When does AI help with a performance review, and when is it risky?
AI helps with a performance review when the manager already has specific evidence and needs help organising it and wording it clearly. It is risky when it is asked to produce feedback from a job title, to decide a rating, or when personal data goes into a tool that is not approved for it.
A fictional 60-person insurance broker, Ashcombe Partners, runs half-yearly reviews. A team leader, Priya, is reviewing an account handler, referred to here as Employee A. Her notes are scattered: 'renewals team, handled 230 renewals H1. client from Marlow firm emailed thanks. missed two deadlines in March (system migration?). quiet in team mtgs but helped onboard new starter. wants to move to commercial lines. needs to improve: chasing docs earlier.' Priya has a form with four headings and a rating scale, and a dozen reviews to write.
How do you write a fair AI performance review step by step?
Gather evidence first, remove identifying details, ask the AI only to organise and word what you supplied, check the language for balance and bias, and make the rating decision yourself. An AI performance review is a drafting aid, not an assessment.
1. Check the rules
Read your employer's AI policy and HR guidance. Use only the approved tool, and include no personal data beyond what that tool is approved for. Many teams use a placeholder such as Employee A and no health, absence or family details.
2. Collect evidence, not adjectives
Write down observed examples, outcomes and dates. 'Handled 230 renewals' is evidence; 'hard-working' is not.
3. Ask the AI to sort, not to judge
Map each piece of evidence to the review headings and objectives. Tell it not to add any example, praise or criticism you did not supply.
4. Balance the language
Ask it to flag vague, personality-based or loaded words and suggest behaviour-based wording.
5. Decide and own it
Choose the rating yourself, discuss it with HR if needed, and rewrite the final text in your own words.
What does a weak and an improved AI performance review look like?
Weak prompt: 'Write a performance review for an account handler who is good but sometimes misses deadlines and is quiet.' Typical result: 'Employee A is a dedicated and reliable team member who consistently exceeds expectations. However, they could be more assertive and proactive in team settings.' It contradicts itself, invents 'consistently exceeds expectations', and 'more assertive' criticises personality rather than a behaviour linked to the role.
Improved prompt: 'Using only the evidence below, map each item to these headings: results, quality, teamwork, development. Do not add examples, praise or criticism. Flag any wording about personality rather than observable behaviour. Evidence: [Priya's notes with the client's name removed].' Result: Results: handled 230 renewals in H1; a client sent written thanks. Quality: two renewal deadlines missed in March, during the system migration, to be discussed for context. Teamwork: helped onboard a new starter. Development: interest in commercial lines; agreed focus on requesting documents earlier in the renewal cycle. Flagged: 'quiet in team meetings' describes personality; consider whether there is a role-relevant behaviour to discuss, or leave it out. Priya decides it is not relevant and removes it.
How do you check an AI-assisted review before it is shared?
Check that every sentence traces to evidence you supplied, that the language describes behaviour rather than personality, and that the tone matches the rating you chose. Then read it as the employee would: would they recognise each example and know what to do next?
Traceability
Delete any praise, criticism or example you cannot point to in your notes.
Bias and balance
Look for words that are often applied unevenly, such as abrasive, emotional, aggressive or lacks presence, and for feedback that judges personality rather than work.
Context
Missed targets during a known disruption, like a migration, should be stated with that context.
Privacy
Confirm that no health, absence or other sensitive personal details went into the tool or into the text.
What are the common mistakes with AI performance reviews?
The most serious mistake is asking the AI to generate feedback the manager never observed. Others: pasting names, health or absence details into an unapproved tool; letting the AI pick the rating; using the same AI phrasing for every team member so reviews read alike; and sharing the draft without reading it as the employee would.
Where does this lesson fit in a Tosheo.ai path?
Performance reviews sit in the Tosheo.ai paths for HR teams and for managers, alongside lessons on AI policy and responsible use of personal data. The hands-on task gives the learner fictional evidence for a review, and the rubric feedback checks for invented content, personality-based language and missing context.