What can it answer?
Start with approved services, FAQs and policies. Separate public information from internal documents. Decide which answers must always be handed to a person.
What data will it receive?
Consider free-text messages, uploaded files and contact details. Collect only what the job requires. A public assistant should not invite passwords, payment details or sensitive customer records.
Where do conversations go?
Identify the model provider, hosting location, logs, access roles and retention settings. Review the provider terms for the actual plan and configuration. Confidentiality cannot be inferred from a chat interface.
How does it fail safely?
Test unknown questions, attempts to override its instructions and requests for information outside the approved scope. Make the route to a person easy to find.
Who keeps it current?
Name an owner for source updates and periodic review. Prices, availability and policies change. A correct answer at launch is not a guarantee that it stays correct.
For help applying this to your site, explore website ai assistants.
This is an implementation checklist. The chosen provider and your use case determine the final data-handling design.