Atlas

Roughly two-thirds of inbound tickets were variations of the same twenty questions, and the answers lived across a help centre, a Notion wiki, and three years of resolved tickets. An earlier off-the-shelf chatbot had made things worse: it answered everything, including things it had no basis for, so agents stopped trusting it and customers escalated anyway.

We treated retrieval and generation as two separate problems with two separate test sets, because a wrong answer is almost always a retrieval failure wearing a language model's voice. Every response has to cite the passages it came from; when retrieval scores below threshold the agent says it does not know and offers a handoff instead of guessing. For anything that touches an account — refunds, plan changes, cancellations — the agent drafts the action and a human approves it. A regression suite of real historical tickets runs before every deploy, so a prompt or model change that quietly degrades answers gets caught before customers see it.

The repetitive tail is handled without a human touching it, every answer is traceable to a source an agent can check in one click, and the cases that genuinely need a person reach one faster than they used to.