AI in recruitment: what to use, and what to leave alone
Recruitment is where AI tools have moved fastest, and where the risk of getting it wrong is best documented. Here is the practical line between the two.
The safe ground: checking a CV against a fact
Does this candidate hold the qualification the role requires. Do they have the minimum years of experience. Do they hold the professional certification the role specifies. Every one of those is a fact with one correct answer, which makes it a rules problem, and rules problems are exactly what automation is good at. Done well, this is arguably the single most automation-ready task in the whole hiring process, and it is worth treating it that way rather than dressing it up as something cleverer.
The risk ground: scoring or ranking overall fit
The trouble starts one step further on, where a tool claims to score or rank a candidate's overall suitability rather than checking a fact against a criterion. That is a judgement wearing the clothes of a measurement. The best-known cautionary tale remains Amazon's internal recruiting tool, scrapped after it was found to be downgrading CVs that mentioned all-women colleges or included the word "women's", having learned the pattern from ten years of male-dominated hiring data (reported by Reuters, October 2018). It is an old example and not a UK one, but the mechanism it exposed, a model quietly learning and repeating historical bias, has not gone away just because the tools have improved.
Checking a fact and guessing a fit are not the same task, and only one of them should be automated with confidence.
What good practice actually looks like
Government guidance on responsible AI in recruitment sets out three practical steps: tell applicants an AI tool is being used and where it has limitations, audit any screening tool for bias at the point you buy it and then again roughly every six months, and keep a manual review route available, particularly because screening tools carry a documented risk of disadvantaging disabled applicants. The ICO's own findings from engaging with employers already using automated recruitment tools add a further two: make sure human involvement in the process is genuine rather than a formality, and update your recruitment privacy notices to actually explain how the tool works and how a candidate can challenge it.
A worked example
A twenty-person retailer uses an AI tool to screen graduate applications. Applied well, the tool checks each application against fixed criteria, degree classification, availability, location, and flags anyone who does not meet them for a person to look at before rejecting. Applied badly, the same tool is asked to rank every applicant by "potential", using patterns from the firm's own past hires, which quietly bakes in whatever bias sat in who got hired before.
Where the edges actually sit
Not every scoring feature is equally risky, and it is worth resisting the urge to treat all of them the same. A tool that ranks candidates purely by a fixed, published set of criteria the business chose itself, years of experience weighted a certain way, say, is closer to automation than it looks, provided the criteria and their weighting are open to scrutiny. The genuine risk sits with tools whose scoring logic the business itself cannot fully see or explain, because at that point nobody, including the employer, can say with confidence what the tool actually rewarded.
The mistake that keeps recurring
The most common mistake is not malicious, it is complacent: a bias audit run once at the point of purchase and never repeated. Recruitment data drifts, application volumes change, and a tool that looked clean a year ago is not guaranteed to be clean now. The six-monthly re-check is not bureaucracy for its own sake, it is the only way anyone would actually notice if that changed.
A close second is assuming a vendor's own marketing answers the question for you. "Bias-tested" on a product page is not the same as being able to say, in your own words, what was tested, when, and against which protected characteristics. If a supplier cannot produce that evidence on request, that absence is itself an answer.
Where Jamie HR sits in this
Jamie HR does not automatically reject candidates, we always have a human in the loop. What it does is keep the record straight around a hiring decision, who reviewed what, and when, so if a candidate ever asks how a decision was reached, you have an answer already, rather than having to reconstruct one.