How an AI receptionist handles an after-hours emergency callout (without making it weird)

Below is a real shape of an after-hours emergency call we modelled for an Inner West Melbourne plumbing customer. The customer-side dialogue is fictional but representative; the receptionist-side dialogue is the actual logic Olivia (Synthly's AI receptionist) runs for plumbing emergencies. Notice what's said, what isn't, and where the design choices matter.

0:00 — The call connects

Two design choices in this opening that matter more than they look. First, "this is Olivia" — we name her. Customers form an immediate mental model of who they're talking to, which dramatically lowers the cognitive load of the rest of the call. Second, "What's happening?" rather than "How can I help?". The first prompts the caller to describe the situation; the second prompts them to summarise. In an emergency, you want description.

0:04 — The customer describes

This is the moment the AI either earns the rest of the call or loses it. The wrong response is anything that sounds like a form-fill — "Can I get your name and address please?" Even though you'll need both, asking for them now signals you've heard the situation as a data-collection task, not an emergency.

Olivia's actual response:

Three things happened in that response. The situation was acknowledged in plain English. The customer was given an action they can take right now to reduce damage. And the conversation was redirected from panic-mode to problem-solving mode, which is the mental state the rest of the call requires.

0:19 — Triage and capability check

While the customer turns the water off, Olivia is doing two things internally: confirming this is a real plumbing emergency (a 'real emergency' for routing purposes is one where damage is ongoing without intervention), and checking which technicians are on-call tonight against geographic proximity to the caller.

0:34 — Pricing disclosure (where many systems get this wrong)

Pricing disclosure is the part of the call that distinguishes good AI receptionists from bad ones. The wrong pattern: "I'll have someone call you back to discuss pricing." That's a 30%+ drop in conversion because the customer immediately starts dialling the next plumber while they wait. The right pattern: state the callout fee, state what it covers, confirm acceptance, dispatch. Most customers will accept; the ones who don't are usually price-shopping a non-emergency and are better filtered out at this stage anyway.

The pricing itself comes from the customer's own configuration. Olivia doesn't invent rates — she reads from the same callout-fee structure the business has set up, which respects after-hours weighting, public-holiday weighting, suburb-distance loading, and emergency-vs-standard distinctions.

1:02 — Confirmation and dispatch

The job is now in the business's calendar (Google Calendar, Outlook, or whichever they've connected). The technician's mobile has received an SMS with the address, the symptom description, the agreed price, and a one-tap callback to the customer. The business owner has received a one-line summary text: "Booked, 14 Smith St Newtown, burst pipe, $295 + parts, Jordan ETA 11:38pm." None of this is happening in a CRM after the fact — it all happened during the call.

What the call does NOT do

Equally important is what Olivia deliberately does not do in an emergency call:

• Does not collect email. Email is irrelevant to an emergency dispatch and asking for it would signal that the AI doesn't understand what kind of call it's on.
• Does not ask for a credit card. AU trades convention is post-job payment for emergency callouts. Asking up-front breaks customer expectations.
• Does not run a satisfaction survey or upsell. Both are appropriate for follow-up SMS, not the emergency call itself.
• Does not transcribe the entire call back to the customer for confirmation. That works for booking a flight; it's strange and slow for an emergency dispatch.

When the AI hands off to a human

There are scenarios where the right design is explicitly to escalate, not handle. Olivia hands off to the business owner via priority SMS in these cases:

• Gas leak suspected. The script changes — get them outside, away from the property, call 000 first if smell is strong, then dispatch.
• Property damage already extensive. Insurance assessor coordination is involved; an AI is the wrong driver for that conversation.
• Customer is distressed beyond the situation. A bereavement, a medical emergency complicating the plumbing situation. Empathy is a real-time human judgment.
• Caller asks to speak to a human. Always honoured immediately, no friction. Olivia tells the customer "I'll get the owner to call you back inside 5 minutes" and pings the owner with the full transcript.

Why this matters more than it looks

After-hours emergencies are 8–15% of the call volume of a typical AU plumbing or HVAC business, but they generate roughly 30–40% of the revenue. They are also the calls where customer loyalty is forged or broken — a customer who had their burst pipe handled calmly at 11pm becomes a customer who recommends you for the next 10 years.

Most plumbing businesses lose this call. They lose it to voicemail, to a confused after-hours diversion, to a partner who isn't sure how to quote a callout. The AI-receptionist design above is the answer to all three failure modes — but only if the design is taken seriously. A bad AI receptionist on an emergency call is genuinely worse than voicemail.