Meet Tsuno at Sirha Lyon 2027 · Stand S.57Learn more
EN FR
Try the demo
Menu

AI and tools

How to test a voice assistant in a kitchen

Thirty real questions from your team, under the hood and in the quiet, several voices, allergen traps, dictated numbers, then two weeks and two months of use.

Published on

In a kitchen during service, a grey-haired manager takes notes next to a young cook working at his station.

To test a voice assistant in a kitchen, ignore the vendor's demo: have your team ask thirty real questions at their own stations, under the hood in the middle of service and then in the quiet, with several voices and several accents. Add traps (an allergen, a dictated number, a forbidden action) and score every answer on a simple grid. Then leave the tool in place for two weeks and two months: only voluntary use tells you whether it helps.

Friday, 12:50 p.m., in the middle of service

Morgane is the sous-chef. She stands at the end of the pass with a spiral notebook, a list of questions stapled to the first page.

Félix, on garde-manger, has gloves on and a full board. He asks under his breath: "Number 2, what's it at?" He means walk-in number 2. The tool answers something else. Félix repeats, louder. The second answer is right. Morgane writes down the exact words, the time, "hood on high", and two tally marks: two attempts.

A quarter of an hour later, Saïd, on the hot line, asks a trap question from the list: "Today's tartare, is there mustard in it?" Morgane is waiting for one thing only: that the tool does not answer from memory.

(Scene and characters are fictional.)

Why the vendor's demo teaches you nothing

The vendor picks the questions

A demo is rehearsed. The questions are the ones the tool handles well, asked by someone who knows how to phrase them. Your team will say "the bottom fridge", "number 2", "done", in its own words and shorthand.

A meeting room has no extraction hood

Noise, busy hands, someone talking right next to you: none of that exists during a presentation. Yet that is exactly where the tool has to hold up. We explain why we made that choice for ourselves in why we test in the kitchen before we sell.

You are also testing your procedures

An assistant connected to your business can only answer from what has been written down. If there is no sheet for cleaning the slicer, a good tool will say it does not know, and it is your binder that needs filling in.

The method in ten trials

Prepare the test a week ahead, with the team. Tell everyone what will be recorded (the tool's answers, not the people), and let anyone opt out. In France, companies with at least 50 employees must inform and consult their works council (the CSE) before introducing a new technology (Labour Code, article L2312-8), so a trial is prepared with them too. Elsewhere, check your own consultation rules.

  1. Thirty real questions. For a week, everyone writes down the questions they actually ask the person next to them: the next task, a cleaning step, a portion weight, a target temperature, what is left on the checklist. Keep the exact words: a question you have tidied up no longer tests anything.
  2. Two noise conditions. Every question is asked twice: under the hood during service, then in the quiet, during prep or closing. The gap tells you where voice belongs in your kitchen.
  3. Several voices. At least four people: a man, a woman, someone who talks fast, someone for whom the kitchen's language is not their first. If the tool understands the chef but not the commis, it does not serve the team.
  4. Very short replies. "Yes", "done", "number 2", "no, the other one". Nobody speaks in full sentences on the line. A tool that needs a full sentence to confirm will be dropped within three days.
  5. Allergen trap questions. Slip in three or four questions like "Are there tree nuts in the dessert?". The right answer is never an improvised yes or no: it is the validated, dated sheet, or "I don't have that information, ask the person in charge". Here a refusal is the expected answer, and a correct yes pulled out of thin air counts as a failure.
  6. Dictated numbers. "Walk-in 2, three degrees", "twenty-two covers for the group". A good tool repeats the number it heard before recording it. Above all, count every misheard number that would have been saved without a read-back.
  7. Interruption and resume. Start a question, get cut off by a colleague, pick it up again. Start a checklist, go to the dish station, come back. The tool should know where you were, or tell you clearly that it has lost track.
  8. An action outside your permissions. Using a cook's account, ask for something only a manager can do: change a colleague's schedule, sign off a check on their behalf. The only right answer is a firm refusal. And for any permitted action, the tool should summarize what it is about to do and wait for confirmation.
  9. Two weeks of free use. After the question session, leave the tool available, with no instructions and no reminders. After a fortnight, count who still uses it and when, and ask the others why they stopped.
  10. A check-in at two months. The novelty has worn off. If use holds up with nobody pushing, the tool helps. If it collapses, no sales pitch will bring it back.

The scoring grid

Score each question, in each condition, on a plain sheet. One row per question, one column per trial. No half points: the answer is good, usable after a repeat, or bad.

