1 Customer Service AI Agents, Live in Production.

Customer Service AI Agents: Drafting Every Reply, Sending None of Them

Support teams absorb a workload with two very different halves. One half is genuinely difficult — a customer whose problem is unusual, or who is angry for a good reason, or whose situation needs someone with authority to make a call. The other half is retrieval: finding the order, reading the account, locating the article that already answers the question, and writing a reply that says only what the record supports. The second half is most of the volume, and because it sits in the same queue as the first, it consumes the attention that the difficult cases need. Every measure the team is judged on — first response time, reopen rate, satisfaction — degrades from that competition rather than from any lack of skill.

An agent-operated support function takes the retrieval and leaves the judgement. The constraint that makes this safe is uniform across every agent below: nothing is sent to a customer without a person approving it. A support reply is a commitment made in the company’s name to someone who will act on it, and an autonomous mistake there is not a wasted minute — it is a promise you now have to honour or retract. So the agents classify the message, look up the account, draft the reply, and stop. They also refuse in specific places: a ticket the customer never confirmed is not closed, a detractor better left alone is not contacted, and a credential notice is written so that a security-aware customer can safely trust it — no links, no urgency, nothing a phishing email could imitate.