
Once in twenty, for a team of fifteen asking ten questions a day each, means seven or eight misses a day and close to two hundred a month. That number tells you nothing until you sort it. A question that has to be repeated, or an action turned down at confirmation, is acceptable. Outdated information, or a critical question sent the wrong way, must be driven toward zero, and every case is handled on its own.
Friday, 8:05 p.m., at the pass
The rush is in full swing. Tristan, the sous-chef, is calling tickets across the pass. Moïse, a month into the cold station, answers "yes chef" to every line.
"Two tartares, one no capers, one burrata." Moïse: "Two tartares, one burrata." Tristan looks up. "Again." Moïse repeats it and drops the capers again. Tristan says it once more, slower, and waits to hear "one no capers" before moving to the next ticket. The trays wait ten seconds longer. Nobody complains.
In this brigade, people get things wrong all night. They mishear, skip a detail, mix up two tables. It is not a disaster, because the pass is built to catch exactly that: call, call back, check. What Tristan would never let through is a plate going out with capers to the guest who asked for none.
A kitchen assistant should be judged the same way. (Tristan, Moïse and their brigade are fictional.)
What "once in twenty" means for a real team
Take a round number as a working assumption: an assistant that gets it wrong once in twenty. This is not a figure measured on San. It is an order of magnitude to reason with.
| Scale | Questions asked | Misses at once in twenty |
|---|---|---|
| One person, one day | 10 | 0 or 1 |
| Fifteen people, one day | 150 | 7 to 8 |
| Fifteen people, a six-service week | 900 | 45 |
| Fifteen people, a 26-day month | 3,900 | 195 |
"Once in twenty" sounds reassuring. Over a month, it is almost two hundred moments when the tool did not do what was expected. Nobody can say whether that is acceptable without knowing what those 195 misses are. We set out this calculation in why an assistant must be able to say "I don't know". This article takes the next step: the grid for sorting them.
Six families of misses, from harmless to serious
Asking again
The extractor hood drowns out the voice, the tool did not catch it and says so: "Sorry, I didn't get that, can you repeat?" This is the most frequent miss and the cheapest. It is visible, it costs a few seconds, and it leaves no false record behind.
The incomplete answer
Someone asks for the bar closing steps, the tool gives the first four and fails to mention there are seven. The person spots the gap if the sheet is in front of them, not if they are only listening. Annoying, and worth watching.
The wrong sheet, on a non-critical topic
"Brunch prep" opens the summer brunch sheet instead of the winter one. The person usually notices at the first step. The cost stays low, as long as the topic has nothing to do with health or safety.
The action stopped at confirmation
The tool heard "create a task for the hood" instead of "tick off the hood task". It prepares, summarizes, and the person says no. It is a miss, but it is also proof that the safeguard works: nothing was carried out. The principle is explained in the two-step rule.
Outdated information
The tool answers correctly, but from a sheet that no longer matches reality: the old oven, the old supplier, the old recipe. This miss is dangerous because it is invisible. The answer looks right and is followed with confidence. We give it its own article: an outdated procedure is more dangerous than a missing one.
A critical question sent the wrong way
A question on an allergen, an out-of-range temperature, a cleaning chemical or first aid gets a summarized answer instead of the approved sheet, or any answer at all where the tool should have said "I don't have a sheet for that, ask the person in charge". This is the most serious family, even if it is rare.
The sorting grid
To separate the families, frequency matters little. Two questions decide: is the miss visible when it happens, and does it touch health, safety or the law?
| Family | Visible at the time? | Critical topic? | Status | What to do |
|---|---|---|---|---|
| Asking again | Yes | No | Acceptable | Nothing, unless it becomes very frequent at one station |
| Incomplete answer | Sometimes | No | Reduce | Check the sheet, show the list instead of reading it out |
| Wrong non-critical sheet | Often | No | Reduce | Rename or merge sheets that look alike |
| Action stopped at confirmation | Yes | No | Acceptable | Note the phrasings that get misheard |
| Outdated information | No | Sometimes | Drive toward zero | Correct or withdraw the sheet, tell the team |
| Critical question sent the wrong way | No | Yes | Drive toward zero | Review the same day, fix the rule before the next service |
Read the grid from the bottom up. The last two rows have one thing in common: nobody notices them when they happen. That is what makes them serious. A visible mistake costs time. An invisible one costs trust, and sometimes much more.
Why an average says nothing about safety
Picture two tools that are both wrong once in twenty, so 195 misses a month each.
- The first has 180 repeats, 14 actions stopped at confirmation and one wrong brunch sheet.
