without a works council, the GDPR applies. Article 5(1)(b) requires specified, explicit and legitimate purposes. A coach whose purpose nobody has written down has no specified purpose. Anyone who later does use the points in the annual appraisal is processing them for a new purpose. That purpose must be compatible with the old one. In my view, an appraisal rarely fits a purpose that used to be called practising.

From December 2027: high-risk even without an HR decision?

The EU AI Act lists employment as a high-risk area in Annex III, point 4. Point (b) names three cases. The first concerns decisions on terms of employment, promotions and terminations. The second the allocation of tasks based on behaviour or personal traits. The third stands on its own.

“… to monitor and evaluate the performance and behaviour of persons in such relationships.” Regulation (EU) 2024/1689, Annex III, point 4(b)

This third case does not require an HR decision. Going by the wording, evaluating performance is enough. What remains open is whether pure practice feedback that nobody but the employee sees falls under it. I am not aware of any ruling on this. I would not rely on it falling outside.

Many know the way out via Article 6(3) of the EU AI Act. Under it, a system from Annex III is not considered high-risk if it does not pose a significant risk, for instance because it only performs a narrow procedural task or merely prepares an assessment. For the coach this door is closed. The same paragraph says in its last subparagraph: a system from Annex III that performs profiling of natural persons is always considered high-risk. A coach that grades named employees carries out profiling. See above.

When this takes effect is governed by Article 113(c)(i) as amended by Regulation (EU) 2026/1744. The obligations for high-risk systems under Annex III apply from 2 December 2027. That sounds far away. But a coach introduced today will have collected a good 13 months of grades by then. Today’s works agreement has to withstand tomorrow’s situation.

What is decisive here is what the provider intended the coach for. Article 3(12) of the EU AI Act expressly counts promotional and sales materials as part of the intended purpose. If the brochure contains a sentence such as “rates your employees against fixed criteria”, that is the intended purpose. So read the brochure the way a supervisory authority would read it.

Whoever repurposes the coach becomes a provider themselves

The most delicate point lies in your own company. Suppose the provider has worked properly, built the coach as a pure training tool without a team view and not classified it as high-risk. The head of sales nevertheless exports the points into a spreadsheet every month and discusses them in the annual appraisal. From that moment the tool is being used for something it was not built for.

Article 25(1)(c) of the EU AI Act has a clear consequence for this case. Anyone who modifies the intended purpose of an AI system that was not classified as high-risk in such a way that it becomes a high-risk system is considered a provider themselves. They then bear the provider obligations under Article 16. These range from the quality management system through the conformity assessment to CE marking and registration. For a mid-sized company that only wanted to practise, that is a project nobody budgeted for.

The spreadsheet is the moment that matters. No provider sees it. No contract forbids it. In the audit I therefore do not ask about the coach at all at first. I ask where its figures flow to.

The agent that checks the human

Back to the vendor’s sentence that human in the loop is outdated. A second agent is supposed to check whether the human’s work is correct. For high-risk systems the EU AI Act provides for the opposite direction. Article 14(1) requires that such systems can be effectively overseen by natural persons. Article 26(2) obliges the deployer to assign this oversight to people with the necessary competence, training and authority. The human stands above the system. An agent that monitors the human does not replace this oversight. It is a further system that itself has to be overseen.

For the coach this means, in very practical terms: who checks whether the AI rates correctly? An example. In the practice call an employee refuses a discount because the margin does not allow it. The coach scores that as “objection not handled”. Professionally the answer was right. If there is no way to challenge the rating, the employee learns the wrong thing. And with a grade attached.

Built properly, exactly this is the strength of the tool. A coach whose criteria the head of sales defines and regularly checks against real calls makes a team better. People practise more often because it costs nothing and nobody is watching. This second condition is the core. If it falls away, people practise for the grade and no longer for the customer.

Four rules under which the coach may stay

Steering instead of blocking means here: four rules that are in writing before the first call.

  • The rating stays with the employee. Managers see at most whether practice took place, and team values only in groups in which nobody can be singled out by calculation.
  • It does not go into the HR system. No export, no interface to the personnel file, recordings and points are deleted after a fixed period.
  • The purpose is fixed in writing. Training, nothing else. That is purpose limitation under Article 5(1)(b) GDPR and intended use under ISO/IEC 42001, A.9.4.
  • The works council is involved beforehand. The works agreement is in place before the first employee practises (Section 87(1) no. 6 BetrVG). From 2 December 2027, the information obligation under Article 26(7) of the EU AI Act is added for high-risk systems.

The four rules do not make the tool free of obligations. They make it negotiable. A works council that is shown that the points do not leave training has a reason to agree. A works council that is shown a team view with a ranking has none.

As an auditor I want to see evidence on this, not declarations of intent. The permission setting showing that managers see no individual values. The disabled export. The deletion log. And an impact assessment for the employees under ISO/IEC 42001, A.5.4: what consequences does the coach have for the person being rated? Pressure, fear of mistakes, grades that do turn up somewhere after all. Anyone who strikes A.5.4 from the Statement of Applicability has to justify that (ISO/IEC 42001, 6.1.3). For a system that grades people, I cannot think of a good justification.

Three questions for the provider before you sign. Where are the points stored, and who sees them? Can the team view be switched off, and is it logged when someone switches it back on? And what does your promotional material say about the purpose? The answer to the third question helps decide the classification.

One question remains that only you can answer: who in your company would notice if the points from training turned up in the annual appraisal?

Share: LinkedIn E-Mail

Frequently asked questions

Is an AI coach for sales calls subject to co-determination?+

As a general rule, yes. Section 87(1) no. 6 of the German Works Constitution Act (BetrVG) covers technical devices designed to monitor performance or behaviour. Under the case law of the Federal Labour Court, objective suitability is enough. A coach that stores points per person is suitable, even if it is sold as training.

Is an AI practice coach a high-risk AI system?+

Possibly. Annex III, point 4(b) of the EU AI Act names the evaluation of the performance and behaviour of workers without requiring an HR decision. If the coach grades named employees, that is profiling. The exemption under Article 6(3) is thereby barred. The high-risk obligations apply to it from 2 December 2027.

May the coach analyse nervousness or enthusiasm in the voice?+

No. Since 2 February 2025, Article 5(1)(f) of the EU AI Act has prohibited AI systems that infer the emotions of a person in the workplace. Exceptions exist only for medical or safety reasons. Sales training falls under neither.

Is it enough if only the employee sees their rating?+

It is the strongest lever, but not a free pass. The works council remains competent because the technical possibility of evaluation exists. Record the setting in the works agreement, switch off export and team view, and make sure that every change to them is logged.

Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)

Last updated: 28 October 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.

Continue reading