A performance review shouldn't be an annual event

Editorial

The traditional performance review rests on a shaky foundation. It runs on a manager's memory of someone's year, written down the week the form is due, and treats that memory as the official record, even though it is one of the least reliable inputs a company has.

I'd go further. The annual review survives because, until very recently, the two things it assumes were basically true, that a manager's memory was the best available evidence of how someone worked and that performance could only be judged at scale on a calendar. AI breaks both at once, and so performance management should stop being an event and become part of the work itself.

Start with the evidence. The vast majority of what an organization produces is unstructured, and that is often where the truth about performance lives. The commits and code reviews, the call recordings, the documents shipped, the tickets closed, the Slack thread where someone quietly unblocked a teammate, etc. The old review ignored all of it and ran on a manager’s rough summary. When a model reads the work directly, the evidence shifts from what a manager remembers to what the person actually produced.

Then, once the evidence changes, the calendar follows. Reviews attach to real milestones (a launch, a quarter of quota, a project delivered) and feedback arrives while it can still change the outcome. The dreaded once-a-year conversation quite often turns into a summary of feedback that already happened. And promotion-readiness can be checked whenever it matters, instead of argued once a year by whoever speaks most convincingly in the room.

Also, there is a quieter win that I think matters most, and it is fairness. A single manager’s rating carries that manager’s blind spots and recency bias. Reading and reconciling many independent signals is the most reliable way we know to correct for that. A wider base of evidence is simply a fairer one.

We have been running performance this way at Comp for a while.

There is one risk I would not publish this without naming. Reading someone’s work continuously, artifact by artifact, can easily slide into surveillance. What keeps it on the right side is what the system is allowed to read, what it is told to ignore, and whether the employee sees the same view their manager does. Get that wrong and the fairness gain becomes a collapse of trust (worse than the annual review you were trying to replace).

To be clear: none of this removes the human in performance review. Someone still decides! When a manager edits the summary the system drafted, that edit should feed back as a signal that teaches the system what good work looks like in that company. And this makes next year’s review better than this one’s. And when it works, managers stop rebuilding the case from memory and spend that time with the person instead.

Next in this series, the same lens on another HR subfunction.

Once again, let us know what you thought of this piece. The newsletter compounds when you do.

Hope you enjoy what follows,

Chris Gerlach · Co-Founder & CEO, Comp