Hospitality group voice agent

The sponsor here was an operations director, and the language was deflection, not models. Two thirds of inbound calls arrived between 18:00 and 21:00, when every member of floor staff was carrying plates. Missed calls were measured; missed revenue was inferred from them.

The brief: answer everything, on every channel, without adding headcount, and never let the agent invent a table that does not exist.

The agent runs on written playbooks rather than an open prompt. Reservations check live availability against the PMS and POS, honour covers and turn times, and confirm by SMS or WhatsApp. Changes and cancellations follow the group's own policy windows. Menu, allergen, parking and opening-hours questions answer from a maintained knowledge base with a last-reviewed date.

Anything involving a complaint, a large group, a deposit dispute or a special-occasion request is escalated by design. The escalation carries the full transcript and the caller's history, so the human does not start from 'sorry, could you repeat that'.

Five weeks to the first venue, then one venue per week. Voice over telephony with sub-two-second turn latency, the same agent brain behind WhatsApp and site chat, and one action layer calling the PMS and POS so all three channels can only do what the systems allow.

Every action the agent takes is logged as an event with its inputs and outputs, which is what makes the weekly review possible: containment rate, escalation reasons, cost per conversation, and the top ten things the agent got asked and could not do.

73% of conversations complete without a person, no call goes unanswered, and direct bookings are up 18% against the same period last year — mostly the evening calls that previously rang out.

Cost per conversation is tracked and capped per venue. The number that convinced the board was not the containment rate: it was that the floor staff stopped answering the phone during service.