
An AI connected to your operations should never carry out an action on its own. It prepares, sums up the effect in a line or two, and a person reads it and confirms. We call this the two-step rule: it catches misunderstandings, wrong people and wrong locations before they reach the team. In Tsuno, every confirmed action is logged and permissions are checked again at the last moment.
Tuesday, 11:05 a.m., an hour before service
Hélène, head chef, is planning the week from the tablet at the pass. She asks her assistant: "Put Amadou on dish duty Thursday night, instead of Rémi."
The assistant prepares and shows: "Schedule for Amadou at Lyon Confluence, Thursday, November 19, 6:00 to 11:00 p.m., dish station, published to the team. Rémi's shift on this station is removed."
Hélène stops at "November 19". She meant the following Thursday, the 26th, the day Rémi has booked off. She corrects it, reads the new version and confirms. Ten seconds of reading saved Amadou a wasted trip this week and an empty dish station the next.
(Scene and characters are fictional.)
Why AI should not act without confirmation
It can misunderstand
A request made between two tickets, over the noise of the extractor hood, is often incomplete. "Thursday" can mean this one or next. "Evening" can start at 5 or at 7. An AI fills the gaps with what seems most likely. Most of the time it gets it right. The rest of the time, nobody notices until the day itself.
It can pick the wrong person
Two Amadous on the team, a Rémi in the kitchen and another front of house: mix-ups are ordinary. When a name is ambiguous, a good tool asks you to clarify instead of guessing. But even a name it recognized correctly is worth reading in full.
It can pick the wrong location
As soon as you run more than one site, the location becomes the first source of error. Names look alike, staff rotate, a manager covers two addresses in the same week. A location named by the person should never be swapped for another with a similar name.
Some actions are hard to undo
Removing a shift, notifying a team, publishing a schedule: once it is out, the message has been read. You can fix it, but trust has already taken a hit. The harder an action is to reverse, the more the read-through matters.
How a two-step action works
In Tsuno, whether you go through San or through your usual AI assistant (ChatGPT, Claude, Mistral), an action always follows the same path.
- The AI prepares the action from what the person asked for.
- It summarizes the effect: what will change, for whom, where.
- The person reads and confirms. From an AI assistant, confirmation comes in a new message: the original request does not count as approval of a summary nobody has seen yet.
- The action is carried out, then written to the log: who, when, what, with the state before and after.
Two safeguards complete this path. First, the AI only acts within the person's permissions: it can only prepare what that person could already do in Tsuno. Second, permissions are checked again at the moment of confirmation: if an access was removed between the proposal and the "yes", the action does not go through.
If the person changes anything, the AI makes a fresh proposal with a new summary. You never confirm a version you have not read.
What stays out of reach, even with confirmation
Confirmation does not make everything acceptable. Some actions carry too much weight to go through a conversation, whoever says yes.
In Tsuno, payroll, invoicing, permanent deletion and legal acts stay out of the AI assistant's reach, even when confirmed. These go through the dedicated screens and the people authorized to handle them, with time to think.
Another limit has nothing to do with confirmation: if someone asks to record an employee's health, pregnancy, disability, religion or opinions, Tsuno proposes nothing. Under Article 9 of the GDPR, these are sensitive data. It offers to log a dated, observed fact instead.
The two-step rule also echoes a wider principle of EU law. Article 22 of the GDPR protects people from decisions based solely on automated processing that significantly affect them. In operations as in HR, Tsuno's doctrine is the same: the software brings the facts together, people decide.
The two traps of confirmation
"Yes, yes" on autopilot
The first risk comes less from the AI than from habit. After three weeks of accurate summaries, people start confirming without reading. Confirmation becomes a button you press to make a window go away. That is exactly when the mistake slips through.
Adding more windows would not help. Better to make the summary so short that people actually read it, and to highlight what is out of the ordinary: an unusual location, a date far ahead, a removal.
The summary that is too long
The second trap is the opposite. A ten-line summary listing every technical field does not get read in the middle of service. People skim it and confirm anyway. Too much information has the same effect as none.
