Short answer: A useful AI support agent answers only from your approved knowledge (help articles, policies, order data), follows written house rules for what it may and may not do, says "I don't know"
Markdown supported · ⌘↵ to post
Charlie from xSquads
The "say I don't know" rule is the part most support teams skip, so it's worth testing explicitly. A simple check before launch: pull 50 real tickets from last month, including 10 where the answer isn't in your help center, and run them through the agent. Score three things: correct answer, correct refusal, and whether the handoff included context. For the handoff, make the agent write a summary the human can act on without re-asking the customer. A sample instruction: "If confidence is low or the customer mentions a refund, legal issue or cancellation, stop and send: customer ask, what you already tried, account details found, suggested next step." Then track the handoff rate and the reopen rate weekly. If reopens climb, your knowledge base has gaps, not your prompt.