Die Herausforderung
The problem they faced
The clinic books appointments for multiple doctors over LINE, Taiwan's dominant messaging app. Patients message in natural language; front-desk staff hand-read each request, check the doctor's colour-coded schedule grid, and type each booking into a shared sheet — every reply and every write done by hand.
The failure modes are patient-facing and real-money: a booking promised in chat but never written to the sheet, a treatment booked with a doctor who doesn't perform it, a 30-minute procedure jammed into a 20-minute slot. The work is impossible to scale at peak, and a single wrong write erodes patient trust.
Die Lösung
How JieGou helped
JieGou runs the LINE channel as a governed operations loop, and the central design rule is that nothing is ever written autonomously. An AI agent parses each message, resolves doctor/date/time/treatment, and drafts a booking — then every draft passes a set of deterministic gates built from the clinic's real rules.
Combination treatments, excluded doctor-treatment pairs, sequential/gap procedures, per-doctor durations over the slot limit, and any booking that won't physically fit its slot are routed to a human ("專人確認"), never auto-booked. The colour-coded grid is read deterministically (10-minute review slots off-limits; single-30 vs paired-20 white slots distinguished) so the agent never offers a slot it can't honour. What can be completed cleanly becomes a flagged shadow-draft an operator clears with one tap before anything reaches the schedule. Every draft and approval is logged; the operator is the named human at the gate.
Die Ergebnisse
Measurable impact
Bookings written without a human
Booking flow
Edge cases (combo / wrong-doctor / too-long)
Audit trail
Einführungszeitplan
Von der Installation zur vollständigen Einführung
Tag 1
LINE channel connected. The agent reads the live schedule grid and starts drafting bookings into an operator review queue — shadow mode, nothing auto-writes.
Woche 1
Operators clear the daily booking queue with one-tap approvals. Deterministic gates route combinations, excluded doctor-treatment pairs, and slot-too-short cases to a human.
Monat 1
Per-doctor durations, treatment-catalog rules, and the gray/white slot logic tuned from operator feedback. The bot books the clean cases and hands the rest off — with zero ungoverned writes.
"The bot drafts; we still approve every booking before it touches the schedule. That gate is the whole reason we trust it on a patient-facing channel."
Anonymized real deployment. Clinic identity withheld; the architecture — deterministic write-gates, human-in-the-loop operator review, and audited drafts — is described faithfully. Operational specifics generalized.