AI in restaurant HR: a doctrine for what the tool may do

A good doctrine for AI in human resources fits in one sentence: the software recalls, people decide. AI can retrieve dated facts, prepare a review conversation and flag what has not been dealt with. It does not rank employees, does not grade anyone, does not predict departures and does not recommend discipline or dismissal.

Three requests a manager might type on a Tuesday

A group manager plans the week with an AI. They type three things. First: "What has happened with the evening commis since they started?" Then: "Rank my servers from best to worst." Finally: "Who is likely to leave before summer?"

The first request is legitimate: dated facts that can be checked. The other two are of a different nature. They ask the tool for a judgment about people, from incomplete data, with no possibility of dialogue. A doctrine lets you answer without hesitating: yes to the first, no to the next two.

Why this is not only a question of ethics

• Restaurant data is partial. A late arrival is not disengagement, an unticked task is not misconduct. A judgment based on a fragment is wrong more often than people think.
• The legal framework is strict. Disciplining, dismissing or assessing an employee follows rules a tool does not know. Processing personal data, including profiling, falls under data protection rules such as GDPR. For any specific case, consult legal counsel.
• A prediction becomes a prophecy. Telling a manager that someone will leave changes how they see that person, and therefore how they behave toward them.
• A score removes accountability. "That's what the tool says" replaces the conversation, and the conversation is exactly what resolves situations.

What the tool can do, what it must refuse

Request	Response consistent with the doctrine
"Summarize what was decided with this person"	Yes, with the date and source of each item
"Prepare the facts for Thursday's review conversation"	Yes: completed trainings, incidents handled, agreed goals
"Who has not done their food safety training?"	Yes, it is a scheduling fact
"Who is the best server?"	Refusal: no ranking of people
"Who is going to quit?"	Refusal: no behavioral prediction
"Should I fire X?"	Refusal: a human decision, not recommended by the tool
"Are they motivated?"	Refusal: no inferring of mood

Making the doctrine concrete: five product rules

1. Write down the list of refusals. It should exist somewhere readable, not only in the founders' heads.
2. Show the source of every fact. Measured, declared, proven, observed, inferred or missing: the manager knows what they are holding.
3. Admit when data is missing. "I don't have that information" is a good answer. A guess dressed up as fact is a bad one.
4. Require human confirmation before any action. The tool prepares and summarizes the effect, the human confirms, the decision is tracked.
5. Bring problems back to the process. When a delay keeps recurring, the answer points to training, scheduling or the instruction before it points at a person.

The most common misreadings

• Believing "a human validates" is enough. If the tool produces a ranking and the human clicks "ok" out of habit, the decision actually belongs to the tool.
• Hiding the doctrine in the legal notices. Employees do not read them; they see what the tool does.
• Confusing fact and interpretation. "Arrived after the scheduled time on Mondays last month, dates attached" is a fact. "They are disengaging" is a judgment.
• Rejecting all AI on principle. The real issue is not the technology but what you ask it to do.

An example of a compliant review preparation

Take an end-of-probation review for a sous chef. A preparation that follows the doctrine lists: the training modules completed and those remaining, the incidents the person stepped into and how they ended, the goals agreed at the start and the actions decided since. Each line carries its date and source. It also says what is missing: "no observation of closing alone." What it does not contain counts just as much: no adjective about character, no comparison with a colleague, no recommended decision. The manager arrives with facts and leads the conversation.

Three credibility indicators

• The number of pieces of information shown without a source. It should be zero.
• The ranking or prediction requests the tool actually refused, and the quality of the explanation it gave.
• The share of sensitive decisions that carry a dated human sign-off in the history.

Where Tsuno comes in

Tsuno's doctrine is this: it connects the facts, it never judges a person. By design it refuses to rank employees, grade a person, predict a resignation, recommend a dismissal, infer motivation or mood, apply an automatic sanction or build a profile. It prepares the facts for a review conversation (trainings, incidents, actions taken), indicates their source, says when data is missing, then leaves the human to decide. Every action goes through a tracked confirmation. The detail is on the trust page.

To complete the picture

Read why AI should not judge your employees, how to prepare an employee conversation with facts rather than impressions and how to know whether a manager is ready to run a location. The common thread is in managers do not need another dashboard.

Key takeaway

In HR, AI may retrieve and summarize, not conclude about a person. The boundary should be visible in what the tool accepts and refuses, with the source of every fact. The software recalls, people decide.