AI in performance reviews: what an AI summary can and can't do

An AI summary of 360 feedback is an overview, written by AI, of everything reviewers said about one person, organised into strengths and areas to develop. It’s the most useful job AI does in performance reviews, because it replaces hours of reading and sorting. It’s also the output to check most carefully, because it decides which comments get noticed.

Most writing about AI in performance reviews, and about AI in performance management more broadly, is about drafting: getting a chatbot to write the review for you. This piece is about the other use, summarising what several people wrote about a colleague, and the questions that come with it.

Does it keep reviewers anonymous? Can it be biased? When is it wrong?

The last section describes how Lynxify.me does it, in enough detail for you to hold it against everything before it.

Key Takeaways

• An AI summary condenses ratings and written answers from several reviewers into themes for one person. It replaces the reading and sorting, and leaves the judgement and the conversation to the manager.
• Anonymity depends on three things: names removed before the text is processed, a summary that reports themes without attributing them, and care with small groups. A team of three is rarely anonymous, whatever the software does.
• A summary inherits the bias in the comments, and it can add its own by folding a lone dissenting comment into the majority view.
• Don’t rely on a summary with fewer than about four reviewers, when reviewers disagree, or for any decision about pay or employment.
• Before buying, ask what is sent to the AI model, whether names are removed, whether your data trains a model, and whether the original answers stay visible next to the summary.

What an AI summary of 360 feedback actually does

In a 360 review, one person gets feedback from their manager, a few peers and, if they lead a team, their direct reports. That’s five to seven sets of ratings and written answers. An AI summary reads all of them and writes a short text: what people agree this person does well, where they’d like to see a change, and often which competencies each point belongs to.

How the summary is built decides how useful it is. Some tools pass the model one block of comments, which tends to produce generalities like “communicates well, could delegate more.” Others build the summary from each question’s rating and written answer, so it can say something about a specific competency and the behaviour behind the score.

What a summary replaces is the slow part: reading twenty answers and finding the pattern. It doesn’t decide which of those patterns matters for this person right now, and it doesn’t hold the conversation. That work stays with the manager.

Does an AI summary keep 360 feedback anonymous?

It can, when three things are true.

1. Names are removed from the responses before the text is processed, so the model works on what was said and never sees who said it.
2. The summary reports themes without attributing them. “Several reviewers mentioned late handoffs” is safe, and “one of the two direct reports said” is a signature.
3. Small groups get extra care. If only two peers answered anonymously, showing their answers per question lets the reader work out who wrote what.

Identity can still leak in ways software doesn’t fully control. A summary that repeats a distinctive phrase points to its author. A detail only one person could know does the same. And in a team of three, everyone can guess, with or without AI.

So tell reviewers plainly what happens to their words: who sees the original answers, what the summary shows, and how small groups are handled. I’ve written more about when anonymous feedback works at a small company, and the short version is that people answer candidly when they know the rules in advance.

Can an AI summary be biased?

Yes, in two ways.

The first is inherited. A summary is made of the comments it was given. If the reviewers wrote vague praise for one person and detailed criticism for another, the summary repeats that difference in tidier sentences. The same goes for the bias problem in 360 reviews more generally: recency, friendliness and who happens to be asked all pass straight through.

The second is added by the summarising itself. A model looking for themes will fold one dissenting comment into the majority view, treat a vivid remark as a pattern, or turn a hedged “sometimes seems rushed” into a firm “rushes work.” Each of those changes the picture a little.

There’s one place a summary reduces bias. It reads the fifth form with the same attention as the first, which a manager at the end of a long day doesn’t. That’s a real advantage for consistency, and it’s the reason I still think the feature is worth having.

When not to trust an AI summary

Treat the summary as a starting point, and be more careful in these cases:

• Fewer than about four reviewers answered. There’s too little material for a pattern, and one voice weighs too much.
• The reviewers disagree. A summary tends to average disagreement into something mild, and the disagreement is usually the interesting part.
• One comment is specific and critical while the rest are general and kind.
• The round has ratings but few written answers. Numbers without explanations give a model nothing to summarise.
• The comments use sarcasm, in-jokes or shorthand the team understands and a model doesn’t.
• The result will feed a decision about pay, promotion or someone’s job.

Imagine a round with five reviewers. Four write that a team lead is “great to work with” and “always helpful.” The fifth describes two handoffs where the lead agreed to a deadline and went quiet until it had passed.

A summary of that round can come back as “generally positive, with some room to improve communication.” The one comment a manager could act on has been smoothed away.

The rule that follows is simple. Before you say anything from a summary out loud, read the original comments behind it.

What the manager still has to do

A summary saves reading time. Four jobs remain.

