AI Hiring

Candidate Ranking

An ordering is a tool for spending attention. It becomes a problem the moment it starts being read as a verdict.

The short answer

  • Ranking exists so that limited review time goes to the applications most likely to reward it. That is the whole claim.
  • Small differences in rank are usually noise. Treat the top group as a group rather than reading position 3 as better than position 5.
  • Always be able to see why. A ranking whose reasoning is not visible cannot be checked, corrected or explained.
  • Sample below the cutoff regularly. It is the only way to find out that your criteria are excluding people you would have hired.

What ranking is for

With sixty applications and two hours, the question is not who is best, it is where those two hours should go. Ranking answers that second question, which is the one you actually face.

The gain is real and it is bounded. You still have to read, and the reading is still where decisions get made. What you avoid is spending the first ninety minutes on applications that a consistent first pass would have placed at the bottom.

Do not over-read the order

A ranked list invites a precision it does not have. Positions 3 and 5 are rarely meaningfully different, and the gap between them is smaller than the gap between two human reviewers reading the same pair.

Work in bands instead. A top group worth reading in full, a middle group worth sampling, and a tail worth checking for parsing failures. That maps onto what the ordering can actually support.

It also protects against a subtle failure: a candidate at the top of the list acquires authority simply by being there, and later evidence gets interpreted more charitably for them. Reading the top group as a group makes that harder to fall into.

Insist on the reasoning

An ordering with no visible reasoning is not reviewable. You cannot tell whether a candidate is third because of something that matters or because of a requirement written carelessly, and you cannot explain the decision afterwards.

So use the underlying evidence rather than the position: which requirements were met, where evidence was thin, what the analysis actually found. When a decision is later questioned, those are the reasons. The rank is not a reason.

This is also the practical route to improving the criteria. Reasons can be argued with; a number cannot.

Sample below the line

Every ranking has a cutoff, formal or not, and everything below it is invisible by default. If the criteria are wrong, that is exactly where the evidence of it sits.

Read a handful from below the cutoff on every role. It costs a few minutes and it is the only mechanism that will ever tell you that a requirement is excluding capable people. Career changers and candidates with unusual paths are the ones most often found there, because their evidence is real but does not present in the expected shape.

Frequently Asked Questions

Deciding where limited review time goes. With sixty applications and two hours, the practical question is not who is best but which applications deserve the reading, and ranking answers that. The reading is still where decisions are made.
Usually not meaningfully. Small differences in rank are mostly noise, and the gap between adjacent positions is typically smaller than the difference between two human reviewers assessing the same pair. Work in bands: a top group read in full, a middle group sampled, and a tail checked for parsing failures.
Read the top group properly, but sample below the cutoff on every role. Everything below the line is invisible by default, so if your criteria are excluding capable people that is exactly where the evidence sits. Career changers and unusual paths are found there most often.
The reasons from the application rather than the position on the list: which requirements were met, where evidence was thin, and what was actually found. A rank is not a reason, and a decision recorded as a number cannot be checked, corrected or explained later.

See the reasoning, not just the order

Applications arrive with the evidence surfaced, so a ranking can be checked rather than trusted.