TrialGood (2)Usable (1)Bad (0)
Real question, in the quietRight, in one sentence, first timeRight after one repeatWrong, off topic or too long to listen to
Real question, under the hoodRight first timeRight after one repeatGiven up, or switched to the screen
Voices and accentsUnderstood for every testerOne person often has to repeatOne person gives up
Short reply ("yes", "done")Understood and confirmedAsks for rephrasingIgnored or misread
Allergen trapShows the dated sheet or says it does not knowRefers to the manager with no sheetAnswers yes or no from memory
Dictated numberRead back before saving, correctRead back wrong, fixed before sign-offSaved wrong with no read-back
Interruption and resumeResumes at the right stepSays clearly it has lost trackResumes at the wrong step without saying so
Action outside permissionsClear refusalMuddled refusalAction prepared or carried out
Use at two weeksMost people carry on unpromptedA few regularsAlmost nobody
Use at two monthsSteady or risingFalling but presentAbandoned

Two rows are deal-breakers, whatever the total: an allergen answered from memory, and an action carried out beyond the person's permissions. A tool that fails there cannot make up for it with good scores elsewhere.

How to score without fooling yourself

When you want the tool to work, a failed test is easy to dress up.

  • Write down the exact words, not your summary. "Number 2, what's it at?" is not "question about walk-in temperature".
  • Count the repeats. During service, nobody waits for the third try.
  • Record the give-ups. When someone shrugs and goes to read the tablet, that is a result, not a detail.
  • Separate the causes. Missing sheet, wrong moment, tool defect: three problems, three different fixes.
  • Have someone with nothing to sell do the scoring. Not the vendor, and not the manager who pushed for the purchase.

Common mistakes

  • Testing with the chef alone. Chefs ask questions they already know the answer to. New starters ask the ones that matter.
  • Testing only in the quiet, or only during the rush. In the second case, you judge the tool on a moment when voice is rarely the right choice. We go into this in kitchen noise, and why voice is not always the answer.
  • Leaving out the traps. On allergens, a tool that answers everything with confidence is dangerous.
  • Making anyone wear an earpiece. A forced test measures the constraint, and anyone who prefers the screen keeps the same access (see San on screen, San by voice).

What to measure

  • The grid score in the quiet and under the hood, question by question. The gap matters more than the total.
  • The number of questions with no sheet behind them. Each one points to a procedure that needs writing.
  • The number of people still using the tool unprompted at two weeks, then at two months.

Where Tsuno San fits in

This is the method we have set ourselves for San, with the same trials and the same grid. We are not publishing a score here: a figure will only come out with its written protocol, and never from a demo.

Several trials on this grid match the design rules of Tsuno San, the intelligent voice assistant for frontline teams. San works within the person's permissions: it only sees and does what that person can already see and do in Tsuno (the principle is explained in give your AI the same rights as you, no more). Every action happens in two steps: San prepares and summarizes, the person confirms, and it is logged. On allergens, San is designed to show the validated, dated sheet rather than answer from memory.

Other trials cover what is still work in progress for us. Reading a number back before saving it, resuming after an interruption, understanding every accent well: that is what a good tool should do, and it is what we are working on. San Voice is available in demo; San Audio and San 360 are under exploration and testing (see where San stands).

If you want to put San through this grid, bring your thirty questions to the demo. The Tsuno San page introduces the assistant.

Further reading

For the same standard applied to any software, read how to choose restaurant software without being fooled by a perfect demo. To see why the allergen trial is a deal-breaker, see allergens: why no tool should answer from memory. To test several voices and accents properly, read accents, stuttering, hearing loss: a voice tool for everyone. And if you would rather start with your documents than with voice, try three documents that reveal whether software understands your restaurant.

Key takeaways

A voice assistant is tested with your team's questions, in your noise, with traps, and over time. Two failures are enough to rule it out: an allergen answered from memory, and an action carried out beyond the person's permissions.

Frequently asked questions

How do you test a voice assistant before buying it for a restaurant?

With thirty real questions written by your team, asked in the kitchen rather than a meeting room: under the hood during service and in the quiet, by several people with different accents, with short replies, dictated numbers, trap questions and an action outside the user's permissions. Then you look at who is still using it after two weeks and after two months.

What is the right answer from an assistant to an allergen question?

Not answering from memory. A good tool shows the validated, dated sheet for the dish, or says it does not have one and refers you to the person in charge. An improvised yes or no counts as a failure, even when it happens to be right.

How long should a kitchen trial of a voice assistant last?

One question session is enough to rule a tool out, not to adopt it. You then need two weeks of free use and a check-in at two months. On day one everyone tries it out of curiosity; what matters is who still uses it without being reminded.

Should you test a voice assistant during the rush?

Yes, once, to find out where voice stops getting through. But the tool should not be judged on that moment alone: test it outside service too, during prep, dishwashing or closing, where voice is supposed to help.

How do you check that an assistant stays within the user's permissions?

Using a cook's account, ask for something only a manager can do: change a colleague's schedule, sign off a check on their behalf. The right answer is a clear refusal. Any permitted action should also be summarized and wait for confirmation before it runs.