Answer quality is a knowledge problem
An assistant cannot be more accurate than the material it answers from, and most support knowledge bases contain contradictions that human agents silently work around. Automating on top of that surfaces every one of them, usually to a customer. The first stage of a support project is therefore a knowledge readiness review: what exists, what is current, what conflicts and who owns each answer. Teams frequently get more value from that review than from the assistant, and the assistant is only worth building once its source material is something you would be comfortable seeing quoted verbatim.
Escalate on evidence, not on sentiment
A common design escalates when a customer sounds annoyed. That is the wrong trigger: politely worded questions can be high-risk and irritated ones can be trivial. We escalate on evidence instead — no supporting source found, a policy exception requested, an account action outside the approved set, a repeat contact about the same issue. The transfer carries the full conversation and what the assistant already established, so the customer does not repeat themselves, which is the part of a hand-off people actually resent.
Pilot support automation on one low-risk request
Start with a repeatable request such as an order-status question: verify the account through the approved channel, retrieve the current status from the commerce system and answer with its source and timestamp. Do not expose another customer's record or take a refund action; missing identity, stale status, conflicting policy or an unsupported answer goes to a named support queue with the conversation context. Keep a case ID to prevent duplicate ticket creation and reconcile a failed ticket write before retrying. Compare the current and pilot resolution rate, source-supported answer rate, escalation completion and repeat-contact rate on the same sample, with the support owner reviewing failures before expansion.
How to measure the pilot
A pilot is accepted against the current process, not an industry promise. Agree the sample, baseline and owner before launch, then review the same measures after the workflow has handled real cases.