Every demo of Alert Hire reaches the same moment. The recruiter watches a CV get graded, watches the screening conversation run, and then asks: so what does it actually decide? It is the right question, and we would rather answer it in writing than have anyone guess. The line between what the AI decides and what a person decides is not a setting. It is the product.
What it decides
- The score. Each application gets a sub-score per criterion and an overall score, with a written justification that quotes the CV. The criteria are the recruiter's; the grader applies them.
- The order of the list. Candidates on a role are ranked by score. That is a reading order, not an outcome.
- Whether to send the screening questions. If the score meets the threshold the recruiter set, the candidate gets the questions that day. If not, they stay in the list, graded and visible, and the recruiter can start the conversation by hand.
- When an answer needs a follow-up. If a candidate's reply does not answer the question, it asks again, once, in different words. It does not invent new questions.
- When something is off-script. If the candidate raises anything outside the screening questions, the AI decides one thing: to stop and hand it to a person.
- How the summary is worded. Two paragraphs for the recruiter, plus the extracted answers. The recruiter can read the full transcript underneath.
What it never decides
- Whether a candidate is rejected. Nobody is declined by a score. A recruiter declines a candidate, or does not.
- Whether a candidate moves stage. Alert Hire can write a stage change to your ATS, but only if you switch that write-back on, per integration, and it is off by default. Even then it is your configuration making the change, and it is logged as such.
- What a candidate is told about an outcome. The AI never says "you were unsuccessful" or "you are through". It asks the questions, thanks the candidate, and tells them a recruiter will be in touch.
- Anything about money. If salary expectation is one of your questions it records the answer. It does not negotiate, hint, or comment.
- Anything about a disclosure. Health, family, disability, a complaint: acknowledged once, routed to a person, kept out of the score.
- Whether to keep going when a candidate asks for a person. It stops. Asking for a person never counts against the candidate.
Why the line is there
Four reasons, in the order we weigh them.
A score is only defensible if a person owns the decision. When a client, a candidate or a regulator asks why someone was not progressed, "the model said so" is not an answer. "A recruiter read the grade, read the reasoning, and decided" is. That answer is only available if the decision was actually the recruiter's.
The law is heading the same way. From 10 December 2026, Australian privacy policies have to describe automated decisions that significantly affect people, and distinguish decisions a program makes on its own from ones it assists. Keeping the AI on the assisting side of that line is simpler to describe and simpler to defend. More on that here.
Recruiters have to be able to overrule it. A tool you cannot disagree with is a tool you stop trusting the first time it is wrong. Every grade in Alert Hire can be pinned, overridden or regraded, and the override sits next to the original score in the history.
AI fails confidently. Models are wrong in ways that read as right. So we put the AI where being wrong is cheap and reversible (a score you can read the reasoning for) and keep people where being wrong is expensive and hard to undo (an outcome told to a human being).
What it looks like on a Tuesday
You open the role. Sixty-two applicants, ranked. Five are above the line and have already answered your questions; one of them asked about working four days and is sitting in the Needs you queue with what they said. You call the top three, reply to the four-day question yourself, and notice that number six scored low on the platform criterion because their CV calls it by an older product name. You pin them above the line, and fix the criterion so it does not happen again. Nothing was decided until you decided it.
The full list of what candidates are told is on the candidate data page. To see the line in practice, book a demo.
Sources
- Australian Human Rights Commission with LexisNexis, AI and recruitment compliance checklist. Checked 10 September 2026.
- Office of the Australian Information Commissioner, Australian Privacy Principles guidelines, Chapter 1: APP 1. Checked 10 September 2026.
- Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products. Checked 10 September 2026.
Every source here was opened and read on the date shown. If one has moved or changed, tell us at admin@alertai.com.au.