AI for HR

AI for HR teams

The part of HR that AI handles well is the part nobody demos, and the part that gets demoed is the part regulators have already started restricting.

Almost every HR AI pitch opens with candidate screening. That is the highest legal risk application in the function and, for most organisations, not the largest cost. Under the EU AI Act, systems used for recruitment, selection, promotion and termination decisions fall into the high risk category, which brings documentation, human oversight and transparency obligations that a lightweight tool will not satisfy. Separately, GDPR Article 22 restricts decisions based solely on automated processing where they significantly affect a person. Neither of these bans AI in hiring. Both make an unexamined screening deployment an expensive way to learn.

Where the return actually sits

• Policy retrieval. Employees and line managers ask the same questions about leave, sickness, expenses, parental entitlement and notice periods every week, and the answers already exist in a handbook nobody reads. Grounded retrieval with a citation to the clause removes a real queue and produces an auditable answer.
• Onboarding orchestration. Accounts, equipment, mandatory training, probation dates and right to work checks are a coordination problem across four systems. Automating the sequence and chasing the exceptions is unglamorous and saves more hours than screening does.
• Case file preparation. Pulling the timeline, correspondence and prior warnings for a grievance or disciplinary case, with a human writing every judgement.
• Job description drafting from an approved framework, with a human owning the final text.

What must stay with a human

Selection, rejection, promotion, performance ratings, disciplinary outcomes and dismissal. Not because a model cannot produce an output, but because the person affected has a right to a human review and an explanation, and because the organisation has to be able to defend the decision in a tribunal using something more than a score. If a tool ranks candidates, treat the ranking as one input a recruiter reads, log who reviewed it, and keep the rejected pool auditable.

Bias does not disappear because the model is new

A model trained on your historical hiring reproduces your historical hiring. The practical controls are narrow: define the protected characteristics you will monitor before launch, measure selection rates across groups on real traffic rather than a test set, and agree in advance what result triggers a rollback. A supplier who cannot tell you what their system was evaluated on is telling you something.

Starting point

Take the twenty questions your HR inbox answers most often, and build retrieval over the handbook that answers them with a clause citation. It is low risk, it is measurable in the first month, and it builds the evidence trail the harder use cases will require. An AI audit maps that first, then the higher risk work with the compliance obligations attached.

Frequently asked questions

Related