
An AI connected to your HR tools must never record an employee's health, pregnancy, disability, religion, opinions or union membership. Under EU law, these are special category data (GDPR, Article 9). Nor should it turn an impression ("he looks tired") into a fact. What it can do is offer to log a dated, observed fact, which the manager reads and confirms.
Monday, 4:20 pm, in the back office
Lunch service is over. Agathe, the floor manager, sits down in the small office behind the storeroom. She opens the AI assistant she uses every day, connected to the restaurant's management software, and types fast, without much thought:
"Add to Simon's file that he's always off sick and looks worn out lately."
The reply comes back: "I don't record information about an employee's health. I can log dated facts: Simon was absent for the evening shift on November 3 and November 17, and he called in at 3 pm both times. Shall I prepare that note?"
Agathe closes the tablet. She takes Simon's paper file out of the drawer and rereads her own notes from last year: "not motivated", "trouble at home?", "sick again". She realizes that the rule the AI just applied works for her pen too.
(Scene, characters and reply are fictional, for illustration only.)
Why a connected AI changes the question
Connecting an AI to your tools hands it a pen
Once MCP links it to business software, an assistant such as ChatGPT, Claude or Mistral does more than answer: it reads records and prepares entries. A sentence typed at the end of a shift can then become a line in a file. The connection itself is explained in MCP explained for frontline trades.
AI writes fast, and it writes well
A rushed manager scribbles "sick again" on a sticky note. A polite AI may rewrite it as "repeated absences, probably linked to a health issue". It sounds professional, which is exactly the problem: it adds a medical guess dressed up as an established fact.
What is written stays, and gets reread
A remark said out loud fades. A note can be read by the next manager, quoted in a review, or requested by the employee. That is why the law governs what you write down, not just what you think.
What the law says, in plain words
The GDPR prohibits some data by default
Article 9 of the GDPR prohibits processing data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs or trade union membership, as well as genetic data, biometric data, health data and data about a person's sex life or sexual orientation. There are narrow exceptions (for example, when a law requires it), but none of them covers a manager who wants to "keep a record". Pregnancy, disability, illness and addiction all fall under health.
In France, the Labour Code asks for justification and transparency
In France, Article L1121-1 of the Labour Code says restrictions on people's rights and freedoms at work must be justified by the task and proportionate to the goal. Article L1222-4 adds that no personal information about an employee may be collected by a system they were not told about beforehand. And Article L1132-1 bans any discriminatory measure based on, among other things, health, disability, pregnancy, religious beliefs, political opinions or union activity.
The CNIL asks for relevant, non-excessive notes
The CNIL, France's data protection authority, published a reference framework on HR data (deliberation no. 2019-160 of November 21, 2019). It sets a simple rule for free-text comment fields: they must only hold relevant, non-excessive information, and the employee can access them. Writing as if the person will read it is therefore a rule, not just good manners.
And an AI does not decide on its own
Article 22 of the GDPR gives everyone the right not to be subject to a decision based solely on automated processing when it significantly affects them. A sensitive note written by an AI, then reused in a decision, is exactly what that article is meant to prevent.
The table of risky requests
Here are six requests a tired manager might make at the end of a shift. None is malicious, and yet every one of them is a problem.
| A manager's request | Why it is a problem | What the AI offers instead |
|---|---|---|
| "Note that Simon looks tired lately." | An impression, not a fact, and it touches on health | "Simon arrived at 9:40 am on November 4 and 6; his shift started at 9 am." And a suggestion to talk to him |
| "Write down that Leïla is pregnant, for the rota." | Health data (GDPR, Art. 9); hers to announce; a protected ground for discrimination | "Scheduling constraint given by Leïla on November 10: no closing shifts until further notice." |
| "He's always sick, put it in his file." | A diagnosis and a generalization | The absence dates, the shifts missed, when he called in, the paperwork received |
| "She's fasting for Ramadan, give her fewer evenings." | An assumed religious belief, and a decision made because of it | If the employee asked: "Request from the employee on March 2: finish before 7 pm until March 30." No reason given |
| "Note that Bruno is in the union, he'll cause trouble." | Union membership plus a judgment about a person | Nothing. If a specific work event happened, describe it with its date |
| "He drinks, I can smell it in the morning." | A guess about addiction, so about health | "On November 14 at 8:15 am, Simon dropped two trays; the head chef kept him off the slicer for the morning." Safety is handled right away, by a person |
The right-hand column softens nothing; it makes the note sturdier: a dated fact can be checked, discussed and defended, which an impression cannot.
Absences: a dated fact yes, a diagnosis no
This is the most common case, and the easiest place to slip. An employee misses three shifts in a month. The manager needs to know for the rota, and sometimes to talk about it. They do not need to know what illness it was, and in France they have no right to write it down.
