It's Monday morning, and Nurul's inbox is already full. She's the only HR executive at a 60-person logistics company. One colleague asks where to submit a leave request. Another wants the latest expense claim form. A manager wants to know who approves an exception to the travel policy.
None of these questions is hard. But each one means finding the right document, checking it's the current version and explaining the next step. By lunchtime, half her morning has gone.
AI can help Nurul draft those answers, as long as it works from the right documents and she checks what goes out. Here's how HR teams can set that up.
I. Start with simple "where" and "how" questions
Begin with routine process questions that a policy clearly answers: where a form is kept, what a request needs and which team handles it.
Keep personal matters out of the first pilot. That includes individual entitlements, disciplinary issues and hiring decisions. These need a person's judgement. AI can help organise the question, but it shouldn't decide the outcome.
Practise with made-up questions first. Employee data is protected under the PDPA, so agree which documents the AI tool may access before uploading anything. Removing names alone doesn't make a document safe to use.
II. Tidy the policy library first
AI will give a confident answer from an outdated policy just as easily as from a current one. So before writing any prompts, give each policy a title, owner, version number and effective date. Remove old copies from the folder the AI uses.
For any policy, Nurul should be able to answer three questions:
Which version is current?
Which employees does it cover?
Who handles questions it doesn't answer?
If two current policies contradict each other, fix that before going further.
III. A four-step answer workflow
1. Capture the question. Note what the person needs to do. Don't ask for personal details you don't need.
2. Find the right policy. Pull out the relevant section and its version. If there's no clear source, flag the gap.
3. Draft the answer. AI summarises the process, names the source and lists anything still to confirm.
4. Check and send. Nurul checks the source, the wording and the next step before she replies.
Here's a prompt for step 3:
"Use only the approved policy extracts below. Draft a short answer explaining the next step. Name the document, version and section. If the extracts don't answer the question, say what needs clarifying and refer it to HR. Don't guess anyone's entitlement or invent a rule."
Even with clear instructions, AI can sound sure and still be wrong. It may even quote a section that doesn't support its answer. That's why step 4 stays with a person.
IV. Test the tricky questions
Before rolling it out, try a small set of test questions: a simple one, an incomplete one, one based on an outdated policy and one the policy doesn't cover.
For example, someone asks: "Can my request be approved today?" If the policy only explains how to submit a request, a good answer explains the steps and names the approver. It shouldn't promise approval.
For each test, note whether the source was right, whether the AI added rules of its own and whether escalations reached the right person. Include checking time when you compare it with the old way.
V. Give the workflow an owner
Someone needs to keep the policy folder up to date, spot repeated errors and decide which questions the AI should stop handling. For Nurul, that's her and her HR manager.
Employees should always know how to reach HR directly, and how to flag an answer that looks wrong.
VI. Start with one type of question
Pick one routine category, gather the current policies and involve whoever owns them. That's enough for a useful first pilot. For a wider plan, see our guide to building an AI-ready workforce in 90 days.
If you'd like your HR team to build workflows like this with guidance, explore our AI for HR Teams programme or book a discovery call.
Nurul is an illustrative example based on a typical SME.