AI in hiring: faster screening without the bias and legal risk

Screening applications is a volume problem, and volume problems are where teams reach for automation. A role that draws hundreds of CVs is exactly the sort of task an AI could triage in minutes, and the appeal is obvious to anyone who has read that pile by hand. The problem is that hiring is not just a volume problem. It is a decision about people, governed by law and haunted by bias, and an AI that ignores either of those turns a time-saver into a liability. Both outcomes are designed in, so it is worth designing for the right one.

The law already treats this as high risk

This is not a grey area waiting for regulation. The EU AI Act explicitly classifies AI used in recruitment and employment decisions as high-risk, which brings real obligations around transparency, human oversight, and record-keeping. On top of that, GDPR's Article 22 limits purely automated decisions that significantly affect people, and a hiring rejection qualifies. In other words, an AI that auto-rejects candidates on its own is not an efficiency, it is a compliance exposure. Any use of AI in hiring has to start from the assumption that regulators are already watching this specific use.

Bias is not a hypothetical here

The classic failure is well documented: a large employer built a CV-screening model, trained it on its own past hiring, and found it had learned to penalise applications associated with women, because that reflected the history it was trained on. The tool was scrapped. The lesson generalises. A model trained on past decisions learns past preferences, including the ones you would never write into a policy. Left unchecked, AI does not remove bias from hiring, it launders it, giving a discriminatory pattern the appearance of objective scoring.

Where AI helps without deciding

None of this means AI has no place in hiring. It means AI belongs on the work around the decision, not the decision itself. Parsing and structuring applications so nobody retypes them. Organising and surfacing candidates against explicit, job-related criteria. Handling the scheduling and the acknowledgements that otherwise eat a recruiter's week. Used this way, AI gives a hiring team its time back to spend on judging people, which is the part that should never have been automated. This is the same principle behind our process automation work: automate the handling, keep the judgement human.

How do you set this up defensibly?

The compliant version is not slower to build, it is just designed in a different order. Start from the criteria, not the tool: write down the job-related requirements a candidate must show, the way you would defend them to a tribunal, before any system scores anything. Then let the system organise applications against those written criteria, surfacing evidence rather than issuing verdicts: this candidate shows the certification, this one shows the required years in a comparable role, this one is missing both. A recruiter reads that organised evidence and makes the call, and the system records what was surfaced and who decided. Rejections are sent by a person, on criteria a person applied.

Two checks keep the setup honest over time. First, test outcomes, not intentions: periodically compare pass-through rates across groups you are legally required not to disadvantage, because bias shows up in the numbers long before anyone notices it in the process. Second, audit the criteria themselves once or twice a year, since a requirement that looked neutral can act as a proxy for something protected. An illustrative example: a mid-sized firm screening for "continuous employment history" is, without intending it, penalising anyone who took parental leave. The fix is not better AI, it is a better criterion, and only a process that keeps criteria explicit ever finds that.

What good, defensible use looks like

Keep a person accountable for every hire and reject, test for bias rather than assuming its absence, use explicit job-related criteria instead of a black box, and keep the records the law now expects. That is not just how you stay compliant, it is how you actually hire better, because it forces the criteria into the open. Treat AI in hiring as governance first and efficiency second, the way we frame it in AI governance for mid-market companies and what the EU AI Act means in practice. If you want to speed up screening without inheriting the bias or the legal risk, tell us how you hire today.