AI Tools for HR: Safe Uses, Limits, Costs

For HR, use AI to draft and to answer routine questions from approved documents, and keep it away from decisions about individual people. Good first jobs are job descriptions, handbook Q and A, onboarding checklists and survey summaries. Screening, ranking and discipline are regulated and high risk: a trained person decides, and legal checks the process. Checked 31 October 2026.

The job in one table

What the job is

HR handles people's pay, health, careers and complaints. It is repetitive in places: the same leave question, the same offer template, the same onboarding steps. It is also among the most sensitive work in a company, because a mistake affects a person's livelihood and carries legal weight.

That split sets the rule. AI can take the repetitive, low-stakes writing and answering. It should not take the judgement, and it should not touch private records without a clear data rule. Anthropic's guidance says agentic systems trade cost and latency for performance and warns of compounding errors in autonomous systems. In HR the cost of a compounding error is a person, so keep the autonomy low.

Which agent fits

If you pick an outside vendor for candidate screening, ask for documentation on how the tool was tested for bias and what you must disclose to candidates. Bias-audit and notice rules for automated hiring tools differ by country and region, so take them from your regulator's guidance or your counsel.

Add-ons that help

Give each connector the narrowest scope: a folder, not a whole drive; a channel, not a workspace; read-only first. Remote servers carry your data to a third party, so check each page's auth and data notes, and read what MCP servers are. The full list is at MCP servers.

Prompts and kit items

The prompt library holds writing prompts by task. The docs writer skill is a pattern for turning policy notes into clear help pages, and the CLAUDE.md starter is where you put the rules every HR draft must follow, such as inclusive wording and the phrases legal wants in every offer letter.

A prompt for a job description should include: the role's real duties, the must-have and nice-to-have skills, the pay range if you publish one, and the instruction "remove requirements that are not needed to do the job and avoid gendered or age-coded wording." Ask the model to list which requirements it removed, so a person can check.

A safe workflow: handbook answers

This is the best first project in HR, and it follows the same pattern as our support build guide.

1. Collect approved documents: leave policy, benefits, expenses, code of conduct. Date each one.
2. Write the rules: answer only from the documents; quote the section; say "ask HR" when unsure.
3. Exclude personal records. The assistant answers policy questions, not "how many days do I have left."
4. Retrieve first, then answer. Search the documents and send only the relevant passages to the model.
5. Log questions with no names attached, so HR sees what people ask.
6. Review a sample weekly, and fix the documents where answers were wrong or unclear.
7. Route sensitive topics to a person: harassment, health, grievances, termination, immigration.

What it costs

From Anthropic's pricing page, checked 31 October 2026: Haiku 5.5 is $0.10 per million input tokens and $0.50 output; Sonnet 5.5 is $2 and $10.

Worked example (token counts are our assumptions). A handbook answer with 2,500 input tokens (question plus retrieved passages) and 300 output tokens:

Even a large company's volume costs little in tokens. The larger costs are: building and securing the assistant, keeping documents current, legal review, an HR platform subscription if you add one and the time HR spends reviewing. See Claude API pricing.

Risks and how to review them

The licence of a model can also set limits. The usage policy attached to Meta's Muse Spark Open, for example, lists as prohibited uses engaging in or facilitating discrimination or other unlawful or harmful conduct in the provision of employment, and collecting or inferring private or sensitive information about individuals without the right to do so. Whatever model you use, read its terms for employment uses. See open source LLMs for the licence picture.

Job descriptions: a worked checklist

A job description is the most visible HR document and the easiest to improve with AI. Use this checklist for every draft.

Ask the model to produce the draft and a separate list of "things I removed or changed and why." A person reads both. The aim is a shorter, clearer post that attracts the right people, not a longer one stuffed with keywords.

Onboarding

New starters ask the same questions: where do I get my laptop, who approves expenses, what is the leave policy. An onboarding assistant can answer these from approved documents, and a checklist generator can adapt a template to the role and start date. Keep three limits. The assistant does not see personal records. It hands off to a named person for anything about pay, contracts, visas or health. And a person welcomes the new starter, because the first week is about belonging, which no tool provides. A simple measure of success is the number of "where do I find" messages HR receives in a new starter's first month, compared with the same figure before the assistant.

Survey comments

Free-text survey answers are valuable and slow to read. A model can group a thousand comments into themes and pull a few representative quotes. Anonymise first: remove names, teams small enough to identify and any detail that points at a person. Ask the model to report how many comments fall under each theme and to quote exactly, and sample the raw comments yourself to check the themes are fair. Never ask a model to guess who wrote a comment, and never feed it free text from a confidential reporting channel. For small teams, even a summary can identify a person; set a minimum group size below which you do not report.

Working with legal and employee representatives

Bring legal into the project before the first pilot, not after the first complaint. Ask three things: which laws apply to tools used in recruiting and to employee data where you operate, what notice staff and candidates must get, and what records you need to keep. Where employees have a works council or union, involve them early; a new tool that processes staff data usually needs consultation. Write down a plain-language statement for staff about what the tool does and does not do, and where to go with concerns. Consultation and notice duties differ by country, so confirm yours with official guidance or counsel.

Measure it

Do not measure an AI tool by how many candidates it rejects. That is a measure of risk.

Common mistakes

• Starting with screening. It is the riskiest use and the one regulators watch.
• Pasting records into a consumer chat tool. Use an approved tool.
• No owner for the documents. The assistant is only as right as the policy files.
• Hiding the tool from staff. Say what it does and where it stops.
• Letting AI write the judgement in a review. The manager's view should be the manager's.

Who should do what

Related pages: AI tools for business for choosing across jobs, AI customer support for the same grounded-answer pattern, AI agent governance for controls.

Our take

HR is a good place for AI on the writing and answering side and a poor place for it on the deciding side. Ground every answer in approved documents, keep personal records out of general tools and put a person on every decision about another person. The model cost is trivial. The legal and privacy work is the real project. We built this from vendor pricing, licence terms and published guidance, are not legal advisers and have marked every legal point for you to check with counsel.