Context
Harbor shops lose inventory of a different kind: unanswered rings. During load-out and after 5pm the line goes to voicemail. Callers book the next operator. Staff who do connect start cold — no identity, no prior ask, no link already in the guest’s phone. A generic chatbot cannot be trusted with safety, Rx masks, pass redemption, or group holds.
Architecture
The system is a voice operations layer, not an open conversational model. Prompts are bound to approved company knowledge: tour inventory, timing, parking, gear, passes, safety, and escalation rules. Out-of-scope or high-risk requests transfer to a human.
Each completed turn writes structured artifacts: caller identity when captured, topic classification, transcript, staff summary, and whether a booking link was sent. Those artifacts land in the CRM so follow-up does not depend on whoever happened to answer.
The May 2026 corpus was classified across ten operational topics (booking, reschedule, tour info, gear, passes, safety, pricing, parking, refunds, Rx). That distribution is how the knowledge base and escalation map were tightened — not a vanity dashboard.
Implementation
Knowledge-bounded telephony
The agent does not invent policy. It answers from the shop’s approved material and refuses or escalates when the question is a judgment call — non-swimmer safety, refund exceptions, large groups.
In-call booking-link delivery
SMS fires in the same session, not as a next-day follow-up. 64 links in 20 days. The text is a conversion object attached to a live conversation, not a blast.
Desk-ready context
A returning human sees name, number, what was asked, what was answered, and what remains open. The guest does not re-explain the call. Staff time measured against that loop was 93 minutes in the sample window.
Intent-tagged CRM write
Contacts are created with topic and disposition. Reschedules, first-timers, pass holders, and group inquiries are separable without a person typing notes after the fact.
Non-negotiables
No open-ended generation on the brand line
The model is not allowed to price unpublished discounts, invent availability, or give medical advice. Those paths escalate.
Measured, not claimed
Revenue is a potential window from average cruise and gear add-on against booking-intent calls. It is not booked cash. The public report includes transcripts.
Sample window May 11–30, 2026. Potential revenue uses average cruise ($89) and gear add-on estimates against booking-intent volume, including a 9-guest group inquiry. Not closed cash.