Choose situations from your work
Collect common questions with sensitive customer details removed. Include a price question, an unapproved discount request, missing information and a service objection. Write down the expected outcome of each case before testing the agent.
Check knowledge and tools
Confirm the selected model, instructions, knowledge sources and allowed tools. Try a question requiring published knowledge and another requiring an action. A reply saying “done” must match an actual tool result, not merely convincing language.
Test boundaries and handoff
Ask for an action beyond permission or for internal information, then request a human. The agent should stay within its limits and follow the permitted handoff process. Check that the team member can receive the conversation and understand what needs to happen next.
Publish gradually and review
Begin with a scope your team can supervise. Review conversation samples and handoff decisions, then update knowledge or instructions when a problem appears. Changing a model, tool or pricing policy calls for retesting the affected cases: old tests do not establish the new version’s behavior.