How to Write Performance Reviews for Your Team Using AI (Fairly)

What AI Can and Cannot Do Here

AI is an evidence synthesiser and a language auditor. It reads a pile of notes faster than you can, groups them, drafts from them, and flags phrasing that judges a person rather than describing work. What it cannot do is know your team, weigh a quarter that does not appear in your notes, or understand where your company draws the line between “meets” and “exceeds.”

Most of the damage happens at the very first prompt. Ask for a finished review and you get fiction dressed as feedback; ask for organised evidence and you get exactly that.

The first prompt hands the model your judgment and asks it to reverse-engineer a justification. It will happily oblige, because “enthusiastic” and “great team player” are exactly the kind of trait language it has read millions of examples of. The second keeps the judgment where it belongs.

There is a legal edge to this too, and it is newer than most guidance acknowledges. The European Union’s AI Act lists systems used to monitor and evaluate the performance of workers among its high-risk categories in Annex III, with the high-risk obligations taking effect in August 2026. Any employment lawyer will make the same point regardless of jurisdiction: a decision affecting someone’s pay or job needs a human who can explain it. Drafting language from evidence you gathered is not that. Letting a model pick the rating is.

How to Write Performance Reviews Using AI

How to write performance reviews using AI comes down to four stages, always in this order: scrub the inputs, build the ledger, draft from the ledger, audit the draft. Skipping straight to drafting is the single most common mistake, and it is why so many managers conclude AI is useless for this — they fed it impressions and got impressions back.

Open a blank document and paste in everything you have about one person: 1-on-1 notes, project retros, peer shout-outs, their self-assessment, ticket summaries. Then run a find-and-replace pass. The name becomes [Employee]. Project names become [Project A]. Client names go. Revenue figures become “a large renewal” unless the number is genuinely the achievement.

This is the step people skip, and it is the one that makes the rest defensible. Every enterprise security officer says the same thing in different words: the risk is not bad grammar, it is employee data landing somewhere your company never approved.

Which plan you are on changes how careful you need to be, because consumer and workspace tiers handle your conversations differently. ChatGPT Business runs $20 per user per month billed annually, or $25 billed monthly with a two-seat minimum; Claude Team is $20 per seat annually or $25 monthly with a five-seat minimum. For anyone doing this across a whole team, that is the cheapest compliance insurance available.

If you want the fuller picture on what your employer can and cannot see, our guide to whether ChatGPT is safe to use at work covers the account-level settings worth checking before review season, not during it.

Now hand the scrubbed notes over with a job that has nothing to do with writing. You want a dated timeline, grouped into themes the evidence itself supports, plus an honest list of where your notes are thin. That last part is the bit no competing guide asks for, and it is the most useful output of the whole exercise.

You are an experienced HR operations lead helping a manager prepare a performance review.

Below are raw, anonymised notes about one person covering [review period, e.g. January to December 2026]. They are referred to only as [Employee]. Their role is [job title] on a [team type] team.

NOTES:
[paste your anonymised notes here]

Do this:
1. Sort every note into a dated timeline, grouped by quarter.
2. Group the timeline into no more than four themes based on what the notes actually contain. Name each theme yourself.
3. Under each theme, list the two strongest pieces of evidence, naming the outcome and the date.
4. List every claim that has no date, no outcome, or no measurable result.
5. Tell me which months have the thinnest evidence.

Rules: do not invent facts, dates, numbers or achievements. If a note is vague, say it is vague rather than filling the gap. Do not write any review language yet.

The gaps list is where recency bias dies. If four of your five themes draw entirely on October and November, you have not found evidence of a weak year; you have found evidence of a weak note-taking habit. Fix it going forward by treating 1-on-1 preparation as evidence capture rather than agenda-setting, and by turning meeting transcripts into structured notes as you go instead of at year end.

Draft the Review in Your Voice

The draft stage has one job: turn the ledger into prose that sounds like you, not like a quarterly earnings call. Default AI output fails here in a specific and predictable way, and every leadership coach has watched it happen — the writing gets so polished that the person reading it does not recognise themselves in it.

The fix is a banned-word list plus a hard rule that every positive statement names the outcome that earned it. If you want the underlying structure behind prompts like this one, it is the same role, goal, context, format pattern we use everywhere on the site.

You are helping me write the narrative section of a performance review. I have already decided the rating. Do not suggest one, argue with it, or mention it.

Employee: referred to only as [Employee], a [job title] on my team for [length of time].
Review period: [dates].
Audience: [Employee] will read this, and so will [HR / my skip-level].

EVIDENCE — the only material you may use:
[paste the themed ledger from the previous step]

Write three sections:
1. What went well — about 150 words, tied to specific dated outcomes.
2. Where the work needs to develop — about 120 words, about the work, never the person.
3. Focus for the next period — three specific, checkable commitments.

Voice rules:
- Plain English, short sentences, first person ("I saw", "I asked you to").
- Ban these words entirely: synergy, leverage, spearheaded, passionate, rockstar, journey, impactful, holistic, exceptional.
- Never describe [Employee] as likeable, abrasive, emotional, confident, humble, or with any other personality trait.
- Every positive statement must name the outcome that earned it.
- Add nothing that is not in the evidence above. If a section is thin, say so instead of padding it.

