Define the questions it should answer
Begin with a narrow task and an audience: for example, helping staff find approved procedures. List the questions it should answer, the source documents it can use and the decisions that stay with a person.
Name an owner for source information. Outdated or contradictory material needs attention before testing. Agree how changes will be checked and made available to the assistant.
Build a representative question set
Use common questions, ambiguous requests, missing information and questions outside the agreed scope. Include conflicting sources and a question with no supported answer. Keep sensitive production records out of early experiments unless their use has been explicitly agreed.
Write down what a good response should contain, which source supports it and when it should ask for clarification. Review the responses against those expectations rather than judging how confident the wording sounds.
Check sources and access boundaries
Ask whether answers are supported by the approved material and whether the source references actually lead to that support. Check that a user cannot retrieve information outside their permitted scope.
For an assistant connected to tools, test permissions for each action separately. Drafting a reply and sending it are different capabilities. Confirm any approval step before the assistant changes a record, sends a message or makes a booking.
Rehearse the handover to a person
Test what happens when the assistant cannot help, a connected system fails or a customer asks for a human. The handover should carry useful context so someone can continue without starting the conversation again.
Example handover: a website visitor asks a question outside the assistant’s sources. It explains the limit, offers to pass the enquiry to staff and includes the conversation context in the handover. Test whether the team has enough information to continue.
Agree acceptance and ongoing ownership
Review useful answers, unsupported answers, handover success and staff review effort. Agree an acceptance threshold for the actual use case, along with the limitations that remain. A staged rollout gives you a way to observe behaviour before expanding access.
Assign responsibility for content updates, monitoring, costs and pausing the system. Keep the representative test set so changes to the assistant can be checked against previous behaviour.
Explore customer service assistants or plan your first engagement.