Cut the Round That Has Never Changed a Decision

Which interview round should you actually cut?

The one whose verdict has never differed from the round before it. That test beats length, cost or candidate complaint, because it asks what a round contributed rather than what it consumed. In most loops it points somewhere specific: the async video answer, the generic behavioural hour, the second conversation with a different job title sitting in the chair.

The funnel data shows the same structure from the other end of the loop. In one large applicant-tracking dataset covering more than 54 million applications and 93,000 jobs, recruiter screens pass roughly 35% of the candidates who reach them while later stages convert far higher, at 95% post-onsite and 81% at offer ¹. Those are stage-to-stage rates among candidates who reached each stage, not an end-to-end funnel, stage names are configured per customer, and the heaviest filtering happens earlier still at application review, which is not one of these numbers.

The 95% is the figure worth sitting with. At that rate the onsite is doing very little independent filtering: the decision has effectively been made before it ends, and the late stages ratify a call already taken. That is a finding to act on, and it says nothing about how demanding an onsite is. The action is not automatically to delete the onsite, because it may be redundant or it may be the round everything else is quietly leaning on, and those need opposite responses. Which funnel metrics still mean anything now covers what else shifted underneath these rates.

Test whether a round has ever changed a decision

Take the last year of closed reqs and, for each round, count the decisions that turned on it. A decision turned on a round when the verdict entering it and the verdict leaving it disagreed, and the later one won. Rounds with a count of zero are confirming. Rounds you cannot count at all are telling you something about your records, which is a finding in its own right.

Run the count, then check which of your rounds rests on an unsupervised document, because that is where the signal has moved. In a blind study run through a UK university's real examinations system, researchers submitted entirely AI-written GPT-4 answers under 33 fake student accounts across five undergraduate psychology modules: 94% of those submissions were never flagged, and the AI grades averaged just over half a classification boundary higher than the real students ². That is coursework rather than a hiring round, the tasks were short-answer and essay questions a language model handles unusually well, and a submission counted as flagged if a marker raised a concern of any kind, so the 94% is the share that drew none at all. The submissions were unedited model output, which the authors call the most detectable possible use of the tool. Treat it as a floor on what passes unremarked.

The async video round has a number of its own now. A peer-reviewed study of machine-scored video interviews reports an uncorrected sample-weighted correlation of .24 with job performance across five organisational samples, against .32 for human-rated structured interviews in the same uncorrected terms ³. Four of the five authors were employed by the vendor whose algorithms were evaluated, the estimate rests on five samples against the 105 effect sizes behind the comparison, and both figures are uncorrected, which makes the comparison internally fair and externally thin. Take .24 as what those five samples could establish about the format. It is not evidence that machine scoring matches a structured interview, and whether the async video round is worth keeping is the case that has to be argued round by round.

What do you do when you hire six people a year?

Use the weaker version of the test, which works at any volume. Have every interviewer state, in writing, the verdict they would give on the evidence they personally collected, before the debrief opens. Then count how often a later round's independent verdict differs from the earlier one. That measures the same property at the round level without needing a year of completed hires behind it.

Two conditions make the count mean anything. The verdicts have to be independent, so nobody reads anyone else's notes first, and they have to be written in the same vocabulary across rounds, so a strong yes in round two is comparable to a strong yes in round three. Neither is expensive. Both are usually missing, and that absence is why the question feels unanswerable. The first round has a one-field version of the same discipline: record what you would have decided from the application alone, before the call, and whether the screen is producing any information becomes countable over fifty of them.

Candidate complaint is the other criterion the genre reaches for, and it needs its context attached. In a 2004 meta-analysis of applicant reactions, 53.5% of samples involved people who were not actual applicants and 90% of the hypothetical-context studies used college students; correlations between justice perceptions and behavioural intentions ran stronger in hypothetical settings than in authentic hiring, and the authors conclude that the role of fairness may be overestimated in studies conducted in hypothetical selection contexts ⁴. Candidate feedback still tells you things nothing else will. As the criterion that picks which round dies, it is weaker than a direct count of what each round changed. The withdrawal signal worth acting on is a specific one anyway: a candidate who is sharp in the remote screen and flat in the onsite is a coverage problem, not a length problem.

Name every round a gate or a sell

Label each round before you remove any of them. A gate round produces evidence and casts a vote. A sell round exists so the candidate meets the people and the work, and it should not vote at all. Most loops that feel too long contain two or three sell rounds quietly voting, which is both where the length came from and where the debrief disagreement comes from.

Relabelling is often the whole fix. A round that everybody values and that has never changed an outcome is usually a good sell round wearing a gate's clothes: keep it, shorten it, tell the candidate what it is for, and take it out of the scorecard set. The loop gets shorter in decision terms without losing the part people actually wanted. Calendar time moves less than that suggests: in SHRM's benchmarking data the median nonexecutive role took 44 days from requisition to accepted offer while the median organisation spent 7 days conducting interviews ⁵, so most of the wait sits outside the rooms you are arguing about.

Cutting a gate round is the heavier move, because removing it removes its evidence. That is safe only once you can point at the cell it occupied and say what will now go uncollected. Giving every round a cell no other round owns is the map that makes an empty column visible before you find it in a bad hire six months later.

On Monday, write your loop out as a list, mark every round gate or sell, and for each gate round write the one piece of evidence it produces that no other round does. Any gate round where that line comes out vague is your candidate. Cut it, run four reqs without it, and keep the per-round verdicts so the next version of this argument has numbers in it instead of opinions.

See how it works >