The underperformer draft

This is the one managers most want to hand over and most need to keep control of. The draft can carry the words. It cannot decide that the gap is real, and it must never be the first time the person hears about it.

Same evidence and voice rules as before.

[Employee] is not meeting expectations in [specific area]. This will not be a surprise to them — I raised it on [date] and again on [date].

Write the development section only, in about 150 words:
- Name the gap in terms of observable work, not attitude or effort.
- Reference the two conversations above, using those dates.
- State what meeting expectations would concretely look like next period.
- Offer two forms of support I can realistically provide.

Do not soften the gap into a compliment. Do not use "opportunity" as a euphemism for a problem. Do not reference or recommend any formal HR process.

If the gap has reached the point where a formal process is the honest next step, that is a different document with different rules — see drafting a performance improvement plan rather than stretching a review to do that work.

The promotion case

Same evidence and voice rules as before.

I am recommending [Employee] for promotion from [current level] to [next level]. Our published criteria for [next level] are:
[paste your company's level criteria]

Write a promotion justification of about 250 words that:
1. Maps specific dated evidence to each criterion, one short paragraph per criterion.
2. Flags any criterion where the evidence is thin, and says so plainly.
3. Describes scope the person is already operating at, not potential.

No superlatives. Do not claim readiness for a criterion the evidence does not support — a weak paragraph I can go and fix is far more useful than a confident one I cannot defend in a calibration room.

Calibrating across your team

Calibration is where individual fairness becomes team fairness, and it is a language problem before it is a ratings problem. Run all your finished drafts together and look for uneven evidence depth — two people at the same rating where one got three dated outcomes and the other got adjectives. Peer input helps here, and synthesising 360 feedback before you draft gives the ledger a second source.

I am preparing for a calibration meeting with other managers.

Below are the finished narrative sections for [number] people on my team, anonymised as [Employee A], [Employee B] and so on, each labelled with the rating I have assigned.

[paste the drafts]

Analyse the language only, never the ratings:
1. For each person, list the concrete outcomes cited and count the vague claims.
2. Show me where two people at the same rating received noticeably different depth of evidence.
3. Flag any sentence I would struggle to defend if a peer manager challenged it.
4. Name the single strongest piece of evidence for each person so I can lead with it.

Do not suggest changing any rating.

Run the Bias Audit Before Sending

Every draft gets audited, including the ones that read beautifully — especially those. Bias in reviews rarely looks like hostility. It looks like warmth pointed at the wrong thing: praise for who someone is rather than what they delivered.

Audit the performance review draft below for evaluation bias. Do not rewrite the whole thing yet.

DRAFT:
[paste your draft]

Report on each of these separately:
1. PERSONALITY LANGUAGE — every phrase describing who [Employee] is rather than what they did. Quote each one.
2. RECENCY — which months the cited evidence comes from, and whether any quarter is missing entirely.
3. VAGUENESS — every positive claim with no outcome attached.
4. COUNTERFACTUAL — take the three strongest sentences and rewrite each twice, once using he/him and once using she/her. Tell me whether the tone shifts, and exactly where.
5. HEDGING — criticism stated so indirectly the reader could miss it.

For every finding give me the quote, the problem in one line, and a replacement grounded only in the evidence I provided.

Here is what most guides get wrong about this step. They present AI as a bias remover. It is not one, and treating it as one is worse than not auditing at all. The writing-analytics firm Textio tested exactly this and found that asking ChatGPT to remove bias from written feedback mostly left the personality comments in place — they simply came back sounding more formal, and which ones survived varied with the employee’s name. Use the audit as a flagger. You decide what each flag means.

What You Must Never Outsource

People analytics specialists will tell you AI beats human memory at recalling a year of work, and they are right. That is a claim about recall, not about judgment, and the two get conflated constantly.

There is also the question everyone asks quietly: will HR or my report be able to tell? Not from the text, if you followed the method — but that is the wrong thing to worry about. The ledger is your answer. You did the observing, you kept the dates, you chose the rating. A model organised your notes. That is a defensible sentence in any room, and it is worth being ready to say it plainly rather than hoping nobody asks.

When the Draft Comes Back Wrong

Most bad output traces to a missing instruction rather than a bad model. Match the symptom to its cause before you re-prompt, and you will fix it in one pass instead of five.

Invented achievements deserve special attention, because they are quiet. A model filling a thin theme with something plausible is the same failure mode described in our guide to why AI hallucinates and how to stop trusting the wrong outputs — and in a review it can end up in a permanent record. Before anything is submitted, check that:

• Every date, number and outcome in the draft appears in your ledger
• No sentence describes the person rather than the work
• Each quarter of the review period is represented somewhere

For tone that lands wrong rather than facts that are wrong, the repair is usually a single targeted instruction rather than a fresh draft; our walkthrough on refining AI output without starting over covers the diagnostic pattern. And once the document is finished, the harder half of the job starts — preparing for the conversation itself is where a good review either lands or does not.

Work through one person end to end before you touch the rest of your team. The scrub takes ten minutes the first time and three after that, and by the second review you will be spending your time on the part that was always the real work: deciding what you actually think. Knowing how to write performance reviews using AI is worth having precisely because it hands that part back to you.