Claude drafts the first reply to travel-agent support messages.

The marketplace's ops team answers travel agents writing in about their bookings: a cancellation that has not come through, a check-in that was refused, a question about a fare type or a service charge. Each reply has to follow the marketplace's own procedures, and the facts are often scattered across the message and earlier turns in the ticket. The assistant's job is to do the reading and write the first draft, so the ops agent starts from a reply rather than a blank page.

One model call does the whole job, with no rules engine and no reply templates. Claude reads the message, the ticket history and every switched-on document in the policy library, sorts the request into one of 32 types, judges its urgency, pulls out the booking details and drafts a reply with next actions. When a procedure in the library covers the case, the draft follows it step by step. Library documents are prompt-cached so repeat drafts reuse them, and a token budget of 150,000 keeps every draft bounded.

Ops teams use it from a console or through an API. The console covers drafting, a searchable history of every draft with who asked and what was sent, the document library and team roles. Ticketing systems call the reply endpoint with their own key, one per integrating system and revocable on its own, and a cheaper classify-only endpoint returns the same read without a draft.

The assistant has no access to the booking system, and the prompt holds it to that. It will not state a booking reference, amount, fare or traveller name that is not in the message, the thread or a document. It will not confirm that anything is booked, cancelled or refunded; it commits to checking instead. It will not quote internal notes. By default a second pass audits each draft against the procedures and corrects it, which takes a draft from about 30 seconds to about 60.

The assistant is deployed behind sign-in and API keys and has been run on real support messages. Its tests cover 122 unit and integration cases and 9 browser flows. An eval sends 16 real messages through the live model, a second suite grades 17 hotel and flight cases from the client's own testing, and a replay harness re-drafts past replies and grades old against new on 12 criteria. How widely ops agents use it day to day has not been confirmed.