A voice agent that survived a 400-room property

The problem

Between 6pm and 9pm the front desk phone queue was three deep. Roughly a third of callers hung up before anyone answered, and the group had no idea what those calls were about — they simply weren’t recorded anywhere.

What we did

The agent sits in front of the queue rather than replacing it. It answers immediately, handles the four things that make up most inbound volume — new booking, modify booking, check availability, directions — and transfers anything else to a human with the conversation so far already summarised on screen.

Three things mattered more than the model choice:

• Latency. Anything above about a second of silence and callers start saying “hello?”. We stream transcription and begin generating before the caller finishes, then handle barge-in when they talk over the agent.
• Escalation. The agent transfers on any hint of frustration, any complaint, and anything involving money it can’t verify. A frustrating agent is worse than a hold queue.
• The transcript. Every turn is logged and reviewable. When a guest disputes what was agreed, there is a record.

The honest part

Sixty-eight percent handled without a human is good, not magic. The remaining third are the calls that were always going to need judgement, and the agent’s job there is to route quickly with context rather than to attempt them.

Two of the four intents took a week. The other two took a month, almost entirely because of edge cases in the property management system’s availability API.