Thinking About AI for Hiring? Ask These Questions Before It Screens Anyone

A recruiter asks a tool to summarise applications against a job description. One candidate's career break is described as a concern even though the job requirements do not make continuous employment necessary. Another candidate uses an alternative document format the tool cannot read, so their experience is omitted from the summary.

The risk is not limited to a system that automatically rejects people. A summary, ranking or colour-coded recommendation can influence a human reviewer. UK government guidance on responsible AI in recruitment highlights bias, accessibility and the need to test tools in the context where they are used.

Request a plain account of the inputs, intended purpose, output and known limitations. Ask how the tool handles accessible formats, career breaks, different names and incomplete applications. Find out whether the vendor uses candidate data to improve models, which parties process it and how corrections and deletion work.

Demand evidence relevant to the actual role and applicant population, not just a general accuracy score. If the tool ranks candidates, understand which factors affect the ranking and whether recruiters can see the original evidence. A black-box score should not quietly become the real selection criterion.

A reviewer needs time, authority and information to disagree with a recommendation. Provide the original application beside any summary, flag uncertain extraction and record the reasons for material selection decisions. Offer an accessible route for candidates to raise errors or request an alternative process where appropriate.

The ICO's work on recruitment tools addresses personal-data handling by providers. Employers should also assess their own responsibilities and the current rules for automated decision-making; the right answer depends on the specific processing and decision.

A first pilot could organise interview scheduling or produce a draft summary for a recruiter without scoring suitability. Test against a representative set that includes varied career paths, accessible formats and applications that should not be ranked.

Track omitted qualifications, correction requests, reviewer disagreement and differences in outcomes where lawful and appropriate. Stop if the team cannot explain or correct material errors. Hiring decisions deserve more than a quick check of a model's answer.