
We will never promise a tool that makes no mistakes, an average success rate, certified allergens, an assistant that takes the manager's place, a guaranteed drop in staff turnover, a setup in a few minutes, teams won over in advance, an AI that thinks like you, or something that works without a network. All of them would help sell, and all of them would be false or dangerous. Here is what we say instead, one by one.
Monday, 9:30 am, in the office behind the kitchen
Josiane runs operations for four restaurants. On her desk, next to a cup of coffee, is the printed brochure of a tool she was shown last week. She is holding a red pen.
She crosses out one line, then a second. Idriss, her head chef, puts his head round the door: "What are you keeping?" She reads what is left: "Whatever they can show me during service. The rest is there to get me to sign."
(Fictional scene and characters.)
Why these promises get made
They answer a real fear
Buying a tool means risking time, money and the team's patience. A bold promise reassures at the moment of signing.
The team tests it from the first service
The brochure is read in the office. The tool is judged in the kitchen, on a Friday night, by people who signed nothing. The first promise that fails in front of them takes all the others down with it.
On some topics, a false promise hurts someone
A setup that runs late is annoying. A wrong answer about an allergen sends someone to hospital. Not all promises carry the same weight, and we treat them accordingly.
The nine promises we refuse to make
1. A tool that is never wrong
It is tempting because the buyer wants reassurance, and a demo in a quiet room seems to confirm it.
It is false. A tool that understands speech in a kitchen will sometimes mishear: the extractor fan, an accent, a three-word phrase thrown out between two tasks. And it is dangerous, because a team that has been promised infallibility stops checking.
What we say instead: San can get things wrong. That is why every action happens in two steps (San prepares and summarizes, the person confirms, and it is logged), why a fragile answer flags missing or conflicting data, and why the assistant has to be able to say it does not know.
2. An average success rate in big type
A figure fits on a slide and looks like proof of measurement, which is why it sells.
It is misleading. An average puts a harmless "say that again" in the same bucket as a mistake about a temperature. On a team that asks a lot of questions every day, even a high rate leaves misses every day, and the average does not say which ones.
What we say instead: no figure without a measurement protocol. We separate misunderstandings, which you fix by asking again, from mistakes on a critical topic, which no average should hide. Tsuno is used at Kajirō, our own restaurant group, but we draw no figures from it: that is not a protocol.
3. An assistant that would certify your allergens
It is tempting because this is the sharpest fear on the floor. A tool that took it off your hands would be worth its weight in gold.
It is dangerous. An assistant does not taste, does not see the recipe that changed this morning or the supplier who switched breadcrumbs. In France, allergen information for food that is not prepacked, such as dishes served in a restaurant, is governed by Decree No. 2015-447 of 17 April 2015. That information is something the restaurant owes its guests. No software can certify it on the restaurant's behalf.
What we say instead: on allergens, San is designed to show, not to summarize. It displays the business's approved, dated sheet, or says it does not have one and refers you to a manager. This is a design principle we are building and testing in kitchens, explained in is there sesame in it?.
4. An assistant that would take the place of the manager or of training
Managers are stretched thin, and a tool that trained people on its own would save them hours.
It is false. San does not demonstrate a technique the way an experienced colleague does, does not calm a panicking new hire, does not decide who is ready to run a station. If you pull the experienced cook off the station because "the tool knows", the new hire is left alone.
What we say instead: San gives access to what the team has written, at the moment of the task. Passing on skills stays human, and Tsuno Compétences is there to organize it on the job.
5. Turnover that would drop for sure
It is tempting because people leaving is the problem that wears restaurant owners down the most.
It is false. Someone leaves over pay, hours, their manager, their commute, a better offer. No tool guarantees they will stay, and there is no reliable data to put a figure on what a departure costs you.
What we say instead: we will track indicators the tool affects directly, such as time to autonomy on opening shifts, the time a manager spends on reporting, or the delay between an anomaly and the action taken. Each result will be published with its protocol.
6. A setup in a few minutes
Nobody has time, and a long rollout is frightening.
It is false for a useful tool. An empty tool is useless on day one, and the question "who writes the procedures?" comes up fast. Promising a few minutes means leaving that work to the owner without telling them.
What we say instead: your teams open Tsuno already filled. You can start with a trial built from three non-sensitive operational documents, in an isolated environment that is deleted on the agreed date. For a group, La Mise au Carré is an on-site engagement: we pull the content out of your documents, a person validates every sensitive match, then everything is reviewed on the floor.
7. Teams that will love it
It is tempting because the owner buys for the team and wants to be sure of the choice.
