Ground the reply in information you control
The assistant retrieves relevant business knowledge before asking the configured AI provider for a response. Website pages, documents, FAQs, services, and policies can supply that knowledge. The process is intended to keep replies focused on your business rather than produce a general answer to any topic.
Retrieval can miss a relevant passage, and an AI reply can still be incomplete or wrong. Add concise, consistent facts and test them. A duplicated document with an old price can create confusion even when the current price appears elsewhere.
Prepare answers for the questions visitors ask
Begin with opening hours, service areas, what each service includes, and how to contact your team. Write policies with the exceptions that change an answer. For example, if delivery varies by location, state what the team needs to check rather than giving an unconditional promise.
Use natural customer wording in your tests. Try a short phrase, a detailed question, and a follow-up referring to the previous message. Review whether the reply remains relevant and whether it offers a practical next step when the information is unavailable.
Distinguish availability from staffing
An automated reply does not require a staff member to type each message. It still depends on a working provider connection, subscription access, and available usage. Do not describe it as unlimited or always available regardless of those conditions.
Human replies depend on your team. If a visitor needs a custom decision, explain that a person must review it. Keep contact options current and avoid a greeting implying that an agent has joined when only the AI is responding.
An answer-review checklist
- Is the reply supported by current business information?
- Does it answer the visitor question before asking for contact?
- Does it state relevant exceptions or missing details?
- Does it avoid inventing a quote, booking, or approval?
- Can a visitor reach a person if the reply is insufficient?
Update knowledge when policies change
When a price, service description, or opening hour changes, update the source and remove superseded information. Reimporting a page does not excuse leaving conflicting documents in the workspace. Retest the affected question and record which answer your team approved.
Do not add private credentials or confidential customer files simply to improve answer coverage. Relevant knowledge and conversation context can be supplied to the AI provider. Review the privacy guidance and upload only information you are authorised to process.
Measure useful answers rather than reply volume
Review the questions that needed clarification and the enquiries your team could follow up. A high count of replies is not a measure of accuracy, customer satisfaction, or revenue. An assistant that asks unnecessary questions may generate activity without helping the visitor.
Start with a narrow trial and expand coverage after correcting gaps. Use the Free plan limits to assess actual usage. If a question keeps needing a person, preserve that route instead of repeatedly adding speculative instructions.
Diagnose a reply before changing the prompt
When an answer is wrong, identify which fact was missing or conflicting. Check the current website source, manually added FAQ, and uploaded document before adding a stronger instruction. If the underlying policy is unclear, the assistant cannot reliably infer what the business intended.
Try the question in a fresh test and then as a follow-up. A reply may be affected by recent conversation context as well as the retrieved knowledge. Record the exact factual correction your team wants, update the approved source, and rerun the relevant questions.
If the provider cannot reply at all, check service configuration, subscription access, and usage separately from knowledge quality. Rewriting an FAQ will not repair an unavailable provider connection. Keep the visible fallback practical and avoid repeatedly asking a visitor for details when the assistant is unable to answer. Review each correction with the staff member responsible for that information.
A controlled trial to adapt
Example to adapt: write a policy with a normal service area and one excluded location. Test both locations and a vague question that does not name an area. Compare the reply with the source your team approved. For the vague question, a clarification may be more useful than an unconditional promise. If the reply changes the exception, correct the source or retain a human review route. Keep the synthetic question and approved reference so the same check can be repeated after a knowledge update.