AI Prompts for HR Teams: Where It Helps and Where It Must Not Go

HR is unusual among business functions: most of its work involves decisions about identifiable people, which puts a large share of it straight into the highest-obligation category under the EU AI Act. Recruitment and employment decisions are named in Annex III as high-risk uses.

That does not mean AI has no place in HR. It means the line between helpful and hazardous is sharper here than anywhere else, and it is worth drawing it explicitly before handing the team a prompt pack.

The bright lines first

These should not be delegated to a model, regardless of how the prompt is written:

• Screening, ranking or scoring candidates. This is a high-risk use with specific obligations, and a prompt in a general chatbot is not a compliant way to do it.
• Performance ratings or promotion decisions. Same category, same reasoning.
• Disciplinary and grievance outcomes. Procedurally sensitive and individually consequential.
• Anything involving real employee data in an unapproved tool. Special-category data such as health, disability or union membership raises the stakes further.

Our guide to AI in recruitment under the AI Act covers the obligations in detail.

Where it genuinely helps

Job descriptions and adverts

Rewrite this job advert for clarity. Remove requirements that are not genuinely essential, flag any wording likely to deter candidates from under-represented groups, and tell me which requirements appear to be proxies for something else. Do not add requirements.

The proxy clause is the valuable one. It surfaces the degree requirement that is standing in for something nobody has articulated.

Policy drafting

Draft a policy on this topic for an organisation of this size in this jurisdiction. Use plain English at a reading level appropriate for all staff. Flag every point that would need a legal review, and every point where I have not given you enough information to be specific.

Policies are where models do their best HR work, because the output is checkable and the structure is conventional.

Preparing a difficult conversation

I need to have a conversation about this issue, described generically with no names. Help me plan it: what to open with, three responses the person may give and how to handle each, and what I should avoid saying. Tell me where I risk straying into something that needs a formal process.

Describe the shape, never the person.

Making sense of free-text survey responses

Here are anonymised survey comments. Identify recurring themes, quote a representative example for each, and tell me which themes are supported by many responses and which by only one or two. Do not infer causes.

The last two clauses matter. Models reliably over-read a single vivid comment, and reliably supply a cause for a pattern that may not have one.

Explaining an entitlement

Explain this entitlement in plain English for an employee audience. Keep every condition intact. State separately what is not covered. Flag anything the source document leaves ambiguous rather than resolving it.

The data habit that makes all of this safer

Work with the shape, not the person. Replace names, dates, identifiers and anything distinctive enough to identify someone in a small team. In HR, re-identification is easy: the employee who returned from long-term sick leave in March identifies exactly one person in most organisations.

Verify before you send

Everything above produces a draft, not an answer. Entitlements, notice periods, statutory minimums and procedural steps must be checked against the current source, because a confidently wrong figure in an HR letter is both an employee relations problem and a legal one. Our verification method covers how to scale the checking to the stakes.

What to tell employees

If you use AI in processes that affect staff, say so. Transparency is both a legal expectation in several contexts and the thing that stops a reasonable tool becoming a trust problem when it is discovered rather than disclosed. That applies to notetakers in HR meetings too, where the transcript of a grievance conversation is a document you may later have to produce. See our guide to AI meeting notes at work.

Frequently asked questions

Can we use AI to sift CVs?

Employment screening is a high-risk use with specific obligations. It is not something to improvise with a general tool, and it needs proper assessment first.

Can we use it to write a dismissal letter?

For structure and tone, perhaps. The reasoning and the decision must be yours, and the letter must reflect the process actually followed.

Do we need to tell candidates?

Where AI materially affects a decision about them, transparency obligations are likely to apply. Check the position for your jurisdiction and process.

Does HR need its own AI training?

Yes. HR sits at level 3 in most organisations because its decisions affect individuals. See our guide to AI fluency levels by role.

Our CPD courses cover AI for HR and people teams, including the Annex III obligations that attach to employment decisions.