Building NightShift: AI That Answers Your Phone at Midnight
How we built an AI voice dispatcher that handles emergency trade calls, collects job details, and sends bookings without a human on shift.
NightShift started with a simple question. What happens to a plumber's business when a pipe bursts at 2am and nobody picks up?
We knew the answer. The customer calls someone else. The job is lost. The plumber finds out in the morning, if at all.
The obvious fix is to hire a night operator. But that does not work for a solo tradie or a small crew. You would spend $200 a night on staff to maybe take one or two calls. So we built the alternative.
The core of NightShift is a voice AI agent built on Retell AI, powered by a custom LLM endpoint we control. When a call comes in, the agent answers within two rings. We named her Sarah. She introduces herself, says the call may be recorded, and asks what is going on.
Sarah is not scripted in the traditional sense. She runs on GPT-4 with a prompt that describes her personality, her job, and how to handle different call types. Emergency calls get a faster, more focused Sarah. General enquiries get a calmer one. The model figures out the tone based on what the caller says.
After the call, a webhook fires. We use GPT-4o to pull the structured data out of the transcript. Name, address, callback number, issue, preferred time, call type. That goes into Supabase and either gets queued for approval or auto-booked into the operator's calendar.
The hard part was not the AI. It was making Sarah sound like a real person on a voice call. Voice is unforgiving. Long responses and robotic phrasing kill the experience. We kept responses under 25 words, raised the model temperature, and put real example dialogue in the prompt so the model had something to match.