
Workplace AI is best judged by what it refuses when someone important asks it to cross a line. Our red lines rest on a few simple principles: serve the work rather than watch the person, collect nothing behind anyone's back, stick to facts, let a human decide. Those principles only count if they hold up against a real dilemma. Here are four, and how we settle them.
Tuesday, 11:15 pm, the dining room is empty
The chairs are upside down on the tables. Mathieu and Salomé, partners in three restaurants, have kept one table at the back, two coffees and a stack of annotated sheets. They are talking about the assistant their teams have been using for a month.
Mathieu has an idea: "If I knew who asks the most questions, I'd know who isn't trained." Salomé puts her pen down: "Or who dares to ask. And the one who doesn't dare, how would you spot them?"
They go quiet for a moment. Mathieu crosses out a line on his sheet. The question he wanted answered was fair: where is training missing? The way he had in mind to answer it was not.
(Fictional scene and characters.)
Why a list of refusals is not enough
A refused feature always comes back through another door
We have already written what San never does, feature by feature. But a list cannot tell you how to answer the request that is not on it yet. For that, you need the principles that produced the list.
Hard requests come from reasonable people
The dilemma rarely comes from a malicious boss, more often from a manager who means well or a group that needs to steer its sites. So the decision has to be made beforehand, calmly, and in writing.
The law sets a floor
In France, the Labour Code requires prior notice to employees and proportionality for any device (Articles L1222-4 and L1121-1, covered in the guide AI, voice and privacy at work). Those texts say what is forbidden. They leave open what a tool should refuse to offer even where the law would allow it.
The principles behind our red lines
- Serve the work, not the surveillance of the person. A feature is judged by its purpose: helping someone do the right task, or producing a view of someone.
- Nothing hidden from the person concerned. Whatever a tool keeps about an employee, that employee should be able to know.
- Facts, never a verdict. Tsuno's doctrine fits in one sentence: the tool brings the facts together, people decide. A machine does not know a person's context.
- The same permissions, no more. San acts within the permissions of the person using it, and every action goes through a logged confirmation.
- Look at the process before the culprit. When something does not hold, look at training, the rota, the instruction.
- Say what is not built yet. A rule announced too early destroys more trust than it creates.
Four dilemmas, and how we settle them
The hundred-site group that wants to open up the history
Say a group of a hundred locations tells us: "We sign if our regional managers can read the questions each employee asks." The request has a logic: spotting sites where procedures are misunderstood.
Our position is to refuse. The questions an employee asks should not turn into a file on them. A commis who knows the regional manager will read their hesitations will stop asking, and the tool will lose the one thing that makes it useful. What we offer instead serves the real need: knowing which sheets are missing, which are badly written, which stations keep coming up in unanswered questions. That is what a good tool should show, and it is what we are working on. Keeping the question history closed to managers is a principle we are building: we will only call it done once it is.
The manager who wants to know who asks the most
This is Mathieu's question. It runs into a simple problem: the number of questions does not measure skill. It measures the station, the season, the quality of the sheets, and the courage to ask. The new hire who asks twenty questions in the first week is doing exactly what you want.
The real need is to know where training is missing. Dated facts already answer it: modules validated, modules still to go, stations where nobody is trained. Tsuno Compétences is there to track them. A per-person counter would add nothing except fear.
The monthly leaderboard, which exists, and why the AI stays out of it
It would be easy to write that Tsuno measures nothing. That would be false. The app shows a monthly leaderboard made of points earned through training, punctuality and completed tasks, to keep teams engaged. It counts actions done, visible to everyone.
We own that, with a clear boundary: the leaderboard is never passed to the assistant. The AI does not use it, and San assesses no one. Here is why. A count of actions shown in the app stays what it is. The same count in the hands of an AI that drafts, summarizes and answers would quickly become an appraisal: "He's near the bottom of the leaderboard, you might want to talk to him." That is exactly the verdict the doctrine refuses.
The temptation to measure engagement
The most appealing request of all. A gauge per employee, green or red, fed by late clock-ins, unticked tasks, how often someone messages, maybe the tone of their voice. On voice, the EU AI Act has already settled it: inferring an employee's emotions has been banned since 2025, as we explain in AI and emotions at work.
For the rest, no law would stop us. We refuse anyway, because an engagement score mostly measures the organization (a rota that chains a close and an open, an instruction never written down) and pins it on the person. We go into this in AI should not judge your employees.
Our decision grid
When a new request comes in, we ask three questions, in this order.