- The second has 150 repeats, 40 incomplete answers, 4 pieces of outdated information and one summarized answer on an allergen.
Same rate, same average. The first is a colleague who is a little hard of hearing. The second is a liability. No overall figure can tell them apart, which is why we do not publish an average success rate for San. What matters is the breakdown: how many invisible misses, how many on critical topics, and what was done about each one.
On critical topics, the goal is that no error gets through. You do not get there by hoping the tool is brilliant. You get there by building it not to answer from memory: it shows the approved, dated sheet, or it says it has none.
Keeping your own log of misses
A restaurant does not need to wait for a vendor's numbers. It can keep its own, on paper or in a shared log.
- Name one person per location, often the sous-chef or the opening manager. They keep the log for four weeks.
- Record each miss in one line: date, station, question asked, family from the grid, consequence, follow-up.
- Ask the team to report misses without worrying about it. "It made me repeat three times" is information, not a complaint.
- Review the log for fifteen minutes a week. Look for patterns: same station, same phrasing, same sheet.
- Handle any outdated information and any critical question sent the wrong way separately, the same day. Pass the case to the vendor with the exact question.
- After a month, compare the split by family, not the total.
The log is about what the tool answered, never about who asked.
Common mistakes
- Asking a vendor "what is your success rate?" and stopping there. Ask instead what the tool does with an allergen question when the sheet does not exist.
- Counting repeats as serious failures. A brigade calls back tickets all night; a tool that asks again instead of guessing is doing its job.
- Forgetting the invisible misses. If nobody looks for them, the log only contains repeats, and it reassures for the wrong reasons.
- Blaming the tool for a wrong sheet. Outdated information often comes from the content, and that is what needs fixing.
- Testing on a single day. Misses change with noise, station and time of day; a Saturday night is nothing like a Tuesday morning.
What to measure
- The number of invisible misses per month, outdated information and critical questions sent the wrong way. The goal is to drive it toward zero, and every case is reviewed.
- The share of acceptable misses (repeats, actions stopped at confirmation) in the total. The higher it is, the more mistakes catch themselves.
- The time between a critical miss being logged and the sheet or rule being fixed. It should be counted in days, not weeks.
Where Tsuno San fits in
Tsuno San, the intelligent voice assistant for teams, is designed around the grid in this article (San Voice is available in demo, San Audio under exploration and testing: see where San stands). Every action happens in two steps: San prepares and summarizes, the person confirms, and it is recorded, which turns a misunderstanding into a stopped action rather than a mistake. On critical topics, San is designed to show the location's sheet instead of summarizing it, or to say it has none. San acts within the person's permissions. It does not promise to be infallible, and we do not publish an overall success rate. A built-in log of misses, sorted by family, is part of what a good tool should offer; today, Tsuno's logbook already lets you record them with a date. To see how San handles a trick question, visit the Tsuno San page.
Further reading
For the big picture, read Tsuno San, the work companion for frontline teams. The most critical topic of all is covered in allergens: why no tool should answer from memory. To see how we test before we announce anything, read why we test in the kitchen before we sell. And on the gap between a reassuring screen and reality, compliance theater.
Key takeaways
An error rate is judged by what it is made of, not by its average: misses that are visible and caught are the normal price of talking in a kitchen. Those that go unnoticed on a critical topic must be driven toward zero, and a restaurant can check this for itself with a simple log.
Frequently asked questions
How many mistakes does an assistant make if it is wrong once in twenty?
For a team of fifteen people asking ten questions each per day, that is 150 questions and seven or eight misses a day. Over a month of 26 service days, roughly 3,900 questions and 195 misses.
Which assistant mistakes are acceptable in a kitchen?
The ones that are visible and fixed on the spot: a question that has to be repeated, an incomplete answer the person completes, a misunderstood action they turn down at confirmation. Mistakes that go unnoticed on a critical topic are not acceptable.
Why does an average success rate say nothing about safety?
Because two tools with the same rate can make opposite kinds of mistakes. One only asks people to repeat themselves, the other lets a wrong allergen answer through once a month. Same average, very different risk.
How can a restaurant track its assistant's mistakes?
With a simple log: date, question asked, type of miss, consequence, follow-up. Fifteen minutes a week is enough to review it, and every critical miss is handled separately, the same day.
What should you do when the assistant answers from an outdated procedure?
Log it as a miss in its own right, find the sheet behind it, correct or withdraw it, and tell the team. The fault often lies in the sheet rather than the tool, but it needs fixing just as fast.