Meet Tsuno at Sirha Lyon 2027 · Stand S.57Learn more
EN FR
Try the demo
Menu

Team and management

Interpreting without judging: the limit of AI in people management

AI can bring facts together and offer a cautious reading of a situation. It must stop before judging a person: that call belongs to the manager.

Published on

In an empty restaurant dining room, a manager prepares a one-to-one, an open notebook and a tablet on the table

AI can help a manager with two things: bringing facts together ("three missing temperature checks on Sunday nights this month") and offering a cautious reading of the situation ("the Sunday night shift is often uncovered"). It has to stop before the third step: a judgment about a person ("Kevin is careless"). This limit comes down to the context the machine cannot see, the bias in the data, EU law and the trust of the team. The tool can line up the facts. The decision stays with the manager.

Tuesday, 3:30 p.m., in the empty dining room

Lunch service is over. Mélanie, the general manager, sits at a table by the window with her notebook and the tablet. She is seeing Kevin, a line cook for eight months, on Thursday for his one-to-one.

On the screen, the month's facts. Three missing temperature checks, all on Sunday nights. One incomplete closing checklist, also a Sunday. Otherwise, completed training modules and two shifts he picked up at short notice.

The first sentence that comes to mind is easy: "Kevin doesn't do his checks." Mélanie writes it down, then crosses it out. She looks at the schedule. On Sunday nights, Kevin has closed the kitchen alone since Didier, the sous-chef, moved to days. Nobody gave him the closing sheet that was updated in September. And the tablet on the hot line was replaced twice this month.

She writes something else: "On Sunday nights, closing depends on one person, without the current sheet." On Thursday, she will open with a question: "How are Sunday nights going?"

(Scene and characters are fictional.)

Why three steps and not two

A fact can be checked

"Three missing checks on Sunday nights in October" can be verified. You open the history, see the dates, find out a check may have been done on paper. Even when it is wrong, it can be corrected.

An interpretation can be debated

"The Sunday night shift is often uncovered" is a reading. It is about a situation, a time slot, an organization. It may be wrong, which is exactly why it must be presented as a hypothesis, with the facts behind it. AI can offer this reading as long as it labels it as one.

A judgment is imposed

"Kevin is careless" is no longer about a situation but about a person. You cannot check it; you believe it or not. Written by a tool, it looks objective, and it follows the person from one review to the next. That is the step AI must not take.

Why AI has to stop before the judgment

The context it cannot see

The tool only knows what was entered. The shift swap agreed during the break, the broken tablet, the sheet that was never handed over, the cancelled bus, the personal problem: none of it is in the data. A manager can ask. AI can only fill the gap with what is likely.

The bias in the data

Team data is not neutral. Whoever closes alone builds up more gaps than whoever works in pairs. The person who logs everything honestly looks worse than the one who ticks boxes without checking. A judgment drawn from those numbers often measures the shift and the honesty of the logging, not the person.

EU law

Two texts set clear limits under EU law.

  • GDPR Article 22 gives everyone the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects or similarly significantly affects them. A disciplinary step, a refused promotion or a change of role decided on the strength of a score alone falls into that zone.
  • The AI Act (Regulation (EU) 2024/1689, Annex III, point 4) classifies as high-risk the AI systems intended for employment and workers management, including those used to make decisions on promotion or termination, to allocate tasks based on individual behavior, or to monitor and evaluate the performance and behavior of people at work. These obligations apply from December 2, 2027: Regulation (EU) 2026/1744, published in the EU Official Journal on July 24, 2026, pushed back the original date of August 2, 2026. Article 5 of the AI Act has already banned, since February 2, 2025, inferring a person's emotions in the workplace, except for medical or safety reasons.

In France, the Labor Code adds that in companies with 50 or more employees, the works council (CSE) must be informed and consulted before introducing means of monitoring employees' activity (Article L2312-38). Other countries have their own rules on workplace monitoring. For a specific case, ask an employment lawyer or your data protection officer.

The team's trust

A team will accept a tool that reminds it of a procedure. It will not accept a tool that judges it. As soon as people believe a machine is writing an opinion about them, they start ticking boxes for the tool instead of doing the work. The gaps vanish from the screens, not from the kitchen.

The rules for AI that interprets without judging

  1. Start from dated facts, with where they come from: recorded, declared, observed, backed by proof. A fact without a date is no use in a one-to-one.
  2. Interpret situations, not people. The subject of the sentence is a station, a shift, a procedure: "Sunday closing", not "Kevin".
  3. Present the reading as a hypothesis, and say what supports it. "All three gaps fall on Sundays, when one person closes."
  4. Say what is missing. "I don't know whether the closing sheet was handed over." Missing data is information.
  5. Suggest a question rather than a conclusion. "Do you want to check who closes on Sundays, and with which sheet?"
  6. Leave the decision to the manager, and every action in two steps: the tool prepares, the person confirms.

