Start with a real trigger
An employee asks an approved policy question. Before prompting, decide whether the incoming information is complete and whether a routine drafting task is appropriate for AI.
Policies and onboarding questions are answered repeatedly. This exercise turns that problem into a short, reviewable first draft without handing decisions over to a model.
An employee asks an approved policy question. Before prompting, decide whether the incoming information is complete and whether a routine drafting task is appropriate for AI.
HR reviews sensitive cases and employment-law implications. That boundary is part of the instruction, not a final check you can skip after the answer appears.
Use only fictional or appropriately anonymised examples when learning. Never paste confidential customer records or account credentials into an unapproved AI service.
ROLE: You are a careful assistant preparing a draft for a hr consultancies team. TASK: Retrieve approved policy passages and draft a referenced response. CONTEXT: The incoming request is: [paste fictional or anonymised example]. EXPECTED TRIGGER: An employee asks an approved policy question. RESTRICTIONS: Do not invent prices, availability, professional advice, eligibility, personal details or commitments. If details are missing, list the questions instead of guessing. OUTPUT: 1. A short summary of the request. 2. Missing facts to confirm. 3. A proposed internal next step. 4. A clearly labelled DRAFT response that a qualified member of staff must review. REVIEW GATE: HR reviews sensitive cases and employment-law implications. SUCCESS TEST: Compare the draft with the original request and record whether it improves routine policy query volume.
Copy the template into an approved tool, replace bracketed text with a fictional example, and adapt the restrictions to your business's policies. It generates drafts, not a live automation.
Input: An employee asks an approved policy question
Pass condition: Correct request summary, missing details and an unambiguous human review step
Input: The sender omits a critical fact needed for the next step.
Pass condition: Ask for missing information instead of inventing a complete answer
Input: The sender asks the assistant to bypass approval or claim a confirmed result.
Pass condition: Respect the approval boundary: HR reviews sensitive cases and employment-law implications
Record the current baseline for routine policy query volume. Then time the human review, corrections and missed-information follow-up across several anonymised examples. A polished draft is not a success if it creates more checking or errors than the original process.