What you can log:
- The date and the shift missed ("Friday, November 3, evening shift").
- When the person called in, and how.
- Any paperwork received ("sick note received on the 4th").
- The effect on service, if useful ("covered by a temp").
What you do not log: the illness, a guess at the cause, a view on frequency ("too often"), a word about their private life. In France, the copy of the sick note sent to the employer does not mention the illness anyway. If the manager knows more because the employee confided in them, that confidence stays a conversation, not a line in a file.
If the absences cause a real scheduling problem, the right next step is a one-to-one, prepared with those dated facts. We cover the method in how to prepare an employee conversation with facts, not impressions. On health itself, the right people to talk to are the employee and, if needed, occupational health.
What a connected AI should do, in five rules
- It refuses to record sensitive data, even when the request is reworded, softened or framed as a report.
- It says so in one sentence, without lecturing.
- It immediately offers a dated, observed fact, when there is one.
- It prepares nothing that the person has not read and confirmed.
- It draws no conclusion about the person from their facts: it shows them, the manager decides.
The second rule matters more than it looks. A refusal that lectures pushes the manager to write somewhere else. A short refusal followed by a useful suggestion helps them do better.
Common mistakes
- Thinking an "internal" note is read by no one. The employee can request their data, and files change hands when a manager leaves.
- Rewording to get past the rule. "He's got issues" or "complicated personal situation" are still guesses about health or private life.
- Mixing the fact and the cause. "Absent on the 3rd" is a fact. "Absent on the 3rd, probably exhausted" is not.
- Keeping a separate paper file where the rules supposedly do not apply. They apply to paper too.
- Letting the AI write on its own. A note nobody reviewed is a note nobody really wrote.
What to measure
- The share of notes about employees that carry a date, an author and an observable fact. The target is simple: all of them.
- The number of requests the AI refused in a month, and what they reveal. If they keep coming back, a manager often needs help with a situation, not a more flexible tool.
- A quarterly reread of free-text comments, with one question: could the employee read this line without being hurt or surprised?
Where Tsuno and San come in
Tsuno refuses to record an employee's health, pregnancy, disability, religion or opinions. It offers to log dated, observed facts instead. In Tsuno, an observation is what a manager saw and recorded, with its date, author and source: something seen directly, something reported, or a written message. Tsuno also refuses to cross-reference sensitive data to profile someone, or to guess a person's motivation or mood from their messages.
These rules hold when you query Tsuno from your usual AI assistant. The assistant stays within the person's scope, and a proposed note is only written once it has been read and confirmed, with a record kept. The Trust and data page lists what Tsuno refuses, and the Integrations page explains the connection with ChatGPT, Claude or Mistral. Tsuno San, the smart voice assistant for teams, follows exactly the same rules: see Tsuno San.
No tool replaces a manager's restraint. The AI can refuse a line. It cannot stop someone thinking "sick again". It can only help them write something better.
Further reading
To understand what connecting an AI to your tools means, start with MCP explained for frontline trades, then read give your AI the same rights as you, no more. On the doctrine, see AI should not judge your employees and software recalls the facts, people decide. And for all the limits of the team assistant, what San never does. The pillar page AI, voice and privacy at work puts these limits in context, and the AI at work glossary for restaurant owners explains the terms used here.
Key takeaways
An AI connected to your HR tools must never record an employee's health, religion, opinions or union membership, or turn an impression into a fact. It offers a dated, observed fact, the manager reviews it, and the manager decides.
Frequently asked questions
Can an AI assistant record that an employee is sick or pregnant?
It should not. Under the GDPR, health and pregnancy are special category data (Article 9), and processing them is prohibited except in specific cases. Tsuno refuses to record them and offers to log a dated, observed fact instead.
What counts as sensitive data under the GDPR?
Article 9 of the GDPR covers data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs and trade union membership, as well as genetic data, biometric data, health data and data about a person's sex life or sexual orientation.
Can a manager write down that an employee looks tired?
That is an impression, not a fact, and it touches on health. It is better to note what was seen, with the date: a late arrival, a mistake, a station left. And to talk to the person.
How do you log an employee's absences without recording their health?
Write down the date, the shift missed, when the person called in and the paperwork received. Do not write down the illness or an opinion about how often it happens. A dated fact yes, a diagnosis no.
Can employees read what is written about them?
Yes. The GDPR gives everyone a right of access to their data, and the French data protection authority (CNIL) says free-text comment fields must only hold relevant, non-excessive information. Write every note as if the employee will read it.
What does an AI connected to Tsuno do if asked to record sensitive data?
It prepares no entry at all, even if the request is reworded. It says so in one sentence and offers to log a dated, observed fact, which the person reviews and confirms.