Evidence-Based Performance Reviews: How to Do It

A performance review is supposed to be an assessment of a year. In practice it is usually an assessment of a manager's memory of a year, written in the week it is due, and weighted heavily toward whatever happened most recently.

Evidence-based reviews fix the input, not the form. Same conversation, same ratings scale — but the manager arrives holding a record rather than an impression.

What counts as evidence

Not "notes." Evidence has three properties: it was recorded at the time, it is specific, and it is checkable by the person being reviewed.

The last row matters more than teams expect. A review where the person contributes their own evidence is both fairer and better-informed — most managers do not see the majority of someone's work, and pretending otherwise is the root of most review disputes.

The problem it solves

Three failures, all of which come from using memory as the data source:

Recency. The last six weeks crowd out the first ten months. Someone who had a difficult spring and an excellent autumn is remembered as excellent; the reverse is remembered as a problem. More on recency bias and how to counter it.

Availability. Visible work is remembered; invisible work is not. The person who kept a critical system running without incident produced nothing memorable, which is precisely the outcome they were paid for.

Reconstruction. Asked to assess a year in an hour, managers reason backwards from an overall impression to the specifics that justify it. That is not assessment, it is rationalisation with examples attached.

How to run it

Collect continuously, review periodically

The whole method is that the collection is not an event. If evidence gathering happens in review week, you have rebuilt memory with extra steps.

Practically: work is linked to priorities as it happens, peer feedback is captured near the moment, and 1:1 notes accumulate. At review time the manager's job is to interpret a record, not to remember one.

Separate the record from the judgement

The record is what happened. The judgement is what it means. Keeping them apart is what makes a review defensible — and it is what lets a disagreement be about interpretation rather than about facts.

A review conversation where both people are looking at the same record is a fundamentally different conversation from one where the manager asserts and the employee recalls.

Write the evidence first, the rating second

Draft the evidence summary before deciding the rating. Reversing that order reintroduces the reasoning-backwards problem the method exists to prevent.

Let the person add to the record

Before the review, not during it. They will surface work the manager never saw — which is the point, not an inconvenience.

What it does not fix

Worth being straight about, because evidence-based reviews get oversold:

It does not remove judgement, and shouldn't. Deciding what a body of work means is the manager's job. Evidence constrains the judgement; it does not replace it.

It does not make ratings objective. Two fair managers can read the same record differently. The gain is that they are now disagreeing about the same facts.

It does not fix a broken rating scale, unclear expectations, or a distribution the company enforces regardless of performance. Those are separate problems and evidence makes them more visible rather than less.

It does not eliminate bias. It narrows the gaps where bias operates unchallenged. That is a real improvement and it is not the same as a solution.

What changes when you do it

Reviews get shorter to write and longer to discuss. The writing is assembly rather than recall; the conversation has more to work with.

Disagreements become specific. "I don't think that's fair" becomes "I think this project mattered more than the record suggests" — an argument you can actually resolve.

The gap between visible and invisible work narrows. Maintenance, enablement and helping colleagues show up in a record even though they don't produce memorable moments.

Calibration gets easier. Comparing two people is possible when both have records; without them, calibration compares how well two managers write.

Getting started without a system

You do not need software to start. A workable minimum:

1. Every manager keeps a running note per report. One line per notable event, dated, added within a week of it happening.
2. Capture peer feedback at delivery, not at year end — one question, at the moment a piece of collaborative work finishes.
3. The person keeps their own list, and it is merged before the review.

That alone removes most of the recency problem, and it will immediately show you how much of someone's work a manager does not see.

The reason to automate afterwards is not accuracy — it is that a record assembled by hand gets abandoned in a busy quarter, and the quarters when it lapses are exactly the ones you most need it for.

Related

• Recency bias in performance reviews — the specific distortion this method counters
• Why annual performance reviews don't work — the case against the annual cycle itself
• Continuous performance management — the operating rhythm this fits into
• How company, department and individual KPIs connect — evidence needs something to be evidence against
• How to audit what your AI assistant did — if any part of your evidence is AI-assembled