How Many Interview Rounds a Role Actually Needs

How many rounds is normal, and why that is the wrong question

Skip the benchmark hunt, because the number would not help. Elapsed time is what the benchmarks actually measure. In SHRM's 2021 benchmarking survey of its member organisations, the median time-to-fill for nonexecutive roles was 44 days, of which the median organisation spent 5 days screening applicants, 7 days interviewing, and 4 days deciding and making an offer ¹. Interviewing is a small slice of a long calendar.

That distribution is worth sitting with, because it reframes the usual argument. The interquartile range around that median runs from 28 to 73 days and the mean is 54, so any single average misleads twice over ¹. More to the point, the days are mostly not the loop. They are waiting on a requisition approval, waiting for four calendars to intersect, and waiting on a decision nobody scheduled. A loop with one fewer round and the same scheduling habits will not be much faster.

So the round count is a proxy for two different complaints that need different fixes. If the complaint is that candidates drop out, the cause is usually elapsed days and silence between stages rather than the number of conversations. If the complaint is that the decision still feels like a coin flip at the end, adding a round has never fixed it, because the new round tends to ask the same thing again in a different room.

The question worth asking instead is what each stage establishes that nothing else does. It is answerable in an afternoon, it produces a shorter loop as a side effect, and unlike a round count it survives contact with a role that genuinely needs four stages.

Name the claim each round is there to test

Write one sentence per stage saying what it establishes that no other stage does. A screen establishes the two or three facts that would end the conversation. A work stage establishes that the person can do the central task. A team stage establishes how they work with other people. Any stage without a sentence of its own is a duplicate, and duplicates are what make loops long.

The test is whether two stages could change the decision in different directions. If a candidate who impressed the panel in stage two was always going to impress it in stage three, stage three is only confirming what stage two already decided. The selection literature puts this in its own terms: a composite built from two or three different kinds of evidence reaches about .61 as a corrected correlation with job performance, and the gain comes from the kinds being different ².

A worked version for an individual contributor role:

• Screen, 20 to 30 minutes. Claim: the constraints match. Location, compensation band, and the two or three things about the role that are not negotiable.
• Work stage. Claim: this person can do the central task, with evidence of how they did it rather than only what came out.
• Team stage. Claim: how they handle disagreement, ambiguity, and a decision somebody else made.

Three stages, three claims. A fourth earns its place only where the role carries a specific risk with evidence of its own: managing people, handling money, or a domain where being confidently wrong is expensive. Where the new claim is about working with AI, the instinct is to bolt on a stage for it, and that is usually the wrong shape (test AI skills without adding an hour to the loop).

Which rounds to cut first

Cut the stage whose claim another stage already covers, then the stage that produces an impression rather than a record. In most loops those are the second behavioural interview and the informal culture chat. Neither is deleted for being useless. They go because the same information is being collected twice and the second collection is the one nobody wrote down.

The culture chat deserves a specific warning, since it is the stage most often defended as costing nothing. Adding a low-diagnostic conversation to good information makes the prediction worse. In a controlled test, students predicting a classmate's semester grades did worse after conducting an unstructured interview, correlating .31 with the outcome against .65 from prior grades alone, and in a further study most participants chose to interview someone answering at random over not interviewing at all ³. That is undergraduates predicting a classmate's grades in a laboratory, a long way from a manager predicting job performance. The mechanism is what travels; the effect size does not.

Two more cuts are usually safe. The panel round where four people ask overlapping questions in sequence, which can become one stage with a fixed question set split between them. And the final conversation with a senior leader that exists to confirm a decision already made, worth keeping only if that person can say no and occasionally does.

Whatever gets cut, keep the claim. If the culture chat was the only place anyone asked how the candidate handles being wrong, that claim moves into another stage with a real question and an anchored scale behind it (what actually makes an interview structured).

What a shorter loop costs, and what it does not

It costs the redundancy that was catching a small number of mistakes, and that is a real trade rather than a free saving. What it does not cost is evidence, as long as each remaining stage keeps its claim and produces a record. The loops that get worse when shortened are the ones where somebody took minutes off every stage and left all of them standing.

Three things belong in place before cutting, all of which cost less than the round being removed. A written claim per stage, so the gap a deletion leaves shows up the same day. An anchored rating scale on the stage doing the heaviest lifting, so its evidence stands up without a second opinion propping it. And a decision rule agreed in advance, because a shorter loop puts more weight on the debrief (collect the scores before the debrief).

The candidate side is not a separate project. A three-stage loop that says what each stage is for, and when the answer comes, reads as a process somebody designed. A five-stage loop with two unnamed stages reads as an organisation that has not decided what it is looking for, which is generally what it is.

Run this per role. A loop is a claim about a specific job, and one template applied to every role describes something other than the position being filled. Where a role genuinely needs an extra stage, the question is the same one: what does it establish that nothing else does (whether an AI-fluency round earns its place).

See how it works >