Read the original answers for every point you plan to raise. Add what the summary can’t know: the reorganisation in March, the project that was cut, the fact that two reviewers joined a month ago. Choose one or two things to talk about, because a person can’t work on six. Then hold the conversation, which is where the feedback becomes useful or doesn’t.

Keep three kinds of text apart while you do this: what the reviewers wrote, what the AI synthesised, and what you think. When they blur together, nobody can tell later whether a sentence was evidence or interpretation. There’s a full walkthrough of the conversation itself in how to deliver 360 feedback results.

Can AI write the performance review itself?

That’s a different job, and it’s worth separating the two. Summarising condenses evidence that already exists: several people wrote down what they saw. A chatbot asked to write a performance review from a prompt has no evidence to work from unless you supply it, so it fills the gap with sentences that sound right.

If you draft with a general AI tool, bring the specific moments yourself, and use it to shorten, reorder or check your tone. Keep names and confidential details out unless your company has approved the tool for that. Then rewrite the result so it sounds like you, because a review that reads as generated tells the person you didn’t spend the time.

Eight questions to ask before you buy an AI summary feature

1. What exactly is sent to the AI model, and are names removed first?
2. Is our data used to train a model?
3. Is the summary built per question or competency, or from one block of text?
4. Can I always see the original answers next to the summary?
5. Is the summary kept separate from my own notes?
6. What happens when only two or three people respond?
7. Could an anonymous reviewer be identified from the summary?
8. Is it included in the price, or sold as an add-on?

A vendor who answers these quickly has thought about them. For the wider buying decision, what small teams need from a 360 tool covers the rest.

How Lynxify.me does it

Here is our implementation, answered against the same questions.

• When it runs: once a feedback round is finished, the manager who owns it can request an AI summary. It produces one summary per person.
• How it’s built: from each question’s rating and written answer, so the summary can speak to specific competencies.
• Names: removed from the responses before they are processed.
• The model: processing is done by an external AI provider, named in our privacy policy, under terms that opt out of using the data to train models.
• Small groups: per-question detail from anonymous reviewers is shown only when enough of them answered to keep it unattributable, and answers that can’t safely be separated are pooled before summarising.
• Separation: the summary sits in its own card on the report, apart from the reviewers’ original words and from the manager’s private notes. One click copies it into your notes as a draft, and the original stays unchanged.
• Freshness: each summary shows when it was generated.
• Original answers: the full written responses and scores stay on the report. The summary is a way in and doesn’t replace them.
• Access: only users with management access to that round can see the summary.
• Price: AI feedback summaries are included in the free plan.

Everything in the earlier sections applies to our summaries too. With three reviewers I’d read the comments and skip the summary, and I wouldn’t use it, or any summary, as the basis for a pay decision.

A summary is a way in

The useful question about AI in performance reviews is which part of the work it takes over. Reading and sorting twenty answers is a good part to hand over. Deciding what matters, checking it against the original words, and saying it to a person are the parts that make feedback worth collecting, and they stay with you.

If you want to see what a summary looks like on real feedback from your own team, you can start free with Lynxify.me. You sign in with Google or Microsoft, and there is no setup call. If you already have a tool, the eight questions above work on it just as well.

Frequently asked questions

What is an AI summary of 360 feedback?

An AI summary of 360 feedback is an overview, written by AI, of everything reviewers said about one person, organised into strengths and areas to develop. It is built from the ratings and written answers that a manager, peers and direct reports submitted. It replaces the reading and sorting, and it leaves the judgement and the conversation to the manager.

Does an AI summary keep reviewers anonymous?

It can, when three things are true: names are removed before the text is processed, the summary reports themes without saying who raised them, and small groups are handled with care. No software makes a team of three anonymous, and a summary that repeats a distinctive phrase can still point to its author. Ask a vendor how each of those three points is handled.

Can I trust an AI summary of performance feedback?

Yes as a way into the feedback, and no as the final word. A summary is least reliable with fewer than about four reviewers, when reviewers disagree, when one specific critical comment sits among mild ones, and in rounds with ratings but few written answers. Before you repeat anything from a summary, read the original comments behind it.

Can I use AI to write a performance review?

You can use it to draft and edit, as long as you supply the evidence. A general chatbot knows nothing about your colleague, so it produces plausible sentences with no real moments behind them. Bring the specific situations yourself, keep names and confidential details out unless your company has approved the tool, and rewrite the result in your own words.

Which 360 feedback tools include AI summaries?

Several do, and they differ in ways that matter more than the label. Check whether the summary is built per question or from one block of comments, whether names are removed before processing, whether the original answers stay visible beside it, and whether it is included in the price or sold as an add-on. Lynxify.me includes AI summaries in its free plan, built per question, with names removed before processing.