What a High-Agency Recruiter Actually Does All Day

There is a question the AI summit circuit keeps dodging: what does human agency actually look like in practice? The phrase gets a great deal of use. It is rarely cashed out into anything a person could be observed doing — a decision they make, the evidence they make it on, the reasoning they could defend afterwards if somebody asked them to.

We decided to answer it in the one place where we can answer it with some authority. Recruitment. What does it take to be a high-agency human recruiter? Not in principle, and not as a value statement. In practice, at the level of a specific judgment a specific person has to make and then stand behind.

Rather than write another document about it, we built a simulator. The Draup Recruiter Simulator is a hands-on, AI-powered platform covering two areas: Compliance and Legal Readiness, and Skill Building, the second of which runs as four live simulation modules. It is in beta, the features will keep evolving, and we would rather hear what breaks than hear that it is interesting.

Agency is what survives a confident machine

Part of the reason the phrase stays undefined is that the two easiest definitions are both wrong, and they are wrong in opposite directions.

The first is refusal. On this reading, the high-agency recruiter is the one who ignores what the system produced and works from instinct. That is not judgment. It is declining to engage with the evidence, and it produces worse decisions than the tool would have, with the added disadvantage that nobody can reconstruct how they were reached.

The second is the human-in-the-loop formulation, where a person sits in the workflow with the authority to approve or reject, and in practice approves. Nothing in that arrangement requires the person to have understood the output. Presence is not agency. A signature is not a judgment. If the only thing the human step reliably produces is a timestamp, the loop is decorative.

The useful definition sits between them and is considerably harder to satisfy. A high-agency recruiter can look at an output and say what it is actually claiming, where that claim could be wrong, what would have to be true for it to be right, and what they intend to do instead. Four separate abilities, none of them innate, all of them trainable — and all of them invisible until someone has to perform them under time pressure with a hiring manager waiting on an answer.

Which is why the vocabulary matters far less than the exercise. You cannot read your way into the ability to challenge a ranked list. You acquire it by challenging ranked lists, being wrong some of the time, and finding out why.

Why a simulator rather than a paper

Our instinct with a question like this one is to write a paper about it. This subject resisted the format, and the reason is worth naming, because it generalises beyond recruitment.

A document can tell a recruiter that an automated screening tool may carry adverse impact, or that a candidate's answers may have been generated rather than lived. What it cannot do is put them in the position of noticing it. Recognition is a different faculty from comprehension. Knowing that a category of failure exists does very little for you at the moment the failure is in front of you wearing ordinary clothes and looking exactly like a normal shortlist.

That gap matters most for the recruiter in the middle of a normal week. Nobody schedules time to consider whether a matching model is systematically undervaluing a particular set of candidates. That call gets made in the gaps between everything else, and it gets made well only if the person has already met the failure mode somewhere safe. Simulation is the cheapest place to meet it — cheap in money, but more importantly cheap in consequence, because the candidate who is wrongly filtered out inside a simulation is not a real person who never hears back.

Compliance as part of the craft, not a separate department

The first area the simulator covers is Compliance and Legal Readiness: NYC Local Law 144, GDPR, and AI bias audits. I am aware this reads like a topic to be handed to legal and forgotten about, and I want to argue against that reflex.

The reason is structural rather than legal. Whatever the obligations turn out to be in a given jurisdiction, the person who configured the tool, ran the search, and either moved a candidate forward or did not is the person who knows what actually happened. The organisation can hold the policy centrally. It cannot hold the specifics of an individual decision centrally, because those specifics only exist in the hands of whoever made it. That asymmetry does not go away by writing a better policy.

So the recruiter needs working answers to questions that used to belong to somebody else. When an automated tool contributes to a screening or ranking decision, has it been examined for bias, when, and what did that examination find? What is the candidate told about it, and at what point? What happens when a candidate asks for a human review, or asks what data was used to assess them? A recruiter who cannot answer is not usually exposed because they did something wrong. They are exposed because the answer exists somewhere in the organisation and they are the one standing in front of the question.

Treating this as craft rather than paperwork also changes the tone of the conversation with candidates. Explaining plainly how a decision was reached is a trust-building act. Being unable to explain it is the opposite.

The four capabilities the modules train

The Skill Building area runs as four live simulation modules. Each one sits at a point where the burden of judgment has quietly moved onto the person operating the tool.

AI interpretation. Matches, scores, rankings and summaries all look like conclusions and are in fact estimates. Interpreting one means knowing what it is a proxy for, where that proxy is weakest, and which candidates a given method will predictably under-rank. The hazard is not that models are wrong. It is that they are wrong in a fluent, well-formatted, confident-sounding way, and fluency reads as reliability to a tired person.

Follow-up writing. Generated outreach is grammatical, fast and frictionless, which is precisely why it is so easy to send something that says nothing a specific candidate needed to hear. The skill is working out what a particular person, at a particular stage, is actually deciding — and writing to that decision rather than around it.

Skills assessment. Judging demonstrated capability rather than filtering for keywords asks something different of a recruiter. It means separating a claimed skill from a shown one, and knowing what a given assessment method can and cannot establish. Every method has a blind spot; using one without knowing its blind spot is how confident mistakes get made.

Candidate fraud detection. Where a plausible-looking application is cheap to produce, the informational value of polish falls. This is not suspicion of candidates, and framing it that way misreads the problem. It is the ability to notice inconsistency in a process that has become easier to game.

Where to begin

If you lead a talent acquisition function, the practical first step is not another AI policy. It is to find out concretely how your recruiters handle these four situations today. Give them a ranked list built on assumptions you know to be shaky and see whether anyone challenges it. Ask what the team would say to a candidate who wanted to know how they were screened. Whatever gap you find is your training agenda, and it will be more specific than any competency framework would have handed you.

Then let people practise somewhere a wrong call costs nothing. That is what the simulator is for. It is early, deliberately narrow, and built to be argued with — and the arguments from recruiters who use it in anger are what will make the next version better than this one. Tell us what you think.