Using AI for Performance Reviews: What to Hand Over and What to Keep
Review season is about to start, and you will be tempted. You have six, ten, fifteen of these to write, a year of half-remembered work to summarize for each person, and a tool sitting open that can produce something polished in thirty seconds.
Most managers have already given in to some version of that. The question is not whether to use AI on reviews, but how to use it to help write your reviews.
Your team has already noticed
An August 2026 survey of 1,034 US corporate professionals, run by the coaching firm Highwire, asked both sides. 78% of managers said they had used an AI tool for employee feedback or a performance review in the past year. Only 16% of individual contributors said they had been told AI played a role in their most recent review. Another 40% suspected it had, or weren’t sure.
So the reviews are being written with AI and received in a fog. People are guessing.
The quality verdict was split, which is honest and useful. 54% found AI-assisted feedback more specific and actionable. About a third found it more generic (34%) or less useful (32%). That is not a tool that ruins reviews. It is a tool that improves some and flattens others, depending on how it was used.
One note: Highwire sells human-led coaching in skills it describes as ones that cannot be automated, including giving feedback. A finding that AI feedback needs a human touch is a convenient one for them. The numbers are still worth reading. Just read them with that in mind.
Why it costs more coming from you
Here is the part that should shape how you use it.
Peter Cardon and Anthony Coman showed 1,100 working professionals the same congratulatory message from a supervisor, described as written with low, medium or high AI assistance. The writing quality held up. The person behind it did not.
- 83% saw the supervisor as sincere when AI assistance was low. That fell to 40% to 52% when it was high.
- 95% found low-assistance messages professional, against 69% to 73% for high-assistance ones.
And the detail that should make every manager uncomfortable: people judged their own AI use leniently, rating it similarly at every level of assistance. They only became strict when the writer was their supervisor. You will grade yourself the way they grade themselves. They will not grade you that way.
Two things follow. First, the line is not “AI or no AI.” Lightly assisted messages kept their sincerity almost entirely. The damage came from AI doing most of the writing. Second, that study tested a congratulations email. A performance review is about someone’s performance. I would expect the penalty to be larger there, not smaller.
What AI is genuinely good at in a review
Gathering the year. The hardest problem in a review is not phrasing. It is The tendency to weigh what happened recently more heavily than what happened months ago, simply because it is easier to remember. In reviews it means the last six weeks decide the rating for the whole year, so a strong spring gets buried by a rough October and a weak first half gets rescued by a good quarter. It is one of the most consistently documented distortions in performance evaluation. . If you kept notes during the year, even scrappy ones, AI is excellent at organizing them by theme and putting the evidence under each one, so March has as much weight as September. The key is to keep notes through the year, AI is only as good as the information you can provide to it.
Checking your draft. Ask it to flag vague praise (“great team player”) that has no example behind it, any claim you have not supported, and words you use differently for different people. That last one is a quiet bias check that is hard to do by eye.
Checking consistency across your reports. This is the manager’s own version of The step where a manager, or a group of managers, compares ratings across people to make sure the same performance gets the same score. Without it, a generous reviewer and a strict one produce ratings that mean different things, and people are effectively graded on who they happen to report to. . Put your summaries side by side and ask where you have described similar work very differently. It will not make the call for you, but it is good at spotting the inconsistency.
The part you cannot hand over
The judgment. What this person actually achieved, what it was worth, where they should go next, and the specific moments that support it.
AI cannot know those things, because they were never written down anywhere it can see. It will fill the gap with plausible, general language, which is exactly the “more generic” a third of employees reported. A review built that way reads finished and says nothing, which is the same failure as a context failure in anyone’s work: accurate in every sentence and wrong about the person.
And the conversation. The written review is the record. The meeting is the review.
A method that keeps the trust
- Collect evidence all year, in your own words. A running note per person. This is the single biggest improvement to any review, AI or not.
- Check your data rules before you paste anything. A review is personal information about a named employee. If your company has not approved a tool for that kind of data, it does not go in. Anonymize, or use the approved tool, or do this step by hand. Your team’s AI policy should already answer this.
- Ask AI to organize, not to assess. “Group these notes by theme and list the evidence under each” is safe. “Write a performance review for Sam” is where the 40% sincerity starts.
- Write the judgment yourself. The rating, the strengths, the growth areas, the examples. Short and specific beats long and smooth.
- Use AI to check, then fix it yourself. Vague phrases, unsupported claims, inconsistency across reports.
- Tell them what it did. One sentence is enough: “I used AI to organize my notes from the year and check my draft for vague language. The assessment is mine.” That closes the 40% fog, and it is a disclosure that matches what actually happened.
Notice that step 6 only works because of steps 3 and 4. You can say it honestly because the substance is yours.
The takeaway
Most managers are already using AI on reviews and most employees have noticed without being told.
Let it gather, organize and check. Write the judgment yourself, and say so. The research on AI-written supervisor messages shows the penalty comes from AI doing the writing, not from AI being involved, so a lightly assisted review keeps nearly all of its credibility and a heavily assisted one loses about half of it. In the one document where your team most needs to believe you meant it, that is not a trade worth making.
Sources: the manager and employee figures are from Highwire’s “Who’s Managing Whom?” (opens in a new tab) survey, released 6 October 2026: 1,034 US full-time corporate professionals at organizations with 100+ employees, 321 managers and 713 individual contributors, fielded in August 2026. Highwire, by Knopman Marks, sells human-led coaching in feedback and related skills, so it has an interest in this result. The release itself returned an error when I tried to read it, so the methodology comes from the coverage of it, which was consistent across several outlets. The sincerity and professionalism figures are from Anthony Coman and Peter Cardon’s study in the International Journal of Business Communication, as described in the University of Florida’s release (opens in a new tab): 1,100 professionals rating a supervisor’s congratulatory message at low, medium and high levels of AI assistance. It tested a congratulations email, not a performance review, and the argument that the penalty would be larger for a review is mine. Recency bias and calibration are long-standing concepts in performance management rather than findings from either study. The method is mine.
Related: Do You Have to Tell People You Used AI? is the disclosure test behind step 6. Building an AI Culture: Your Systems Count Twice as Much as Their Mindset is the argument that the review process is the thing your team actually reads, which is why this one matters more than most of what you write.