Evaluating people fairly

When you evaluate people, your memory isn't a neutral record. The halo effect is one of the strongest distortions: one vivid, positive impression colors everything else you think about a person, including traits it says nothing about. It works in reverse too: one bad moment can darken a whole review.

Taras is the case in point. In July, when the payment page broke, he carried the queue, and Leo thanked him in the company chat. Since then it's easy to think of him as the most careful coordinator on the team as well. But speed in a crisis and public praise say nothing about care. Carefulness shows in reopened tickets and survey answers, and there Taras has the team's weakest record.

The defence is to judge each criterion on its own evidence. Ask, for every claim you're about to write: which fact shows this trait, not a different one?

The recency effect makes what happened last weigh more than what happened for months. In a review season at the end of August, July is vivid and February is a blur.

The check is simple: put the striking week next to the rest. Taras's week of 12 July, the payment-page fault: 2 reopened out of 152 closed, 1.32%, with 24 of 25 survey answers positive. A great week. But his other July week, 5 July, was 9 of 93, 9.68%. And January to June he reopened 9.51% and got 78.10% positive answers on 790, above 8% reopened every month.

So the week is the exception, not a new level. It still belongs in the review, as a fact about how he handles a flood. It just can't replace six months of evidence. A good habit for any review: before you write about a period, compute the same measure for the whole period and for each part of it.

Even a real, measured number can mislead when it rests on little data. Taras's 96% satisfaction in the week of 12 July comes from 25 survey answers. With a base that small, a single answer moves the rate by 4 percentage points, and one unusual week can swing it far from the person's usual level. His 78.10% from January to June rests on 790 answers: far steadier evidence.

The same goes for weight in the whole year. The July week is one of about thirty weeks Taras worked this year and 5.37% of his tickets. Counted into the seven-month total, it moves his reopen rate only from 9.51% to 9.07%. That's the honest size of its effect.

So ask two questions of any striking figure in a review: how many observations is it based on, and what share of the whole period does it cover? If the answer is "few" and "small", report it as a fact, not as the verdict.

Halo and recency push in the same direction with Taras: one vivid, recent, publicly praised week tempts you to rebuild the whole review around it. The fair answer is neither to drop the week nor to let it decide.

The week is real evidence of something specific: how Taras handles a flood under pressure, which matters for a support team. It goes into the review with its numbers: 152 tickets closed during the payment-page fault, 2 reopened.

But it isn't evidence that his quality has improved. His other July week (9.68% reopened) looks like January to June (9.51%), the week is one of about thirty, and its 96% satisfaction rests on 25 answers. Saying "his quality has improved" would turn an exception into a trend.

A fair line sits both facts side by side: his strength under pressure, and his usual reopen rate as the area to work on next year.

Memory keeps the dramatic and the recent and drops the rest. A fact log keeps the rest: dated, observable work facts and their impact, good and bad, for every person, across the whole period.

One line per fact is enough: a date, a behavior anyone could have seen, and what it changed. "14 April, Bohdan: three hours pairing with a seasonal hire on refund tickets; her hand-ups on refunds stopped the next week." Not labels ("slow, not a team player"), not impressions ("great month!"), and never guesses about private life ("maybe something at home?").

Two rules keep it fair. Everyone gets entries, not just the people who stand out, or the log simply records your attention. And no hidden tracking: it holds work facts the person could see and discuss, nothing else. If you haven't kept one since January, rebuild what you can from the monthly facts, the incident log and your one-to-one notes, then start fresh: five minutes each Friday.

The third common distortion in reviews is similar-to-me bias: we rate higher the people who work the way we do, because their way looks obviously right to us.

You were the fast coordinator who grabbed hard tickets. Taras is fast too. Bohdan takes about 35 minutes to reply and argues with the AI reply assistant. Without numbers, Taras feels like the better coordinator. But look at what doesn't resemble you: Bohdan logged the most mentoring hours this year, 36, while the seasonal hires learned the job; quiet Maryna has the fewest reopened tickets on the team. None of it looks like you, and all of it counts.

A related trap: different working styles get read as character. Bohdan questioning the AI drafts may be exactly the care that wrote the list of what a draft must never say. Before a style becomes a minus, ask what it produces for families and for the team.

The practical defence against halo, recency and similar-to-me is a fixed process, decided before you look at anyone.

1. Write the criteria first, before opening anyone's numbers: quality for families, volume, help to the team, reliability. Criteria written after you've seen the data tend to bend toward your favorite.
2. Same ruler, same period, for all four: January to July for everyone. Using July alone would be the recency effect in policy form, and in July each person worked only two weeks.
3. Ask each person for a short self-assessment first. It brings in facts you didn't see and shows where your views differ.
4. Get a second reader. Iryna reads all four drafts and asks, line by line: what's the fact? An adjective without a fact gets sent back.

None of this makes evaluation mechanical. Judgement stays; it just has to stand on the same evidence for everyone.

When someone else's number looks weak, we tend to explain it by who they are ("he's slow"); when our own does, we blame circumstances ("I had the hardest week"). This gap is one of the best-documented errors in judging people, and it bites hardest in reviews.

So before a weak number becomes a judgement about a person, check the situation behind it: which ticket types they handled, what else was on their plate, what blocked them, whether the brief was clear. Bohdan's first replies are slower than Taras's. But this year he also logged 36 mentoring hours while the seasonal hires learned the job, and he owns video-outage tickets, which take longer. Writing "slow" would punish exactly the work the team needed.

That doesn't excuse everything. If the situation explains part of the gap, say so; the rest is a fair topic for his plan.

How it fits together: halo, recency and similar-to-me distort memory; a fact log, one ruler, self-assessments, a second reader and a situation check correct it. The next lesson looks forward: development plans and a team that runs while you're away.