Your service menu, a customer’s preferences, tomorrow’s appointment slots and a thumbs-down on a call are different kinds of information. Keep them separate, give them clear sources and let each one change through the right workflow.
Start with four information boundaries
- Approved business facts: services, current prices, location, timings and policies, reviewed by the owner.
- Live operational state: availability, booking status and service progress from the system that owns the transaction.
- Private customer context: relevant preferences and history, scoped to the correct business and verified customer.
- Review material: transcripts, ratings and proposed corrections awaiting review.
A booking slot should not remain “available” because an old document says it is. A customer preference should not become a general answer for another caller. And a frustrated rating should not overwrite a business policy.
Give every business fact an owner
Keep the source, review date and responsible person alongside a fact. Timings may change for a holiday. A service may be discontinued. A price may need approval. The retrieval process should favour current, approved information and show conflicts to the owner.
When onboarding documents disagree with the website, do not quietly choose the longer answer. Ask the business to resolve the conflict. Retain the correction history so the team can understand why an answer changed.
Treat external sources as governed inputs
A public page is not automatically licensed for copying into a permanent knowledge store. Use content the business owns or is permitted to provide, and follow the rules of any API or platform being used.
For Google Places in particular, storage, reuse and attribution are restricted. A place identifier can be retained under the documented exception, but that does not permit retaining the entire response. The standard terms also restrict text-to-speech use and AI-model improvement. Keep Maps-derived content out of phone answers and model-learning pipelines unless the applicable agreement expressly permits the intended use; owner-provided or independently licensed information is the default. Review the Places API policies and the Maps Platform terms against the intended use.
Use the LLM for conversation, not invented facts
The model can ask a clearer question, explain an approved service description or summarise a call. It should not fill a missing price, warranty or appointment slot from general knowledge. If the business has not supplied an answer, a transparent handoff is a valid outcome.
Make feedback a review workflow
Let an owner flag a transcript and supply the correction with a source. Then review whether it is a business fact, a private customer update or a one-off exception. Remove unnecessary personal data and check whether the correction conflicts with an existing rule.
- Save the rating and note as review material.
- Approve a scoped correction with its supporting source.
- Test the original question and related questions.
- Publish a new knowledge version.
- Monitor the outcome and keep rollback available.
This is knowledge maintenance. It is not automatically model training. A thumbs-up does not prove the answer is factually correct, and a thumbs-down does not tell the system what the right answer should be.
Test the questions that should not get an answer
Include other customers’ data, unsupported prices, outdated documents and instructions hidden inside imported text. Check that the assistant respects the business boundary, asks for clarification and avoids treating source documents as instructions.
A clinic receptionist needs an additional clear boundary: administrative help is separate from clinical advice. The safest useful knowledge base is the one that knows both what it can answer and when the business needs to take over.