It is a promise about other people's feelings. On day one, everyone tries it out of curiosity. By day twenty, only the people it really helps are still using it. Promising buy-in pushes you to impose the tool, then to measure buy-in by the number of logins.
What we say instead: look at voluntary use after two weeks and after two months. A team told about the tool before day one, and free to say it does not help, will tell you the truth. And our principle is that voice stays an option, never an obligation.
8. An AI that thinks like you
Every owner dreams of a double who would decide the way they do.
It is false and risky. San does not think for you: it finds what your business has written, within the permissions of the person asking. Lending it your judgment ends with handing it decisions that belong to you.
What we say instead: Tsuno brings the facts together, people decide. When the AI suggests a connection, it is flagged as an inference, never presented as a fact. When nothing has been recorded, Tsuno says so instead of filling the gap.
9. Something that works everywhere, even without a network
It is tempting because cold rooms, basements and storerooms often have poor coverage.
It is not true for us today. The assistant relies on online services: without a network, it does not answer.
What we say instead: keeping critical procedures and checklists available during an outage is what a good tool should do, and it is what we are working on. We will announce it when we can show it.
The summary
| The tempting promise | The risk | What we say |
|---|---|---|
| A tool that is never wrong | The team stops checking | It can be wrong, hence the two-step confirmation |
| An average rate in big type | It hides critical mistakes | No figure without a protocol |
| Certified allergens | Someone in danger | The approved sheet, or a referral to a person |
| Replacing manager or training | The new hire left alone | San gives access to what the team wrote |
| Guaranteed lower turnover | A promise no one can check | Direct indicators, with a protocol |
| Setup in minutes | An empty tool | Tsuno delivered filled, trial or Mise au Carré |
| Teams won over | A tool imposed on people | Voluntary use at two weeks and two months |
| An AI that thinks like you | Decisions handed over | Facts brought together, people decide |
| Everything works offline | A failure mid-service | Work in progress, announced when it can be shown |
Common mistakes when reading a vendor's promise
- Believing a figure without asking how it was measured, where, on how many people and for how long.
- Judging a tool on a demo in a quiet room. Ask to see it during service, or on your own documents.
- Never asking what the tool refuses to do. A vendor with no answer to that question has probably not thought about it.
- Mixing up a demo with a delivered product. In our case, San Voice, the voice interaction, is available as a demo; San Audio and San 360 are being explored and tested.
What to measure
- Mistakes on critical topics, one by one, with their cause (often a missing or outdated sheet), rather than an average.
- Voluntary use after two weeks and after two months, station by station.
- Time to autonomy for a new hire on an opening shift, before and after, with a written method.
Where Tsuno and San come in
What we can put in writing fits in a few lines: San acts within the person's permissions, it prepares and they confirm, and it is logged; Tsuno shows its sources and says when data is missing; payroll, invoicing, permanent deletion and legal acts are out of the AI's reach. Everything else, we present as work in progress. To see it on your own use cases, the Tsuno San page presents the assistant and Security and trust brings our commitments together.
Further reading
The principles behind these refusals are set out in our red lines: the ethics of AI at work, and the full picture in the guide AI, voice and privacy at work. For the assistant's own limits, read what San never does and why an assistant must be able to say "I don't know". For how we run trials, see why we test in the kitchen before we sell. Refusing charming voices follows the same logic: why we do not do avatars or cute voices.
Key takeaways
A promise the team can disprove at the first service costs more than it brings in. We would rather say what the tool does, what it does not do, and what we are still building.
Frequently asked questions
Can an AI assistant like San make mistakes?
Yes. A tool that understands speech in a kitchen can mishear or give a wrong answer. That is why every action goes through preparation, a summary and a confirmation, and why a fragile answer has to say so.
What is Tsuno San's accuracy rate?
We do not publish an average rate. A single figure lumps a misheard phrase together with a mistake on a critical topic. Any figure we publish will come with its measurement protocol.
Can San guarantee allergen information?
No. Telling guests about allergens remains the restaurant's responsibility. San's design principle is to show the business's approved, dated sheet, or to say it does not have one and refer to a manager.
How long does it take to set up Tsuno?
We do not promise a time in minutes. Tsuno goes live already filled with your procedures. You can start with a trial built from three non-sensitive operational documents, then, for a group, with La Mise au Carré, an on-site engagement.
Does Tsuno San work without an internet connection?
The assistant needs the network to answer. Keeping critical procedures and checklists available during an outage is what a good tool should do, and it is something we are working on.
Does a workplace assistant reduce staff turnover?
No one can guarantee that. People leave over pay, hours, management and the job market. We would rather track indicators the tool affects directly, such as time to autonomy on opening shifts, and publish them with their protocol.