Business Problem
ACSI · Hospitality · 2020–Present · Public
Thousands of emails, six languages, one team
A hospitality membership business receives a constant stream of customer email across several products and many languages — refunds, cancellations, login resets, billing questions, order changes, technical problems, account updates and general questions. Answering each by hand is slow and inconsistent; answering with a raw language model is fluent but ungrounded — it invents order states and policies that don't exist.
The job is to read each email in its own language, classify it, look up the real facts where a reference is given, and draft a grounded, policy-correct reply — while keeping a human firmly in the loop for anything sensitive.
Lightweight for most, heavy only when needed
Most of the work is deterministic and cheap: cleaning the message, short-circuiting spam and auto-replies, detecting the language, classifying the product and issue, and looking up facts. Only two stages — drafting and quality-checking — use a language model. That keeps the economics honest: the system is lightweight for the common case and spends only where judgement is needed.
Crucially, the agent reviews every draft before a simulated send, and feedback drives a staged-automation governance model: a case-type only becomes auto-send-eligible once its acceptance history earns it.