Performance review with AI: prepare it without exposing your staff
Nine annual reviews in two weeks, one of them dreaded. A team leader prepares them with AI, from the assessment to the written record, without anything that would identify her staff.
Thursday, 6:15 pm. The last call of the day has just ended on the contact centre floor, and Sophie opens the calendar for this year's round of performance reviews: nine meetings in two weeks, with the nine customer service advisers in her team. Eight will go smoothly. The ninth worries her. This adviser, with the company for eight years, is going to ask her again for the team leader role. She will have to say no, and also talk to him about the quality of his calls, which has been slipping since the spring.
On her desk lies a notebook filled over the course of the year. The temptation is familiar: paste everything into a consumer AI tool and ask "help me prepare this review". Along with the employee's name, his results, and a note scribbled in April about three weeks of sick leave.
Here is how Sophie, a fictitious team leader, prepares her review round differently, with the requests she actually makes.
Performance reviews: what a manager is entitled to write
A performance review produces a judgement on someone's work that will weigh on their salary and their career: some of the most sensitive personal data in working life.
In many countries, employment law requires appraisal methods to be relevant to their purpose and employees to be told about them in advance; some also require a regular career development interview alongside any annual review. In Switzerland, the Code of Obligations limits the data an employer may process about an employee to what concerns their suitability for the job or the performance of the contract. Everywhere, data protection law applies on top: process only what is necessary, and remember that employees can ask to see what has been written about them, a right that both the GDPR and the Swiss Federal Act on Data Protection (FADP) provide.
One rule runs through all of this: health, family life, opinions and trade union membership have no place in an appraisal. Sick leave is not a lack of commitment. If that information is in your notes, it should go no further: not into the written record, and not into an AI.
Finally, your company's own rules come first: if AI is not allowed there, raise the subject rather than working around it.
One week to prepare your performance reviews
Monday, 8:30 am: a space for the review round
In IA Confidential, Sophie creates a space she calls "Annual reviews". In the space's Assistant tab, she writes instructions that will apply to every conversation she opens there, ahead of her general preferences:
- stick to dated, observable facts, never character traits;
- assume nothing about an employee's private life, health or motives;
- refer to each person by a letter, from A to I;
- a direct, respectful tone, short sentences, no HR jargon.
In the Privacy panel, she adds her anonymisation instructions: always mask the first names of the team, the name of the client the contact centre works for, and the name of her company. These are the words that, even on their own, would allow someone to be recognised.
Finally, she checks the models: by default, her requests are handled by open models installed on servers in Switzerland, which pass nothing on to third parties. For this work, she needs nothing else.
Tuesday evening: from notebook to factual assessment
For each of the nine advisers, Sophie rereads her notes and does a first sort by hand. First names become letters. The adviser who worries her becomes "D". The note about sick leave is crossed out: it has no place in an assessment, and no place in an AI. The quality figures stay, without the name of the internal tool that produces them.
She then opens one conversation per employee and pastes in her cleaned-up notes. This is the request she reuses nine times, and the first example to copy:
Here are my notes from the year on an adviser, referred to by the letter D, and the three objectives set for him last year. For each objective, say what the notes show, with the date or fact that supports it, and write "nothing in the notes" when there is nothing. Separate facts from my impressions, which you should flag as such. Do not invent any fact, do not give any overall rating and do not draw any conclusion about the person.
D's assessment comes back as a table. It reminds Sophie of something she had forgotten: in February, D trained two new starters on his own. It also points out that "seems less engaged" is an impression, not a fact. She removes it. She checks every line against her notebook: for another adviser, the assistant had placed in March an incident that happened in May.
Wednesday: preparing the difficult conversation
The assessment does not say how to turn down someone who has been waiting two years for this promotion. Still in the space, Sophie opens a new conversation. Here is the second example to copy:
I am preparing an annual review with an experienced adviser who is hoping for a team leader role. I am going to tell him it will not happen this year, for two factual reasons: the quality of his calls has dropped since the spring, and he has not yet run a team meeting. He also has real strengths, especially in training new starters. Suggest a thirty-minute outline: how to open, when to announce the decision, how to present the two reasons without judging the person, and how to finish on a concrete plan for the year. Add the three most likely reactions and, for each, a possible response. Do not suggest any promise I might not be able to keep.
