AI for HR teams: what to hand over, and what stays with a person

A marketing draft that lands badly costs you a campaign. A hiring process that quietly excludes people costs them a job. That asymmetry is why AI for HR teams needs handling differently from AI anywhere else. The question is which parts of the work you can hand over without handing over the judgement that makes the outcome defensible.

The drafting, which is most of the week

Job descriptions are the obvious start, and the gain is not the prose. Give it the last three you wrote, the real responsibilities and the salary band, and ask it to strip out every phrase that describes a personality, not a job. Then ask what a strong candidate might find off-putting. Half of what comes back you already suspected. The other half is the reason three good candidates never applied.

I watched a head of people read one of those lists aloud last November, in a borrowed meeting room, off a laptop propped on a stack of spare chairs. She reached the fourth line, stopped and said: "That one is mine." It was in all three of the job descriptions she had brought with her.

Interview scorecards repay the effort more. Ask for the evidence an interviewer would need to score each criterion at each level, and a woolly conversation about fit turns into something two people can compare afterwards. Policy drafting works the same way: a structured first version of a family leave or hybrid working policy, which your employment adviser then reads line by line.

Then there is the message nobody wants to write. A rejection after a final-stage interview, a change to someone's reporting line, an outcome nobody hoped for. Write your own blunt version and ask for a rewrite that keeps every fact and drops the defensiveness. The instruction that works is warmer, not vaguer, because vagueness in that message is exactly what produces the follow-up meeting. It is the request I get most often in a session, always at the end, from someone who has already typed the thing three times.

Bias is not a closing caveat

These systems learned their patterns from existing writing, and existing writing about hiring is not neutral. Ask for a profile of a strong candidate and you may get one that assumes an unbroken career history, a particular register of confidence, and a settled idea of what a good communicator sounds like.

None of it arrives labelled. It arrives as fluent, reasonable prose, which is the difficulty. Two habits help. Write your criteria first, in your own words, and use the model to test them, not to invent them. Then hand it your shortlist rationale and ask it to argue that the rationale disadvantages a particular group. Sometimes it will be wrong. Sometimes it will be right in a way that is inconvenient at four o'clock on a Thursday.

Neither habit makes a process fair. A person still has to look at who is actually getting through.

The decisions that stay with you

Who is hired, how someone is rated at year end, who is placed at risk in a restructure. Those are not tasks to delegate, and no amount of prompt skill changes it. Anthropic's own guidance asks users to validate critical information before high-stakes decisions, and few decisions are higher stakes for the person on the receiving end. Read the cases where the answer is not to use one at all before you write your function's policy.

The conversation you are already behind on

Claude Cowork, Anthropic's agent for general office work, began as a desktop app in January 2026 and reached phones and the browser six months later. Tools moving that fast reach a function long before its rules do.

Your colleagues in legal and risk are working through a stricter version of this question. The comms and marketing team are working through a looser one. Those conversations go better in one room.

Write your own rules down before the first time you need them. A rule invented during a grievance is not a rule, it is a defence.