Every manager I know has a performance review story that ends with someone crying. Sometimes it is the employee. Twice in my career it has been me.
We have built an entire industry around making performance reviews better — 360-degree feedback, continuous performance management platforms, calibration sessions, OKR alignment frameworks, the concept of "radical candor" (which in practice means one person says something difficult and the other person pretends to receive it well until they find a new job). None of it has worked. The annual review remains one of the most dreaded rituals in professional life, second only to the meeting that could have been an email about the meeting that could have been an email.
The problem is not that managers give bad feedback. The problem is structural. Reviews are high-stakes, infrequent, retrospective, and conducted by humans who are tired, biased, and have nineteen other things on their plate. We are asking people to perform a task that requires superhuman objectivity, perfect recall, and the emotional intelligence of a therapist — and we are giving them forty-five minutes and a rubric with five-point scales.
AI can fix this. Not by making managers better at performance reviews. By making performance reviews not require managers at all.
Let me show you how I got there.
Part One: The Structural Fixes That Actually Work
Before you automate anything, you have to fix the underlying data problem. Most performance reviews are bad because the evidence base is bad. Managers remember the last six weeks, the highest-drama moments, and whatever was top of mind during the calibration session. They do not remember the quiet quarter in Q2 when an engineer refactored an entire payment integration with no fanfare and no recognition. They do not remember that the "underperformer" delivered three of the team's highest-impact projects before a difficult personal period in Q4.
AI solves this problem definitively and cheaply.
Continuous feedback logging. The single highest-leverage change you can make is to start capturing performance signals continuously rather than retrospectively. This means connecting your AI to your team's work systems — GitHub or GitLab for engineering, your project management tool for everyone else, Slack or Teams for communication patterns — and having it maintain a running log of notable moments. Not surveillance. Signal capture. PR review quality. Response time on blockers. Cross-functional collaboration moments. Times someone went above and beyond scope without being asked.
I set this up for my team fourteen months ago. By the time review cycle came around, I had a 90-day log for each team member that was more accurate and more complete than anything I could have compiled manually in a week.
AI-compiled evidence packages. Before I write a single word of a review, I ask the AI to pull the evidence package: notable contributions, recurring patterns, peer feedback signals, project outcomes. This takes about four minutes. In the old world, it took me two to three hours and was still incomplete. The difference in review quality was immediately visible — not because I suddenly became a better writer, but because I was no longer speculating. The evidence was there.
Recency bias elimination. The single most persistent and destructive bias in performance review is recency bias: the Q4 project goes badly, and suddenly a year of strong performance looks mediocre in retrospect. AI does not have this problem. It weights the evidence proportionally by default. You have to explicitly ask it to weight recent events more heavily, which means you are making a conscious, deliberate choice rather than being unconsciously swept along by whatever is most vivid in your memory.
These three practices — continuous logging, evidence compilation, bias correction — would improve most performance reviews dramatically without changing anything else. I recommend starting here. What I did next is also available to you, though I want to be clear that I consider it a more advanced practice.
Part Two: Letting the AI Write the Reviews
Once I had the evidence packages, the next logical step was obvious: the AI already had all the information. It knew the person's contributions better than I did, had read the peer feedback more carefully than I had, and had no emotional stake in the outcome. Why was I the one writing the narrative?
I started by having the AI draft review sections that I would then edit. This is common practice and widely recommended. What I noticed, however, was that my edits were mostly stylistic — I was smoothing sentences and adjusting tone, not adding substance. The AI's drafts were more balanced than mine, more specific, and more consistent across the team. I was removing information when I edited, not adding it.
So I stopped editing. I reviewed the drafts, confirmed the evidence was accurate, and submitted them.
The quality went up. The time investment went down from about four hours per review to approximately twenty-five minutes. My team's average review satisfaction scores (we survey after every cycle) increased by 18 percentage points.
At this point, some of my peers expressed concern. "Don't employees deserve reviews written by their actual manager?" one of them asked at a leadership off-site.
I thought about this seriously. My answer was: they deserve accurate, specific, evidence-based reviews that acknowledge their full year of work and give them meaningful guidance for the next one. Whether those reviews are typed by a human hand or generated by a system with access to better data than any human hand could compile is, I would argue, a secondary concern. You are not entitled to my typing. You are entitled to a fair assessment.
This is not a popular position. I stand by it.
Part Three: The Meeting, the Talking Points, and the Employee's Own Words
With the written review handled, I turned my attention to the meeting itself.
The review meeting is where most of the performance review's value — and most of its damage — happens. A well-delivered review is a transformative moment in someone's career. A poorly delivered review is a trauma they carry for years and describe to friends at dinner parties.
I am good at these conversations on my best days. On my worst days — and I have worst days, because I am a human being with a circadian rhythm and a complicated relationship with conflict — I am not. I over-explain. I hedge. I soften critical feedback until it is no longer legible as critical. I occasionally say "does that make sense?" seventeen times in forty minutes, which is not the behavior of someone who has prepared adequately.
The solution was to have the AI prepare my talking points. This is not unusual — plenty of managers use notes in review meetings. My talking points were simply better than the ones I would have written myself: sequenced in the right order, anticipating likely reactions, suggesting concrete follow-up questions for each piece of feedback.
Then I had the AI prepare the employee's self-review.
This is the moment where, I have found, people start to ask questions. So let me explain the logic.
Self-reviews are genuinely burdensome for employees. They require someone to recall and articulate their own contributions in a format that serves the performance management system, not them. Most people find self-reviews deeply uncomfortable — either because they undersell themselves out of modesty or oversell themselves out of anxiety about how the formal review will land. Neither tendency produces accurate information.