Fact, cautious interpretation, forbidden judgment

Fact (verifiable)Cautious interpretation (a hypothesis about the situation)Forbidden judgment (about the person)
Three missing temperature checks on Sunday nights this monthThe Sunday night shift is often uncovered"Kevin is careless"
Four late arrivals of 10 to 15 minutes on 7 a.m. openingsThe 7 a.m. openings have a transport or timing problem"She's unreliable"
Allergen module not completed, scheduled three weeks agoTraining never found a slot in the schedule"He doesn't care about the job"
Two incomplete closing checklists after services of more than 120 coversClosing runs over when service is busy"He cuts corners"
Few messages in the team channelNo reading: message volume says nothing about the work"She's disengaged"
Two shifts picked up at short noticeThe schedule relies on a few volunteers"He's the best on the team"

The last row matters as much as the others: a positive judgment is still a judgment. Thanking someone for two specific covers beats handing out a title.

Common mistakes

  • Sliding from the situation to the person. "Sunday is uncovered" becomes "Kevin doesn't do Sundays" on the second pass. Check the subject of every sentence.
  • Treating missing data as a failure. A check that was not logged is not always a check that was not done. Ask.
  • Letting a score sum up a person. A count of actions can motivate a team; it does not replace a one-to-one.
  • Assuming human sign-off is enough when it is automatic. A manager who approves without reading what the tool suggests is not really deciding.
  • Not telling the team what the tool does. A tool whose purpose nobody knows quickly turns into surveillance.

What to measure

  • The share of one-to-ones prepared from dated facts rather than impressions.
  • The number of gaps fixed by changing the organization (schedule, sheet, training) rather than by blaming someone.
  • The number of facts corrected by employees themselves before a one-to-one: a sign that both sides can speak.

Where Tsuno fits in

Tsuno's doctrine fits in one sentence: Tsuno brings the facts together, people decide. The app prepares one-to-ones from dated facts: completed training, checklists done or not, temperature checks, recorded absences, late arrivals and missed tasks, agreed tasks and goals. It separates recorded facts, declared facts, a manager's observations, attached proof and missing data, and an inference made by the AI is flagged as such.

Tsuno's AI does not score an employee in place of their manager, does not predict who will leave, and does not infer mood or motivation. It refuses to record an employee's health, pregnancy, disability, religion or opinions, and offers to log dated, observed facts instead. The app does show a monthly leaderboard, made of points earned through training, punctuality and completed tasks, to keep teams engaged. It is never passed to the assistant: the AI does not use it and does not rank employees, and San, the team voice assistant, does not assess anyone. Every action happens in two steps, the tool prepares, the person confirms, and it is logged. The full framework is on the trust page, and the assistant on the Tsuno San page.

Further reading

For the big picture of how AI supports managers, read AI for frontline managers: analyse, advise, reassure. The core position is set out in AI should not judge your employees, and its HR version in software recalls the facts, people decide: a doctrine for AI in HR. On how this can drift, see why management software can quickly feel like surveillance. For the conversation itself, how to prepare an employee conversation with facts, not impressions. And for the assistant's limits, what San never does. The general framework is set out in the pillar page AI, voice and privacy at work. To see what the tool can prepare without judging, read what AI can gather for a manager.

Key takeaway

Useful AI in management brings facts together and offers readings of situations, presented as hypotheses. The judgment about a person stays with the manager, because only the manager can see the context, ask the question and answer for the decision.

Frequently asked questions

What is the difference between a fact, an interpretation and a judgment in management?

A fact is dated and verifiable: three missing temperature checks on Sunday nights this month. An interpretation offers a reading of a situation: the Sunday night shift is often uncovered. A judgment labels a person: he is careless. AI can help with the first two, never with the third.

Can AI assess an employee's performance?

It should not do so in place of the manager. Under EU law, Article 22 of the GDPR restricts decisions based solely on automated processing that significantly affect a person, and the AI Act classifies AI systems used to monitor and evaluate workers' performance and behavior as high-risk.

When do the EU AI Act rules on high-risk AI in employment apply?

Obligations for the high-risk systems listed in Annex III, which includes employment and workers management, apply from December 2, 2027, after the postponement adopted in 2026 (Regulation (EU) 2026/1744). The ban on inferring emotions in the workplace has applied since February 2, 2025.

Why can't AI see the full context of a team situation?

Because it only knows what was entered. A shift swap agreed verbally, a broken tablet, an instruction that was never passed on or a personal difficulty appear in no data. The manager can ask the question.

Does Tsuno score or rank employees?

The app shows a monthly leaderboard, made of points earned through training, punctuality and completed tasks, to keep teams engaged. It is never passed to the assistant: Tsuno's AI does not use it, does not score or rank employees, and San does not assess anyone.