Roughly half of a support queue is the same fifty questions asked in different words, and answering them by hand is expensive and dull. An AI agent that resolves those — actually resolves them, with the right answer and the right action — frees the team for the half that needs a person. The trap is measuring deflection rather than resolution: a customer who gave up is 'deflected' too.

What it should handle

Order and delivery status, with live data from the order system. Account questions it can look up. Password and access resets it can trigger. How-to questions covered by documentation. Policy questions with a documented answer — returns windows, cancellation terms. Booking changes. Each of these has a clear answer or a clear action, and the agent should take the action, not just describe it.

What it should route

Complaints. Anything involving money beyond a documented refund rule. Bugs and outages. Angry customers. Anything the customer has already asked twice. Vulnerable customers, where detectable. The routing should carry the conversation and any lookups the agent already did, so the human starts ahead.

Measure resolution honestly

Resolved means the customer confirmed it was resolved or did not return within a window, not that the conversation ended. Track resolution rate, escalation rate, customer rating per conversation, and the return rate — how many resolved conversations come back within seven days. Deflection with a high return rate is a cost moved, not saved.

Feed the knowledge base from the escalations

Every escalated conversation is a question the agent could not answer. Cluster them weekly, write the answers, add them to the knowledge base, and the escalation rate falls. The support team's job shifts from answering the same question to writing it down once — a better job, and a cheaper one.