How to Use AI to Write a Performance Review — Yours or Someone Else's

Performance review season arrives the same way it always does — with a deadline that seemed far away until it wasn't, and a blank text field asking you to summarise months of work in a few paragraphs that sound professional, specific, and neither too modest nor too self-promotional. If you're a manager, multiply that by however many people report to you.

This is exactly the kind of task AI tools are genuinely good at helping with. Not because AI knows anything about you or your team — it doesn't — but because the hard part of a performance review usually isn't knowing what to say. It's finding the words to say it clearly, in the right register, without spending two hours on something that should take twenty minutes.

Here's a practical approach to using ChatGPT or Claude for performance reviews, whether you're writing your own self-assessment or writing reviews for others.

Why Performance Reviews Are Hard to Write (And Why AI Helps)

The specific difficulty with performance reviews is that they ask you to be simultaneously accurate, diplomatic, specific, and appropriately formatted — all while your brain is also tracking whether you're coming across as arrogant, too self-deprecating, insufficiently positive about a colleague who had a rough year, or too vague about someone who needs to hear something difficult.

That's a lot of simultaneous processing for what is, at its core, a writing task. And writing tasks are one of the more reliable things AI tools can assist with, as long as you stay in the driver's seat. The AI can't replace your judgment about what actually happened or what someone actually needs to hear. But it can take your rough notes and shape them into something readable. It can help you find language that's direct without being harsh. It can help you turn a bullet list of accomplishments into a paragraph that actually sounds like an assessment rather than a grocery list.

The key is giving it enough to work with. If you put in nothing, you get nothing useful back. If you put in the specific details — what someone actually did, what the outcome was, where they struggled — the output becomes genuinely helpful.

Writing Your Own Self-Assessment

Start by doing a brain-dump before you touch any AI tool. Open a notes app or a doc and write down, without editing yourself, the things you worked on over the review period: projects you contributed to, problems you solved, things that went well, things that were harder than expected, skills you developed, moments where you stepped up. Give yourself ten minutes and don't filter. Include specifics — numbers if you have them, names of projects, rough timelines.

Then take that dump to ChatGPT or Claude and give it a prompt along these lines:

What comes back will probably be too polished and slightly too generic — AI tends to smooth out the rough edges that actually make writing sound like a specific person. Your job after that is to edit it back toward your own voice, put your actual numbers and specifics in where they've been softened, and remove any phrases that sound like they came from a management textbook ("demonstrated a commitment to cross-functional collaboration" is the kind of thing AI loves and humans find meaningless).

The useful thing about this approach is that the hardest part — getting something on the page — is already done. You're editing rather than generating from nothing, and editing is significantly easier.

A Specific Problem: Sounding Too Modest or Too Much

Self-assessments have a particular calibration problem. Many people, especially those who aren't naturally self-promotional, write assessments that undersell what they actually did. Others write in a way that sounds disconnected from the team effort. AI can help calibrate both.

If you think your draft is too modest, try this:

If you're worried you're overclaiming or coming across as someone who takes all the credit:

These are the kinds of second-opinion checks that are genuinely useful — not because the AI has some special insight into your workplace, but because asking the question out loud (even to a machine) forces you to look at the draft from a different angle.

Writing Reviews for Your Team

If you're a manager writing reviews for others, the challenge is different. You may have six or eight people to write about, you probably have notes from throughout the year that are scattered across emails and documents, and you need each review to feel specific and individual rather than copy-pasted from a template.

The same basic approach applies: gather your notes on the person first. What projects did they work on? What went well? What were the development areas? Any specific moments — positive or challenging — that defined their year? Then use AI to shape that into review language.

One thing to watch: AI has a strong pull toward softening critical feedback into something toothless. "There are opportunities to further develop their communication approach" is the kind of phrase that sounds like feedback but communicates almost nothing to the person reading it. If your notes say someone missed deadlines consistently or struggled with client relationships, you need the review to actually say that clearly, even if diplomatically. Push back against AI's tendency to euphemise.

You can counter this by being explicit in your prompt:

A Note on What Stays With You

There are parts of performance reviews that AI cannot and should not write for you. If you have a substantive concern about someone's performance — something that affects their role, their future with the organisation, or their colleagues — the judgment about how to frame that, and whether the review is the right place to raise it, belongs to you. AI can help you find the words once you've made those calls. It should not be making the calls itself.

Similarly, a self-assessment that describes things you didn't actually do is not a time-saving measure — it's a different kind of problem. The AI doesn't know what's true; it only knows what you tell it. If you feed it accurate notes, you get accurate output you can shape. If you feed it wishful thinking, you get polished wishful thinking.

Worth saying clearly because it's sometimes forgotten when something is generated quickly: the words in your performance review carry your name and your credibility. Review what comes back carefully. Change what doesn't sound like you or isn't accurate. Don't submit a first draft because it arrived fast.

A Few More Practical Uses

Once you've drafted a review, AI can help in a few other specific ways. You can paste a completed draft and ask Claude to check whether the feedback for each person is specific and tied to examples, or whether any sections are vague enough that they could apply to anyone. This is a useful consistency check when you've been writing for two hours and your judgment about specificity has started to blur.

You can also use it to check tone across multiple reviews. If you've written six reviews and you're worried some came out harsher than others for similar situations, paste both and ask whether the framing is consistent. You don't want someone to notice their review is written in a noticeably different register from their colleague's.

And if you're writing a review for someone who had a genuinely difficult year — one where the right message is "we want to support you getting back on track" rather than "we're documenting problems" — AI is useful for finding language that's honest about what happened while still orienting toward a forward path. That particular balance is hard to strike when you're writing under time pressure, and a draft to edit from is much easier to work with than a blank field.

The broader point is that performance reviews are high-stakes writing that most people do under time pressure without much practice. Using AI as a drafting assistant — not a ghostwriter, not a replacement for your judgment, but a way to get past the blank page faster — is a reasonable use of a tool that's already available to you. The review that results is still yours. You're just not starting from nothing.