What she takes away above all is a piece of advice she would not have followed by instinct: announce the decision in the first five minutes, rather than after a list of compliments that D would hear as a preamble. One suggested response promises "a review in six months": Sophie strikes it out, as that is not hers to decide. She rewrites the opening in her own words and jots down on a card the two dated facts she will cite.
Thursday lunchtime: rehearsing the appraisal
Sophie knows she tends to over-justify herself when contradicted. So she asks the assistant to play the employee.
Play the adviser, disappointed and a little defensive. I will start the review, and you react as he might, without turning him into a caricature. After ten exchanges, stop and tell me where I was vague, where I over-justified myself, and where I let the decision seem negotiable.
The feedback is blunt: twice, she added another argument instead of rephrasing the objection. She starts again, more briefly.
Out of curiosity, she then picks an external model for a second reading. Before anything is sent, the Confidentiality filter detects a piece of sensitive data left in the text and offers three routes: answer with a confidential model, message untouched; send an anonymised version, which she can reread and correct; or send it as it is. The default choice remains the confidential AI. She keeps it.
Friday, 4 pm: the written record after the review
The review took place at 2 pm. D was disappointed, said so, then offered of his own accord to run the weekly team meeting in January. Sophie took notes by hand, with his own words in quotation marks.
In D's conversation, she pastes these notes, still without a name, and asks for a written record following her company's template: review of objectives, decision and reasons, objectives for the year, the employee's comments reproduced word for word. The assistant writes it as a document, which she edits and then exports to Word or OpenDocument. The layout is rebuilt from the text: Sophie transfers the content into the official form, where she adds the name and her signature. D's name never went through the AI.
She rereads it with one question in mind: if D ever asks to see what was written about him, does every sentence stand up? A sentence about his "lack of ambition" disappears.
The following Monday: leave nothing lying around
The records are signed. Sophie deletes the conversations from the review round: they only ever existed in her browser, and there is no longer any reason for them to stay there. She keeps the space and its instructions for next year, and for the career development interviews in the spring.
What AI does not do in a performance review
AI helps you put things in order, find the right wording, and practise. It does not appraise anyone, and it must not.
It can get things wrong and make things up with confidence: shift a date, attribute to an employee something he never said, turn an impression into a fact. Every line of an assessment is checked against your notes, because you are the one who signs the record and answers for it.
It can also reproduce biases, judging one communication style more harshly than another. The decision on a salary increase, a promotion or a formal warning remains with a person who can explain it. And some situations gain nothing from going through an assistant, even a confidential one: a suspicion of harassment, a conflict involving someone's health. There, you talk to HR, occupational health or a lawyer.
What stays confidential
With a confidential model, the notes you paste and the answers are processed on servers located in Switzerland, under the Swiss Federal Act on Data Protection, outside the reach of the CLOUD Act, and nothing is kept on our side. Your conversation history stays in your browser: we could not produce it if asked to, because we do not have it. A deleted conversation no longer exists anywhere.
With an external model, nothing leaves without going through the Confidentiality filter, and you decide: confidential model, an anonymised version you reread, or the message as it is, knowing the risk. For named appraisals, stay with the confidential models.
Two precautions specific to managers. Whatever you share must not allow the employee to be recognised: a letter rather than a first name, nothing about health or private life, even with a confidential tool. And since the history lives in the browser, anyone using the same session on the same computer can read it: on a shared computer on the floor or in the managers' office, work in a personal session and sign out when you leave.
Further reading
Recruitment raises the same questions earlier on: see our guide on summarising CVs with AI. To know what to remove from your notes before handing them to a tool, keep the list of data you should never paste into an AI to hand, and read the difference between anonymisation and pseudonymisation: replacing a first name with a letter is not always enough.
To rehearse your next difficult conversation on an invented case: open the chat.