- Who is served? The task at hand, or a view of a person?
- Would the person concerned know, and would they accept it if the feature were explained out loud, at the pass, in front of their colleagues?
- Can the real need be met without looking at anyone in particular?
If the answer to the first is "a view of a person", or the second is no, we do not build it. The third question is how we find what to offer instead.
| Dilemma | What we are asked for | Principle at stake | Our answer | What we offer instead |
|---|---|---|---|---|
| Hundred-site group | Read each employee's questions | Nothing hidden, serve the work | No | Spot missing or badly written sheets (in progress) |
| Worried manager | Know who asks the most | Facts, not a verdict | No | Track training validated and still to validate |
| Monthly leaderboard | Use it to talk about people | People decide | Never passed to the assistant | A count of actions, visible to all, inside the app |
| Engagement | A gauge per employee | Proportionality, AI Act Art. 5 | No | Look at the process: rota, instruction, training |
What we choose not to sell
Tsuno does not score a person in place of their manager, predict who will resign, recommend who to dismiss or infer someone's mood or motivation from their messages. No sanction goes out without a person confirming it, and San assesses no one. These refusals are already listed on our Security and trust page.
A manager view of the questions asked to San is a separate case. Keeping it out is a principle we are building, as explained above, and we will only present it as done once it is. Some buyers will want these features and may find them elsewhere.
A frontline assistant only helps if people dare to use it. When you bring an assistant into a kitchen, teams rarely worry first about noise or battery life: they worry about being listened to, then judged. A surveillance feature, even one used once a year, proves them right, and a customer whose teams stop asking questions has nothing left to gain from the tool.
Common mistakes when you write your own red lines
- Thinking a switched-off option is enough. A surveillance feature that exists, even turned off, will be switched on one day in a crisis. At Tsuno, these refusals are neither a setting nor a module: the tool is built not to do them.
- Making an exception for the big customer. The first exception becomes the rule, and every employee of every customer pays for it.
- Mixing up aggregated and anonymous. On a site with four people, "the evening team's questions" points at someone.
- Writing principles without cases. A principle that has never met a real request has not been tested yet.
What to measure
- The calls you have made: each time a request touched a principle, write down the decision and the reason in a log you keep up to date.
- Your team's objections when the tool is introduced, noted word for word, and how they change after two weeks.
- Voluntary use, station by station. A team that feels watched stops asking.
Where Tsuno and San come in
These principles rest on what Tsuno already does. San acts within the person's permissions, and every action is prepared, summarized, confirmed and logged. Tsuno refuses to record sensitive data about an employee and refuses to rate a person in place of their manager, predict a resignation or infer a mood. The monthly leaderboard exists in the app and stays out of the AI's reach. What is still being built, such as a question history closed to managers, we say so plainly. The assistant is presented on Tsuno San.
Further reading
For the full picture, read the guide AI, voice and privacy at work. For the stance that follows from it, see a companion, not a supervisor, and for the other side of our honesty, the promises we will never make. The HR doctrine is laid out in software recalls the facts, people decide.
Key takeaways
A red line only counts if it holds when an important customer asks you to cross it. We settle each case with three questions (who is served, would the person know, is there another way) and we accept not selling what would turn an assistant into a supervisor.
Frequently asked questions
What does ethical AI at work actually mean?
It means an AI whose limits hold when someone makes a specific request: an important customer, a manager under pressure, a vendor keen to close a deal. It serves the work, it does not watch the person, it relies on dated facts and it leaves the decision to a human.
Can an employer find out which employee asks an assistant the most questions?
In France, no personal information about an employee may be collected by a device they were not told about beforehand (Labour Code, Article L1222-4), and any restriction of their freedoms must be proportionate to the aim (Article L1121-1). Our position is that a per-person question counter has no place in a workplace assistant.
Does Tsuno rank employees?
The app shows a monthly leaderboard made of points earned through training, punctuality and completed tasks, to keep teams engaged. It counts actions done. It is never passed to the AI assistant, the AI does not use it, and San assesses no one.
Can AI be used to measure employee engagement?
Inferring a person's emotions in the workplace has been banned in the EU since February 2, 2025, except for medical or safety reasons (Regulation (EU) 2024/1689, Article 5). Beyond emotions, an engagement score blends the organization with the person, and ends up blaming the person for the organization's flaws.
What does Tsuno do when a customer asks for a feature that breaks these principles?
We say no, we explain why, and we look for the real need behind the request. There is almost always a way to meet it without looking at one particular person.