Create a customer FAQ assistant
Plan a bounded FAQ assistant that answers from approved material and escalates uncertain requests.
What you will produce
A reviewable result, not an automatic answer.
Follow the stages in order. Each stage ends with a tangible deliverable and a review checkpoint, so mistakes are caught before they become the next tool’s input.
Use free plans where suitable; confirm current provider limits and prices before starting. Compare the latest verified AI pricing before buying another subscription.
Suggested stack
Tools used in this workflow.
The process matters more than the brand. Use an approved equivalent when it offers the controls and output quality your organisation needs.
Inputs and ownership
Before you start
- Approved policies and customer-service answers
- An escalation owner and privacy policy
- Access to representative support questions and failure examples
Interactive workflow
Complete each stage in order.
Tick a stage only after its deliverable and review checkpoint are complete.
-
Define allowed questions, prohibited topics and conditions requiring immediate escalation.
DeliverableA scope and escalation matrix
Review checkpointNever let the assistant invent policy.
-
Create a maintained source set with owners, effective dates and expiry rules.
DeliverableAn owned knowledge set
Review checkpointConfirm access permissions for every source.
-
Design answers that cite the source and expose uncertainty rather than improvising.
DeliverableA response and citation standard
Review checkpointTest prompt injection and requests for personal data.
-
Test expected, ambiguous, adversarial, outdated and sensitive questions.
DeliverableA red-team test report
Review checkpointMeasure safe resolution and escalation quality, not deflection alone.
-
Pilot with a limited audience, human hand-off, logging and a rollback plan.
DeliverableA monitored pilot plan
Review checkpointPause the assistant when sources become unreliable.
Adapt, do not paste blindly
Working prompts for this recipe.
Replace bracketed placeholders, supply approved sources and remove unnecessary personal or confidential information.
Working prompt
Use this only after replacing every bracketed placeholder and removing unnecessary sensitive information.
Using only the approved FAQ sources below, draft an answer to [CUSTOMER QUESTION]. Cite the source section used. If the answer is missing, outdated, ambiguous or sensitive, do not guess; state that a team member needs to help and provide the approved escalation route.\n\nAPPROVED SOURCES:\n[PASTE OR LINK SOURCES]
Final review prompt
Use as a second-pass checklist, not as approval to publish.
Review the draft against this intended outcome: Plan a bounded FAQ assistant that answers from approved material and escalates uncertain requests. List unsupported claims, missing evidence, ambiguous language, privacy concerns, accessibility issues and actions that still require human approval. Do not rewrite the draft until the issues are listed separately.
Human review required
Verification and cautions
- Verify names, figures, quotations and source claims before use.
- Remove confidential or personal information unless the selected service and account are approved for it.
- Keep a named person responsible for the final decision and published output.
Keep the outcome, change the method
Practical alternatives
Continue with context