AI in HR: The Line Between Helpful and Hazardous

The distinction that keeps you out of trouble

AI applied to HR administration is straightforwardly useful. AI applied to decisions about individuals — who gets an interview, who gets promoted, who is at risk — carries legal and ethical exposure that is disproportionate to the efficiency gained.

This is not a technical judgement. It is that decisions affecting people's livelihoods need to be explainable and defensible, and probabilistic scoring is neither.

Safe and genuinely useful

• Answering policy questions from your own handbook, with citations
• Interview scheduling across multiple diaries
• Document handling — contracts, right-to-work, onboarding paperwork
• Drafting job descriptions, offer letters and internal communications for human review
• Summarising feedback and survey responses into themes
• Reminders and compliance tracking for training and certification

None of these decides anything about a person. All of them remove administrative load from a function that is usually understaffed.

Where the risk concentrates

Automated candidate scoring, ranking or filtering. Attrition prediction that flags individuals. Performance scoring. Anything that produces a number attached to a person and then influences a decision about them.

The core problem is that models learn from historical decisions, and historical hiring decisions encode historical bias. A model trained on who you hired before will efficiently reproduce whatever pattern that was, including the parts you would not defend.

Add to that the difficulty of explaining a model's output to a rejected candidate, and the regulatory direction of travel on automated decision-making, and the risk-reward is poor.

If you use AI in recruitment at all

1. Use it to surface, never to exclude. Ranking a longlist is different from rejecting people automatically.
2. Keep a human decision at every stage that affects a candidate's progress, and make that real rather than a rubber stamp.
3. Test for disparate impact across protected characteristics, before and periodically after deployment.
4. Document how it works in terms you could explain to a candidate or a tribunal.
5. Tell candidates what part of the process is automated.

The internal knowledge use case is the easy win

The highest-value, lowest-risk HR AI project is nearly always a question-answering system over your own policies. “How much parental leave am I entitled to?” asked at 9pm, answered accurately with a link to the policy.

It removes a large volume of repetitive queries, it improves consistency of answers, and it makes no decisions about anybody.

Data protection specifics

HR data is sensitive by nature and often includes special-category data. Access control, retention limits and a clear position on what leaves your infrastructure are prerequisites rather than refinements.

Be especially careful with anything that indexes employee files into a searchable system — the permission model there needs to be exactly right before it holds a single document.

Frequently asked questions