
AI can gather for a manager what is scattered: logs done or missing, overdue tasks, open incidents, questions left unanswered, instructions the team has not seen yet. Every item must carry its origin, because a recorded fact, a declared fact and an inference are not worth the same. And some things are never compiled: judgments about a person, rumours, sensitive data, emotions.
Thursday, 3:10 p.m., at the pass
Lunch service is over. The dish machine is still running. Benoît, the manager, sets the tablet down on the stainless steel at the pass and calls over Halima, the cook who runs the cold station. Tomorrow he holds the weekly team huddle. He wants to start from something solid.
He asked: "What's been left open in the kitchen this week?" The list fits on one screen.
- Temperature logs: done every day, except Wednesday evening for the cold-station prep fridge.
- Task "clean the hood filters": due Monday, still open.
- Incident open since Tuesday: a cracked hotel pan, reported with a photo.
- A question asked on Monday by a commis about the new tartare garnish, with no written answer.
- The updated tartare recipe card: published Monday, not yet read by two people on the evening team.
Halima reads it and points at the second line. "I told Gaël about the hood on Tuesday, we had to wait for the cleaning product." Benoît adds it as a comment on the task: a declared fact, dated, with its author. For Wednesday's log, nobody remembers. The list simply says so: no data.
What stands out is what is not there: no "who's doing a bad job", no "the evening team is less serious". Just facts with their source, and the gaps. Benoît knows what goes on tomorrow's agenda. This scene is fictional.
Why managers struggle to pull the facts together
Every fact lives somewhere different
Logs are on the tablet, tasks in the app, incidents in a notebook or a message, questions in the heads of whoever asked them. Gathering it all means looking everywhere. So people end up looking only at what is within reach.
Memory keeps the latest event
Without a list, a manager mostly remembers yesterday, or whatever annoyed them. A quiet week with one incident on Friday turns into "a bad week". A dated list corrects that bias.
Facts and impressions get mixed up
"Nobody ever cleans the hood" is an impression. "The hood task has been open since Monday, with an explanation given on Tuesday" is a fact. When the two get mixed, the team conversation turns defensive.
What AI can usefully gather
- Logs done or missing. Temperatures, opening and closing checklists, goods-in checks. A missing log is information in itself.
- Overdue tasks. With their due date, the person assigned and the comments. When someone has explained the delay, the explanation stays attached to the task.
- Open incidents. What was reported, when, with what proof, and what has been done since.
- Unanswered questions. A question about a recipe or a procedure that was never settled always comes back at the worst moment.
- Unread instructions. An updated recipe card the team has not seen yet is not being applied yet.
- Scheduling facts. Recorded absences, late arrivals and no-shows, gaps between scheduled and worked hours. These are facts about the schedule, not about someone's worth.
Where each fact comes from
A fact is only worth something if you know where it comes from. On the Trust and data page, Tsuno sets out six kinds of fact and never mixes them up.
- Recorded fact: logged without anyone typing it, like a timestamp or an action the tool itself recorded.
- Declared fact: someone says it is done. Useful, but declared.
- Attached proof: a photo, a document, a delivery note tied to the fact.
- Observation: what a manager saw and noted, with its date, its author and its origin (seen first-hand, reported, written message).
- AI inference: a suggested link, flagged as such, never presented as a fact.
- No data: nothing was recorded, and the tool says so instead of filling the gap.
When a compilation is fragile, it must say that too: missing data, conflicting sources, declared facts with no proof, or too little history to call something a pattern.
The facts at a glance
| Type of fact | Usual source | What you can do with it | What you must not do with it |
|---|---|---|---|
| Temperature log done or missing | Recorded or declared on the tablet | Chase the log, check the unit, review the checklist | Conclude that a person is careless |
| Incomplete checklist | Declared, with the time | See which items get skipped, and when | Rank the closing teams |
| Overdue task | Recorded, with comments | Revisit the due date, ask what is blocking it | Show it without the explanation given |
| Open incident | Declared, often with attached proof | Decide on corrective action, call a repair technician | Look for someone to blame before the cause |
| Unanswered question | Declared or written message | Write or fix the recipe card, answer the team | Hold the question against whoever asked it |
| Unread instruction | Recorded | Bring it up at the pre-shift briefing, check it is clear | Discipline someone for not reading it |
| Absence or lateness | Recorded or manager's observation | Adjust the schedule, talk to the person | Draw a conclusion about their motivation |
| AI inference | Suggested link | Ask a question, check on the floor | Treat it as an established fact |
| No data | Nothing recorded | Say what is missing, get it logged | Read it as "all good" |
What AI must not compile
Gathering work facts is useful. Building a file on a person is not, and it is often unlawful.
