Move the Decision to the Stage That Still Produces Evidence

Which stage is actually deciding your hires right now?

Look for the stage after which almost nobody is turned down. In Ashby's benchmark set, recruiter screens pass roughly 35% of the candidates who reach them, and post-onsite and offer stages pass 95% and 81% ¹. A stage that passes nearly everyone is ratifying a decision rather than making one. The decision happened earlier, usually at a document review nobody counts as a stage.

Two cautions before you build on those numbers. They are passthrough rates among candidates who reached each stage, so they do not multiply into anything end to end, and the heaviest cut of all happens at application review, which is not in the series at all ¹. The sample is one vendor's own customer base, weighted toward venture-backed technology companies, so the levels will not transfer to your funnel.

The shape does. A loop that filters hard at the top on a document and then converts at 95% after the onsite has put its decision in the place where it collects the least evidence, and has staffed several hours of interviewer time to confirm what was settled upstream. The onsite feels decisive because it is where people argue. By then the argument is about a pool that was assembled somewhere else. That shape is measurable one round at a time by comparing the read going in with the decision coming out, which is whether the screen overturns the resume read applied to the first round of the loop.

So the useful question is not how many rounds to run. Which single stage would you defend to a candidate who asked why they were cut, and did that stage give them anything to do? If the answer names a stage where a person produced work you specified, the loop is sound and probably too long. If it names a document, the loop has no evidence stage at all, and dropping the resume screen in favour of a work sample becomes a live option.

What the resume screen can still be trusted to do

Sorting on facts that are true or false: dates, licences, location, work authorization, a named employer somebody could call. That is a real job and it removes real volume. What the screen cannot do any more is tell you who is good, because the parts of a document that used to carry that signal are now the cheapest parts of it to produce.

The evidence there was thin long before generated prose arrived. A meta-analysis of 81 independent samples put prehire work experience, which is most of what a resume screen reads, at a corrected correlation of .06 with later job performance, and experience with relevant tasks, jobs or occupations did no better at about .07 ². Those are corrected figures, so the near-zero result is not an artefact of under-correction. Read the scope carefully: this is experience measured at hire, which is exactly what a screen sees, and it says nothing about licensure or a legally required minimum qualification, which stay in the screen for reasons that have nothing to do with prediction.

What changed recently is the rest of the page. The summary, the achievement bullets and the register were never strong evidence, and they are now free to produce, so a screen that weights them is partly sorting applicants by which drafting tool they opened. Keep the checkable spine and stop grading the prose: the sequence is worked through in reading an application in order of what you can check.

Narrowing the screen's remit also gives you something to test. A screen that only claims to establish scope can be audited against its own claim, which is the practical route into knowing whether your screen is throwing away the wrong people. A screen that claims to identify quality cannot be audited at all, because nobody wrote down what quality meant before the reading started.

Put a price on the evidence stage before you place it

The cost per candidate decides where an evidence stage can sit, so work it out before you draw the funnel. An hour of a hiring manager's time against a twelve-hundred-application pool is not a plan. SHRM's benchmarking of its own member organizations puts median cost per hire at $1,244 for nonexecutive roles against a mean of $4,683 on a badly skewed distribution ³, and that definition excludes interviewer and manager time entirely.

That skew is the whole problem in one line. A process at the 25th percentile spends $354 a hire and one at the 75th spends $4,375 ³, so an evidence stage costing forty dollars a candidate is trivial in the second world and impossible in the first at any real volume. Work out your own number first: candidates entering the stage, cost per candidate, and how much of that cost is somebody's hour and how much is a vendor's invoice.

Then place the stage where the arithmetic survives. Three shapes work in practice:

• After a cheap checkable gate, on the fifty to eighty people who cleared it. The common case, and the one that keeps per-candidate cost bounded.
• On everyone, when the pool is small or the role is genuinely unreadable from paper. Expensive per head, cheap in total.
• Before any screen at all, which only works when per-candidate cost is close to zero. Price it before dismissing it: assessing everyone up front instead of screening is arithmetic, not ideology.

Building a bigger screen runs into a ceiling. Modelling a screening stage made only of methods that scale to volume, with the structured interview left out, Berry and colleagues found the validity-maximising composite topped out at .51 against .61 when the structured interview was in the mix, and pushing that screen-only battery past a four-fifths adverse impact ratio dropped it to .39 ⁴. Those are modelled ceilings, optimally weighted in a way no real screening stage is. They are not numbers to expect from a live process, and resume screening is not in the model at all. The comparative point is the usable one: the methods that scale are collectively weaker than the one that does not, so the top of a funnel cannot be made to carry a decision by adding more of itself.

How do you rebuild the loop around one evidence stage?

Take out one interview round for every evidence stage you add, and say out loud which stage now carries the hire. Adding a stage without removing one is how loops reach seven hours and lose the candidates who had options. The rounds you are cutting were mostly there to compensate for a screen that stopped discriminating, so cutting them is not the sacrifice it sounds like.

1. Write down the stage. One sentence: the hire is decided at X. If two people on the team write different sentences, that disagreement is the finding, and it explains most of your debrief arguments. 2. Audit the artifact. For stage X, name what the candidate produced and which conditions you set. No artifact means no evidence stage, whatever the stage is called on the pipeline board. 3. Count the cost. Candidates entering X, multiplied by cost per candidate, with interviewer hours priced at their real rate rather than at zero. 4. Cut a round. Pick the interview round that has never changed an outcome. Most loops have one, and the round worth cutting is often the second conversation about fit. 5. Size the gate above it. Set the cheap top-of-funnel pass so the number reaching X matches what step three says you can afford.

Cannot afford any evidence stage at step four? Keep one round and make it the evidence. Structured interviews came out top ranked at .42 in the corrected 2022 estimates, with job knowledge tests at .40 and work sample tests at .33 ⁵. Those are pooled correlations with supervisor ratings, not accuracy rates, and the credibility bands around them overlap heavily, so the order is no argument that a work sample is the weaker stage. Take .42 as a reason to structure the round you already run, and nothing more.

One failure mode deserves naming before you start. A team that moves the weight onto an assessment it cannot afford to run at volume will quietly start using the resume screen to decide who gets assessed, which puts the decision back where it began with an extra stage bolted on. The order matters: price the evidence stage, then size the gate in front of it, then write the criteria that gate uses.

See a sample report >