Split the Screen Into a Cheap Pass and an Expensive One

Why does a flat minute per application fail now?

Because a flat budget assumes applications are interchangeable, and they stopped being that. Volume rose sharply: Ashby's benchmark set reports the average recruiter processing 291 applications per hire against roughly 100 in early 2021, with the share of applications reaching an interview falling from about 7-8% to between 3.6% and 4.7% depending on role type ¹. A uniform pace spends identical attention on a duplicate submission and on the one person worth a call.

Read those numbers with their limits attached. They come from one vendor's customer base, weighted toward venture-backed technology companies hiring knowledge roles, and 291 is a ratio of two organization-level totals, so it is never an individual applicant's odds. Duplicate and automated submissions inflate the numerator, which is exactly why a falling interview rate is not evidence that candidates got worse or that screening got sharper. What it does establish is that the pile grew faster than the hours available to read it, which is the condition a pacing rule has to survive.

Workload moved in the same direction. SHRM's 2026 recruiting benchmark reports a 67% increase in median requisitions per recruiter at extra-large organizations, and over two in three organizations reporting difficulty filling open positions ². The brief never publishes the underlying requisitions-per-recruiter figure and attributes the rise to nothing in particular, so read it as a direction and nothing more. It counts openings, so it measures no reading time at all; what it adds is a second squeeze on the same hours, more requisitions per recruiter on top of the larger pile behind each one. The volume side is unpacked in why applications per opening tripled.

The deeper failure is what a flat minute buys. Thirty seconds is enough to form an impression of a document and not enough to check a single claim in it, so a uniform pace guarantees that every decision rests on impressions. A fast scan worked as a proxy while the fluency on the page came from the person who would show up, and that is the half of the premise which stopped holding: polish is now cheap for anyone to produce, and it is the first thing a thirty-second read rewards. Splitting the pass fixes that without adding hours.

Run the cheap pass on questions with a yes-or-no answer

A question belongs in the cheap pass when a second reader would reach the same answer without exercising judgment. Employment dates and continuity. Location and work authorization. A licence or registration an issuing body lists. A stated minimum somebody wrote down before the pool arrived. Each settles in seconds, settles the same way every time, and can be explained to the person it excluded.

Anything you would have to weigh stays out of the cheap pass:

• Years of experience thresholds. A meta-analysis of 81 independent samples put prehire work experience at a corrected correlation of .06 with later job performance, and about .07 for experience with relevant tasks, jobs or occupations ³. It is fast to read and it is not carrying what a threshold implies it carries. Keep it only where a licence or a legal minimum requires it.
• Impact, achievement depth, polish. All scored from prose, all cheap to produce, and none of them settles the same way twice.
• Whether it reads as generated. Untrained evaluators asked to separate model-written text from human writing performed at chance, and three quick training methods lifted them only to about 55% ⁴. That was 2021-era output judged by crowdworkers rather than by recruiters reading in their own field, and the paper measures nothing about how those odds have moved since. Spending expensive minutes on this question is spending them on a coin flip, and the fix for hiring managers rejecting people for sounding like AI sits in how the process is designed.

The cheap pass has one more property worth protecting: it produces the same answer on application four hundred as on application four. That consistency is the entire reason it can be run at volume by whoever is available, and it evaporates the moment a judgment call is smuggled into it.

How long should the expensive pass actually take?

Five to ten minutes, and it should end with a decision about whether to spend an hour, not with a rating. In that time you can open one named claim and check it, read the single artifact the person points at, and write two sentences saying what you would ask them. Anything longer is an interview conducted without the candidate present, and the extra minutes buy confidence that nothing in the document earned.

A worked shape for a mid-level operations role. Minute one: confirm the dates and the employer names against whatever public record exists. Minutes two to four: pick the most specific claim on the page, something like a named system migrated or a named client handled, and see whether the rest of the document is consistent with it. Minutes five to seven: open the one link they gave and read it the way you would read a colleague's work. Minutes eight to ten: write the question you would ask if you had them on the phone. If you cannot write that question, the application did not clear the pass.

The method for the middle of that generalizes. Pick the one claim whose truth would change your decision, and put the minutes on it. Sampling the page evenly is what the flat pace already does. Where a claim is about tooling rather than employment, verifying an AI-use claim in ten minutes with no assessment budget is the same discipline applied to a harder subject.

The ceiling is worth saying out loud. Neither pass tells you who is good, because the document stopped carrying that. What the expensive pass chooses is who is worth spending real evidence on, which is a smaller and more defensible claim than picking winners.

Set the pace from your interview capacity, not the pile

The calendar sets the pace, so start there. Count the first-round conversations your team can genuinely hold in a fortnight, add a margin for the ones that do not convert, and that number is what the expensive pass is allowed to hand forward. Everything upstream is sized to produce it, and the seconds per application are whatever that arithmetic leaves.

1. Take one closed requisition. Count the applications that genuinely needed a second look, judged now with the outcome in front of you. 2. Compare it to capacity. If the count is far above the interviews you could hold, the cheap pass is too permissive. Far below, and it is cutting people the expensive pass should have seen. 3. Tune the cheap pass, not the clock. Adjust the yes-or-no criteria until the two numbers meet. 4. Record the split. Applications in, cleared the cheap pass, read expensively, interviewed. Four numbers per requisition, and after three requisitions you have a screen you can reason about. 5. Cap the shift. Reading application four hundred at the same standard as application four is a claim about attention nobody can support. Break the pile, and put the expensive pass at the front of a session while the attention is there.

Criteria built on keywords stop working as the pile widens, because the words are free and every applicant now has them. That is the failure described in what replaces a keyword filter that stopped working. Read back your yes-or-no criteria and name the record that settles each one. A criterion with no record behind it is a keyword.

The share of the pile that earns expensive minutes is usually uncomfortable. On a large pile it is a small multiple of the share that already gets an interview, which the same benchmark set puts between 3.6% and 4.7% ¹. That feels brutal only until you notice the interviews were already that few, and that the rest of the reading was going into decisions which never happened.

See a sample report >