Updated 2026-11-18 · Research-based guide · Includes a practical decision worksheet
An AI support assistant should make it easier for a customer to reach a correct answer. Begin with a narrow set of routine questions, a reliable knowledge base, and a clear route to a person. Do not start by connecting every customer record or allowing the system to make commitments on the business’s behalf.
Separate answers from actions
Explaining your published opening hours is different from approving a refund, changing an order, or updating account details. List the actions the assistant may take and those requiring human approval. A reply can be drafted automatically while a refund decision stays with the authorized employee. Permission boundaries should be explicit in the workflow.
Prepare a small verified knowledge base
Start with current service descriptions, delivery rules, contact routes, and approved policy language. Give every document an owner and a review date. Remove contradictory versions. If an answer depends on an individual order or contract, route it to a person or an appropriately authenticated system. A general language model’s memory is not your policy source.
Create an escalation policy
Define triggers such as an unclear request, a disputed charge, repeated misunderstanding, a threat to safety, or a question outside the approved scope. Explain to customers when they are using automation and how to contact a person. The handoff should include the original message and relevant conversation context without unnecessarily spreading personal data.
Test the difficult requests
Use fictitious messages for missing order numbers, spelling errors, an angry customer, conflicting instructions, and attempts to change the assistant’s rules. Include a question whose answer is absent from the knowledge base. The correct behavior may be to ask for clarification or escalate. A confident invented answer is a failure, even if it sounds polite.
Compare total operating effort
Include setup, knowledge-base maintenance, human review, missed conversations, integration costs, and per-message charges. Track resolution quality and successful handoffs alongside response speed. A short response time is not a success if customers must repeat their request to a second channel. Confirm plan limits and data handling directly with vendors.
A simple pilot for a US service business
For an appointment-based business, let the assistant explain the service menu and draft a reply about availability. Keep actual booking confirmation in the scheduling system. Test a changed appointment, an unavailable slot, and a cancellation request. Staff should be able to see what happened and step in before a customer receives the wrong commitment.
Decide whether to expand
Review a sample of completed and escalated conversations. Look for repeated missing information and unnecessary handoffs. Fix the knowledge base before increasing autonomy. Keep a manual route during outages and explain expected response times honestly. Expand only after the business can support the process and account for the customer data it uses.
Put it into practice
- List five routine questions and the verified source for each answer.
- Write a human-handoff message before launching the assistant.
- Test unknown, ambiguous, and sensitive requests with fictitious data.
- Review outcomes regularly and keep the manual contact route working.
Put this guide into practice
A repair shop tests three support questions: opening hours, an overdue repair and a refund dispute. Only the verified hours answer should be automatic. The other two need an accountable staff handoff with the original message attached.
| Check | Evidence to record |
|---|---|
| Verified answer source | Observation, source or test result; responsible person; follow-up date. |
| Escalation owner | Observation, source or test result; responsible person; follow-up date. |
| Customer context preserved | Observation, source or test result; responsible person; follow-up date. |
This is a proposed exercise using illustrative circumstances, not a report of a test or visit performed by our publication.
Sources and further reading
Published by Sadddam Hussain. Research-led guidance for US readers. Read our editorial standards.
