AI assistants
Know when the assistant should hand off.
An AI assistant can help visitors find information and describe their needs. A useful design also makes it easy to reach a person when an answer needs judgment or confirmation.
Give the assistant a limited job
Start with a defined task, such as explaining approved service descriptions or routing an inquiry. Identify the information it may use and who maintains it. Avoid asking the assistant to invent pricing, availability or promises that the business has not approved. A clear scope makes testing and maintenance more manageable.
Define the handoff conditions
Plan for missing information, conflicting answers, complaints and requests outside the assistant’s scope. A visitor should also be able to request a person directly. The handoff message should explain what happens next and avoid promising an immediate response unless the business can provide one. Responsibility for following up must be assigned.
Collect only useful context
Ask for the minimum information needed to route the request: the service, a brief description and a contact method the visitor chooses. Explain why you need it. Avoid asking for sensitive details that are unnecessary for an initial inquiry. Decide how the conversation is stored and which people may review it before launch.
Illustrative example: an unsupported request
A fictional agency assistant receives a request for a guaranteed campaign outcome. It explains that the available information does not support that promise and offers to send the question to the team. It passes along the visitor’s stated goal, without inventing a quote or result. This is a design example, not a deployed client system.
Test the exceptions, not only easy questions
Use fictional conversations with incomplete wording, different language preferences and unanswered questions. Check whether the assistant stays within scope and whether the handoff reaches the right person. Review errors and corrections after launch. A fluent answer is not evidence that the answer is accurate or that the request was resolved.
Illustrative handoff messages
Visitor message: “This question needs our team’s review. Would you like to share an email address so we can follow up?” Team summary: “Service requested: campaign planning. Question: can this budget support two markets? Language preference: Spanish. Contact method: email.” Send only the context the visitor agreed to share, with a named follow up owner. These are sample messages, not a live client system.
Further reading
This article presents Diffusion’s practical planning recommendations. The resource below provides additional context.
NIST · AI Risk Management Framework