First Round Collects the Claim, Final Round Watches the Work

What can the first round actually establish?

Everything a conversation can settle, and nothing beyond it. Round one is for scope and fit: what the role is, what the person has actually done, what they want next, and whether the claimed experience matches the job. It should also produce one thing the rest of the loop uses, which is a specific project worth testing later, named by the candidate rather than picked off a resume.

That is a real job and it is worth doing well. Structured interviews estimate at .42 against .19 for unstructured ones in the 2022 re-analysis of the selection literature, so the distance between two interviews is larger than the distance between an interview and most other methods ¹. Same questions, same order, a rating scale everyone shares. A first round run that way is the cheapest reliable evidence a hiring process can buy.

What round one cannot do is settle whether somebody works well with a model, because every answer to that question is now preparable. A candidate can draft, refine and rehearse a story about catching a hallucination the night before, and the interviewer across the table has no way to separate the person who does it from the person who prepared to describe it. That is not a candidate integrity problem. It is what happens when a question has a knowable good answer and the answer is free to obtain.

So use round one to pick the target. Ask what they built or shipped in the last six months that they would be happy to walk through, get the specifics on paper, and stop. Whether the loop needs an AI-fluency round at all is the question this decision sits inside.

What should the final round do instead of asking again?

Produce evidence a conversation cannot. That means work: a slice of the real job, done with the candidate's own assistant, while somebody watches what they choose. The questions in that round are about decisions made in the last forty minutes rather than decisions made two jobs ago, which is the one class of answer nobody can prepare in advance and nobody has to remember accurately.

Watch where the follow-ups can point. Instead of asking "how do you verify AI output", the reviewer asks why the candidate accepted the third paragraph without checking it and rewrote the second one twice. Instead of asking "what would you refuse to delegate", the reviewer asks about the moment the assistant offered to do the whole thing and the candidate either took it or did not. Those questions have no rehearsed answer available, because the event happened in front of both people.

The evidence base supports the format without overselling it. Work samples estimate at .33 in the same 2022 re-analysis, which replaced a .54 figure that had circulated since a 1974 review and that its own lineage abandoned; 53 of the 54 underlying studies tested people already doing the job rather than applicants ¹. Read .33 against .42 carefully. The credibility intervals overlap heavily and the paper argues the pattern is coherent, so this is not a demotion of work samples, and none of that literature is about AI-open assessment, because none exists yet.

What the format buys is a different kind of record, which is why redesigning the interview so AI assistance becomes signal changes what the last round can be asked to decide.

Build the final round out of one task and a clock

One task, forty to sixty minutes, the tools the job uses, and a reviewer who watches without taking part. Pick material the role touches this quarter. Put one thing in it that is confidently wrong. Tell the candidate the assistant is allowed and that the decisions matter more than the polish. Then leave them to it and read what happened afterwards.

1. Material from the actual job. A generic puzzle measures puzzle skill. Give an analyst a real filing, a recruiter a real requisition, a revenue-cycle candidate a real denial. Judging an output requires knowing what a normal output looks like here, and generic material removes exactly that. 2. The assistant open. Banning it tests work without AI, which is a different job from the one being filled, and it converts the exercise into a compliance test. 3. A planted error. Something the field would catch and the model will happily build on. This is the part the whole round turns on, and it is the part most take-homes leave out. 4. A reviewer who stays quiet. The reviewer's job is to note what got delegated, what got checked, and where the candidate stopped. Prompting them defeats the measurement.

Run it live or asynchronously depending on what the role is like, and be honest that the two produce different evidence: take-home or live session is a real trade rather than a preference. For engineering roles specifically, running a coding interview with the assistant open covers the mechanics that differ from a written task.

What does this cost, and what does it replace?

Less than adding a round, because it replaces one. Median time-to-fill for nonexecutive roles sits at 44 days, and the median organisation spends 7 of those conducting interviews against 5 screening and 4 deciding ². Swapping a second conversation for a working session moves no days at all. What changes is what the last hour produces.

The number to keep in view is the spread rather than the median. The interquartile range runs 28 to 73 days and the mean is 54, so a claim that the average hire takes 44 days is wrong twice over. Those are 2021 figures from SHRM member organisations, collected before generative AI reached hiring at all, and they measure calendar days from requisition to offer acceptance.

The same benchmarking data explains why this reads as a bigger change than it is. Work-sample interviews were used by 12%, 11% and 9% of organisations to assess executive, middle management and individual contributor candidates, and simulation exercises by 6%, 6% and 7%, while in-person interviews ran at 79%, 78% and 76% ². Those are self-reports against SHRM's own definitions, so treat them as an upper bound on structure. A job-shaped final round is still a minority practice, which means most loops are replacing a conversation they already run instead of adding a stage.

There is a limit on how far this format travels. A working session is one piece of evidence rather than the whole decision, and stacking three of them repeats the mistake of stacking three conversations. Two or three genuinely different kinds of evidence is the shape that holds up. Whether that means three rounds or four is a separate question, worked through in how many interview rounds a loop actually needs, and the bar the final round is graded against should already be written down by level, as in the three bars at junior, mid and senior.

See how it works >