Budget the Whole Loop in Candidate Hours, Not Round by Round

How many candidate hours is a loop allowed to cost?

Somewhere between four and six hours, end to end, for a non-executive role. That covers the application, a screen, one scoped exercise, a panel and the scheduling overhead around them. No professional body publishes a whole-loop figure, so treat that range as a working default you own, not a standard you can cite at a hiring manager who wants a sixth round.

There is no standard because the genre prices one stage at a time. How long an interview should run. How long a take-home should take. Each answer is defended on its own terms and each is roughly reasonable, so the total grows by one reasonable increment at a time until it lands well past anything the team would have agreed to up front, and no single person made that decision.

Employers already measure what a hire costs. SHRM's benchmarking data, from a random sample of its member organizations, puts median cost-per-hire at $1,244 for nonexecutive roles, with the mean in the same table at $4,683, and the definition sums agency fees, advertising, job boards, referral costs, travel, relocation, recruiter pay and the recruiting system, divided by hires ¹. Candidate hours are not in that list. Neither is interviewer time. That distribution is heavily right-skewed, which is why the median and the mean sit so far apart, and why no single figure describes what a hire costs.

So write your own number down before the next intake meeting. The argument coming is about a sixth stage, and you will lose it if the only figure in the room is the one attached to that stage alone.

Track reviewer minutes in a second column

Write two numbers beside every stage: what it costs the candidate and what it costs your side to read. The second column is the one that moved. A polished submission now arrives faster than a reviewer can work out what it demonstrates, so a stage priced as cheap on the candidate's clock can be the most expensive hour in the loop on yours.

Run one live req through the list. Stage, candidate minutes, reviewer minutes, with the second number counting notes and scheduling as well as the meeting itself. Filled in for one engineering loop, it looks like this.

• Application. Eight candidate minutes on a short form, three reviewer minutes if a person reads it at all.
• Recruiter screen. Thirty candidate minutes, closer to forty of yours once notes and rescheduling are counted.
• Take-home. Four candidate hours as briefed, twenty minutes as actually spent, and an hour or more of reviewer time to work out which of those two happened.
• Panel. Three candidate hours, six of yours across four interviewers and the debrief that follows.

The third row is the one that should stop the meeting. It is the stage everyone defends as respectful of the candidate's time and it is now the worst trade in the loop: least evidence per candidate hour, most reviewer hours per candidate hour. Once a stage costs a candidate more than an hour and hands you a document you then have to interrogate, the honest options are to shrink it or to pay for it, and paying candidates for take-home time is a narrower decision than it first sounds.

Which stages actually earn their hours?

A stage earns its hours when it produces evidence no earlier stage produced and someone can name that evidence in one sentence. Run that test stage by stage on a live req. If two stages return the same kind of evidence, the second is buying confirmation rather than information, and confirmation is what you end up paying candidate hours for by accident.

Two failures dominate. One is duplication: two conversations in which a prepared candidate narrates past work are one conversation held twice, whatever the question lists look like. The other is newer and easier to miss, because the artifact still arrives on time and still looks like the thing it used to be.

Exercises assembled from public problem banks are the clearest case of the second. Evaluated against 115 Python problem statements taken from HackerRank, OpenAI's Codex solved 96% of them zero-shot and 100% few-shot, with clear signs it was reproducing memorized code rather than synthesising it ³. The memorization half is the load-bearing one: a task drawn from a public bank was testing a lookup before any model touched it. That is a 2022 model on a curated benchmark, so treat the figure as a floor. The candidate hours are real either way, and what the exercise proves has thinned to almost nothing.

Volume is not the defence people assume either. Around 13% of hires in one large applicant-tracking dataset included a take-home component ², and that dataset skews to venture-backed technology employers, so the practice is a minority one even where it feels standard. If yours has to survive the evidence test, design against the harder constraint: testing AI skills without adding an hour to the loop.

Who pays for the hours nobody counts?

The candidates least able to absorb them, which makes the total a fairness question and not only a courtesy. Unpaid weekday hours land hardest on people already working, people with caregiving duties and people who cannot move a shift. Those hours never appear in a funnel report, because someone who cannot spare them quietly stops replying.

The reflex when a loop gets too long is to cut the exercise, and candidate-reaction evidence does not support that reflex. Pooling favorability ratings across studies and rescaling everything to a five-point scale, applicants rated interviews highest at 3.84 and work samples next at 3.63, ahead of resumes at 3.57 and references at 3.33 ⁴. Those means come from people rating written descriptions of methods, not sitting through them, from studies published up to 2004, and each rests on between five and ten ratings. Read it as a ranking. That ranking puts the job-shaped exercise near the top of what candidates accept, above the resume screen and the reference call.

So spend the budget where the evidence is and take the hours back somewhere else. A fifth conversation, a second culture round, an async video answer nobody watches closely: those are the hours a shorter loop should buy back, which is the whole of which interview round to cut.

On Monday, take one open req, write both columns, and mark every stage that cannot name evidence unique to it. Cut, shorten or pay for each one. Then put the total in the posting, where it does the most work: a candidate who knows the loop is five hours can plan for it, and a hiring manager proposing a sixth conversation has to argue against a number that is already public.

See a sample report >