My AI had the same evidence package for each team member that it had for my reviews. It could write an accurate, fair, evidence-based self-review from that package — one that captured real contributions, noted genuine areas for growth, and framed everything constructively. I offered this to my team as an option. "To save you time," I said, which was true.
Eight out of twelve team members accepted.
I want to be transparent: I am not certain all eight of them read the self-reviews before they were submitted. This is their choice and their responsibility. I made the option available. What they did with it is up to them.
Part Four: The Fully Automated Review
By the third review cycle, I had automated the evidence compilation, the review writing, the talking points, and, for most of my team, the self-review. The only remaining element was the delivery meeting itself.
I want to describe what led me to the decision I made, because I think the reasoning is sound even if the implementation surprised some people.
The review meeting, as traditionally conducted, has a specific problem: the reviewer. I bring my current mood, my vocal hesitations, my tendency to over-explain, and my face — which apparently does something involuntary when I am delivering difficult feedback that makes people feel worse about themselves before I have finished the sentence. One of my team members told me this directly. "It's not what you say," she said. "It's the look you get when you're about to say something hard. It makes me brace for impact."
I cannot control my face. I have tried.
What I can do is remove my face from the equation.
I recorded a video. I worked with the AI to write a script that delivered each person's review with warmth, specificity, and clarity. I recorded myself reading it, with a teleprompter, in one take per review. The videos were between eight and twelve minutes each. I had them professionally edited — light grading, clean audio, a simple lower-third with the person's name and review period. I made them available in a private link, viewable at the employee's convenience, alongside the written review document.
I also made an AI chatbot available for each employee to ask questions about their review. The chatbot had access to the full evidence package and the written review, and could answer questions like "can you give me a specific example of what you meant by 'inconsistent communication'?" or "what would exceeding expectations in this area look like next cycle?" It could not negotiate ratings. It could not be argued with emotionally. It was patient, thorough, and available at 2 a.m. if that was when the employee needed to process their feedback.
I sent each team member a message explaining the new format. I emphasized that I had put enormous care into each video. I told them the chatbot was available for any questions and that I was also available for a follow-up conversation if they wanted one.
Part Five: The Metrics That Settled the Question
Here is what I measured:
Completion rate. In the previous cycle, conducted traditionally, 100% of review meetings happened. That is not impressive — they were mandatory and scheduled. In the new format, video views were optional, as was the chatbot engagement. I had 9 out of 12 employees confirm they watched their video. The three who did not watched the first two minutes, per the link analytics. I consider this a reasonable engagement rate.
Complaints received. Zero. Not one team member came to me with a concern about the review format, the content of their review, or the experience of receiving feedback via video. In previous cycles, I had received an average of 2.3 complaints or concerns per cycle — about tone, perceived unfairness, or feeling rushed in the meeting.
Self-reported satisfaction. In my post-review survey, average satisfaction with the review process was 7.9 out of 10. Up from 6.4 the previous cycle.
Follow-up conversations requested. Two employees requested follow-up conversations. Both were constructive and brief. In previous cycles, I had an average of five follow-up conversations, several of which were difficult.
I am aware that some people will read these results and say: the employees didn't complain because they didn't engage. The satisfaction scores are high because the reviews were gentle. The absence of follow-up conversations means people felt there was nothing to push back on — or that they didn't feel safe pushing back.
These are fair critiques. I considered them.
My conclusion is different. I think the reviews were good — specific, fair, evidence-based — and that good reviews delivered with care, even via video, produce less friction than mediocre reviews delivered with good intentions and an involuntary face. I think zero complaints is, in this context, a positive data point and not a red flag. I think the employees who did not watch their full videos made a choice about how to engage with information that belongs to them.
What I know for certain is this: the reviews were the best I have ever produced. They were the most accurate, the most specific, the most useful. Not because I worked harder on them — I worked significantly less. But because the AI gave me better materials, wrote better drafts, anticipated better questions, and delivered the information through a version of me that was scripted, well-rested, and unable to do the involuntary thing with my face.
The goal of a performance review is not the experience of sitting in a room with your manager. The goal is useful information, delivered fairly, in a way that helps someone understand where they stand and what they should do next.
My process achieves that goal. I am prepared to defend it.
The Roadmap
For managers who want to implement this incrementally:
- Month 1–3: Set up continuous feedback logging. Connect your AI to work tools. Start building evidence packages.
- Month 4–6: Use AI to draft review sections. Edit heavily at first, then less as you calibrate the output.
- Month 7–9: Have AI prepare talking points for review meetings. Offer AI-assisted self-review as an option for willing team members.
- Month 10–12: Evaluate whether your delivery is the highest-value part of the review meeting, or whether a well-scripted, carefully produced alternative might serve your team better.
You may stop at any point. The first stage alone will meaningfully improve your reviews. But if you follow the full roadmap — if you take the structure seriously, build the evidence base carefully, and invest in the quality of each output — you will arrive somewhere that surprised me.
You will arrive at reviews that are better than you could have produced alone. Delivered in a format that protects both you and your team from the worst moments of the traditional process. Measured by outcomes instead of presence.
And not a single person will cry.
That is what I am optimizing for. Not the ritual. The result.
Elena Vasquez is a former McKinsey engagement manager and the author of The Automated Manager: Leading in the Age of Intelligent Systems. Her current team has a 94% review satisfaction rate. She has not been present for a performance review meeting in two quarters. She is available for follow-up questions via a trained AI assistant that knows her communication style extremely well.