A good summary fits on one or two lines. It answers four questions: what, for whom, where, when. It names the person and the location in full. And if there is a side effect (a shift removed, a team notified), it says so in one sentence.
Good summaries, in five examples
| Request | Expected summary |
|---|---|
| "Put Amadou on dish duty Thursday night." | Schedule for Amadou at Lyon Confluence, Thursday, November 26, 6:00 to 11:00 p.m., dish station, published to the team. |
| "Create a task to check the walk-in cooler seal." | Task "Check the walk-in cooler seal", closing team, Lyon Confluence, due tonight at 11 p.m. |
| "Report the leak under the dishwasher." | Report: leak under the dishwasher, dish station, Lyon Confluence, now. |
| "Note that Justine was absent yesterday." | Dated fact for Justine: absent from the evening shift on November 16, called in at 4 p.m. No reason recorded. |
| "Schedule Rémi's review for Tuesday." | Review meeting for Rémi with Hélène, Lyon Confluence, Tuesday, November 24, 3:00 to 3:30 p.m., visible to Rémi. |
Notice what keeps coming back: the full location name, a date with its weekday, a start and end time, who can see it. Nothing technical, nothing that sends you to another screen to understand it.
Common mistakes
- Confirming the request instead of the summary. "Yes, that's what I asked for" means nothing if you have not read what the AI understood.
- Accepting a summary with no location. With more than one site, a summary that does not name the location is incomplete.
- Letting the AI choose between two people with the same name. If the name is ambiguous, the person should clarify.
- Believing confirmation covers everything. Payroll, invoicing, permanent deletion and legal acts stay outside the conversational path.
- Never looking at the log. The record is there to understand a mistake afterwards, not just to reassure.
What to measure
- The share of summaries corrected or rejected before confirmation. If it drops to zero for weeks, ask whether people still read before concluding the AI is perfect.
- Corrections made after a confirmed action. Each one tells you which part of the summary was not read.
- Reading time. Beyond a few seconds during service, the summary is too long.
Where Tsuno and San fit in
The two-step rule is not specific to San: it applies to every AI action in Tsuno, including from your usual assistant. The AI prepares and summarizes, you confirm, the action is logged with the state before and after, and your permissions are checked again at the moment of the "yes". San Voice, the voice interaction, is currently available as a demo: it is the best way to hear a summary read aloud and then confirmed. The Tsuno San page introduces the assistant, and Trust and data walks through the path of an action.
Further reading
Confirmation only makes sense if the AI works within your permissions, no more and no less: that is the subject of give your AI the same rights as you, no more. To understand how an AI is connected to a business, read MCP explained for frontline trades. The assistant's other limits are gathered in what San never does, and the doctrine in the software recalls the facts, people decide. For the big picture, see the pillar article Tsuno San: the work companion for frontline teams.
Key takeaways
An AI that acts alone will eventually get the person, the location or the date wrong, and nobody will notice before service. Two short steps, a readable summary and a "yes" given with full knowledge are enough to keep people in control.
Frequently asked questions
Can an AI change a schedule or create a task without my approval?
In Tsuno, no. The AI prepares the action and summarizes its effect, and nothing goes out until the person confirms. The confirmed action is written to the log, with who did it, when, and the state before and after.
Why not let AI handle simple tasks on its own?
Because AI mistakes in operations are rarely dramatic: the wrong first name, the wrong location, the wrong date. They go unnoticed until the next shift. Confirmation is the moment when someone who knows the floor catches them.
What happens if my permissions change between the proposal and the confirmation?
In Tsuno, permissions are checked again at the moment you confirm. If an access was removed in the meantime, the action does not go through.
What can an AI never do, even with confirmation?
In Tsuno, payroll, invoicing, permanent deletion and legal acts stay out of the AI assistant's reach, even when confirmed. These go through the dedicated screens and the people authorized to handle them.
What does a good confirmation summary look like?
It fits on one or two lines and answers four questions: what, for whom, where, when. It names the person and the location in full and flags anything that will be hard to undo.