Hospitality · Europe · LLM triage · extraction
300 reservation emails a day, on autopilot.
A central reservations team read 200 to 300 emails a day across 10+ hotels just to work out what each one was. Now every email is sorted as it lands, bookings are extracted, and replies wait for one-click approval.
Two nights in May, family of four
Hello, we would like 14 to 16 May for two adults and two children, sea view if possible. Could we check in early?
- 300/day
- Guest emails sorted as they land
- 10+hotels
- One reservations team, one queue
- 9categories
- Every email sorted into one
A hotel chain's reservations team handled every guest email for 10+ properties: bookings, questions, payment proofs, changes, complaints and automated notices, 200 to 300 a day in the same mailboxes. Before anyone could act on an email, someone had to read it and decide what it was. Threads changed meaning as guests replied, so whole conversations were read again. An earlier internal tool had been lost in a platform migration.
Every email is now sorted into one of nine categories as it arrives. About a third of the volume, fixed-format notices from booking engines, is handled by plain rules with no AI at all. Reservations are extracted with dates, rooms and guests checked against each hotel's own room list, so nothing is invented. Threads are re-sorted when a reply changes their meaning, and replies are drafted for one-click approval. Every booking still gets human eyes before it reaches the hotel system.
- 01 · What we foundAssess
Mapped the team's day across the chain. Reading to sort was the bottleneck: every email had to be understood before any work could start.
- 02 · What we builtBuild
Sorting, extraction, re-sorting when threads change, and drafted replies, built as one system between the mailboxes and the hotel booking system.
- 03 · Where it standsLive
300 emails a day sorted as they land. The team approves and handles the exceptions instead of reading the queue, and grows with volume instead of headcount.