- Judgments. "Unreliable", "not motivated", "bad attitude" are not facts. They have no place in a compilation, even if a manager wrote them in a message.
- Rumours. Something heard second-hand, with no date or author, does not become a fact because a tool filed it. A reported observation must say it is reported, and by whom.
- Sensitive data. Health, pregnancy, disability, religion, opinions: these are special categories of data under Article 9 of the GDPR. Tsuno refuses to record them about an employee and offers to note dated, observed facts instead.
- Emotions. Under EU law, Article 5 of Regulation (EU) 2024/1689, the AI Act, has banned inferring a person's emotions in the workplace since 2 February 2025, except for medical or safety reasons. A tool should not infer someone's mood, motivation or loyalty from their messages.
In France, the Labour Code adds a simple rule: no information concerning an employee personally may be collected by a system that has not been brought to their attention beforehand (Article L1222-4). Facts are compiled with tools the team knows about, not behind its back.
Three practical uses
Preparing a team huddle
Start from the week's list, remove what has been settled, keep three topics. Each topic is a dated fact, and the conversation is about what happens next: who does what, by when.
Preparing a one-to-one
For an individual conversation, gather the facts for the period: training signed off, agreed goals, hours, incidents, actions decided last time. The manager adds their own observations and leads the conversation. The method is set out in how to prepare an employee conversation with facts, not impressions.
Preparing the Monday review
For an operator running several locations, the same compilation, site by site, shows where to spend time this week. See Monday at 8 a.m.: what a multi-site operator really needs to see.
Common mistakes
- Compiling everything that exists. A fifty-line list prepares nothing. Three to five topics are enough for a huddle.
- Stripping the explanation from a delay. A fact without its comment becomes a reproach.
- Mixing kinds of fact. A reported observation presented as a recorded log ruins trust in the whole list.
- Showing the compilation to the whole team unfiltered. One person's scheduling facts are discussed with that person, not in public.
- Forgetting what is missing. A missing log is a topic, not a blank to skip.
What to measure
- The share of huddle topics that rest on a dated fact rather than an impression.
- The number of questions left unanswered for more than a week. It should drop once they become visible.
- How long open incidents take to close, over a month.
Where Tsuno fits in
Tsuno answers a manager's questions from the location's facts: checklists, temperature logs, tasks and due dates, incidents, the logbook, absences and lateness recorded from the Today page, training signed off. Answers can show the facts used and where they come from, and say when data is missing. You can ask from inside Tsuno or from the AI assistant you already use, with the same permissions. Knowing whether each person has read an instruction, or whether a question is still waiting for an answer, is something a good tool should make just as easy to find.
Tsuno San, the intelligent voice assistant for teams, is the way in on the floor: that is often where reports and questions start. San Voice is available in demo. The details are on the Tsuno San page.
The AI neither scores nor ranks employees. The app does show a monthly leaderboard, made of points earned through training, punctuality and completed tasks: it is never passed to the assistant, and the AI does not use it. Tsuno brings the facts together. People decide.
Further reading
For the four roles of AI for managers, read the pillar article AI for frontline managers: analyse, advise, reassure. The line between interpreting and judging is covered in interpreting without judging: the limit of AI in people management. On incidents that keep coming back despite reports, see why restaurant incidents keep repeating, and for the doctrine, AI should not judge your employees.
Key takeaways
AI is useful to a manager when it gathers dated work facts, each with its origin, and says what is missing. It compiles no judgments, rumours, sensitive data or emotions: it prepares the conversation, and the manager leads it.
Frequently asked questions
What information can AI gather for a restaurant manager?
Dated work facts: logs done or missing, overdue tasks, open incidents, questions left unanswered, instructions not yet read, recorded absences and late arrivals. Each one with its date, its author and where it comes from.
What is the difference between a recorded fact, a declared fact and an AI inference?
A recorded fact is logged without anyone typing it, like a timestamp. A declared fact is someone saying it is done. An inference is a link suggested by the AI: it must be flagged as such and never presented as a fact.
What should AI never compile about an employee?
Judgments about the person, rumours, sensitive data such as health, pregnancy, disability, religion or opinions, and emotions. Under EU law, inferring emotions in the workplace has been banned since 2 February 2025.
Can you prepare a performance review with AI?
You can ask it to gather the dated facts for the period: training signed off, agreed goals, hours, incidents. The manager reads them, adds their own observations and leads the conversation. The AI does not write an appraisal.
Can a list of facts compiled by AI be used to discipline someone?
No. A compilation is there to prepare a decision, not to make it. Any decision that goes against an employee remains a human act, based on checked facts and a conversation with the person.