AI Customer Support Agent

One authoritative source per topic, with outdated versions retired rather than left alongside. Contradictory documents produce contradictory answers, and this is where most of the accuracy came from.

Policies chunked so a clause is never split mid-sentence, each chunk carrying its source URL so answers can cite where they came from and be checked.

Order status, stock and shipping come from narrow read-only function calls rather than the model's guess. Anything touching refunds was deliberately left out of scope until the correction rate proved itself.

Low retrieval confidence, frustration signals, complaints and anything about money escalate to a human — with the transcript, what the agent found and what it was unsure about attached, so the person starts where the agent stopped.

The agent first answered the team's own questions, then drafted replies for human approval so the edit rate could be measured, and only went autonomous on the question types where corrections were already near zero.