Human-in-the-loop hiring: AI that assists, never decides

Can OpusHire automatically reject a candidate?

No. There is no configuration, permission level or administrator setting that makes OpusHire reject a candidate without a person deciding. Screening and assessment minimums raise advisory flags for human review; they are not gates. Rejection is always an action a named user takes, recorded as such.

This is worth stating flatly because "human-in-the-loop" is usually a description of how a customer is expected to use a product, not a property of the product. Those are very different things. A platform with an auto-reject toggle is one busy quarter away from having it switched on, and the human who was nominally in the loop never sees the candidates the filter removed.

What does "computed in code, not by a model" mean?

It means the number that influences a decision is produced by rules you can read, not generated by a language model. An AI model may word an interview question, draft a narrative summary of a candidate, or write a readable paragraph about a reference — but it never picks an outcome and never produces the score attached to one.

The distinction matters because language models are not calibrated instruments. Ask one to score a candidate out of ten and it will return a plausible number that is not reproducible, not explainable, and not defensible when a rejected candidate asks why. Rules-based scoring is reproducible by construction: the same inputs give the same result, and every contribution to the total can be pointed at.

It also means rankings decompose. A recruiter can see which requirements a candidate met, which they missed, and how much each counted — and can change the weighting and watch the pool re-rank. A ranking you can argue with is a ranking worth having.

Why is there no timer that ages candidates into rejection?

Because a timer is an automated decision wearing a disguise. A candidate who sits untouched for thirty days and is then moved to a terminal state has been rejected by software, not by a person — the fact that the mechanism is a clock rather than a model changes nothing about who actually decided.

Plenty of systems do this, usually framed as pipeline hygiene, and the appeal is obvious: stale candidates clutter a board. But the candidate experiences it as a rejection nobody made, and the organisation loses the ability to say who decided and why.

OpusHire surfaces stalled candidates so a person notices them. It does not resolve them.

How do you know which parts were AI-assisted?

Because they say so. Every AI contribution to a score narrative or written summary is labelled as AI-assisted and human-reviewable wherever it appears, in the recruiter view and the candidate view alike. There is no surface where a model’s output is presented as though a person wrote it.

Labelling is doing two jobs. It tells a reviewer how much weight to give what they are reading — an AI-drafted summary deserves a different kind of attention than an interviewer's note. And it means a candidate exercising their rights over automated processing can be told accurately what was automated, rather than being given an answer assembled after the fact.

Why build the constraint into the architecture instead of the policy?

Because policies are edited by whoever is under pressure that quarter, and architecture is not. A commitment that survives only while everyone remembers to honour it is not a commitment a buyer, a regulator or a candidate should rely on — and it is not one worth advertising.

The practical consequence is that these properties are testable rather than promised. There is no auto-reject setting to audit for, because the code path does not exist. There is no model-generated score to calibrate, because scores come from rules. That is a narrower product than some competitors offer, deliberately.

It also travels well. Rules restricting decisions made about individuals by solely automated means are tightening in multiple jurisdictions, and the honest position under all of them is the same one: the system ranks, measures and evidences, and a person decides.

What does this cost you in practice?

Time, in one specific place: somebody has to look at the flagged candidates. OpusHire will not clear a pipeline for you overnight, and if your goal is to reduce a thousand applicants to ten without anyone reading anything, this is the wrong platform and we would rather say so now.

What you get for that time is a hiring record you can defend. Every ranking decomposes into requirements, every interview has a recording and a transcript, every assessment score is standardised and comparable, and every privileged action sits in a tamper-evident audit trail. When someone asks why a particular candidate did not progress — a hiring manager, a regulator, or the candidate — there is an answer, and it is not "the system scored them low."

Frequently asked questions