Out of hours: the first job for an AI voice agent in Australia

Contact centres run on queueing maths, and the maths punishes any roster that is not sized for the busiest hour. Take a centre receiving 30 calls an hour at 4 minutes each, which is 2 erlangs of work arriving every hour. To answer 80 per cent of calls within 20 seconds, the target most of the industry sizes to, the Erlang C formula says you need 4 people on the phones. With 4, the average wait is 21 seconds. With 3, it is 107 seconds. With 2, arrivals equal capacity and the queue never drains.

That roster costs the same in every hour you are open, and most of those hours are nowhere near the peak. So the cost forces a choice about which single window to pay for, and nine to five wins. The 6pm caller gets the voicemail.

In Australia the window is narrower than it looks. AEST is UTC+10, AEDT is UTC+11 and AWST is UTC+8, so a line staffed nine to five in Sydney closes at 3pm in Perth, and at 2pm once daylight saving starts on the east coast. Out of hours for a Perth customer begins in the middle of the afternoon.

After five, the agent is competing with an IVR menu and a voicemail box that someone has to wade through in the morning.

That is a low bar, and it is why out of hours is the most common first step we see when an Australian business tries voice AI. Nobody on the day roster is displaced, and the agent only has to beat a dead line.

Putting a voice agent on the night line looks like an easy decision, and the easy version of it is the one that fails.

The trap is that better than voicemail is such an easy target that people aim at it. A general-purpose agent gets pointed at every after-hours call, and it swaps one broken experience for another. The caller who rang about a missed delivery is held in a five-minute conversation that ends nowhere, and they would have preferred the voicemail.

The first people to lose faith are your own staff. They read that transcript at 8am, and from then on they route around the agent. A pilot the team does not trust never becomes anything more than a pilot.

It is narrow on purpose. The agent takes a small set of queries it can finish end to end, and every other call gets the voicemail it gets today, until a measured pass rate says otherwise.

• Pick queries with a definite end: rescheduling an appointment, or confirming whether an invoice has been paid. A missed-delivery complaint has no end the agent can reach alone, so it stays out.
• Leave everything else on voicemail, unchanged. The caller outside the set gets exactly what they get today.
• Write each query as a test scenario before a real caller meets it. A scripted caller plays the query against the agent, a judge grades the result, and the scenario runs again on every change.
• Hold each query type until it passes 99 per cent of graded calls. That figure is our threshold rather than a regulation, and below it the query stays in testing and the caller stays on voicemail.
• Only then widen the set, one query type at a time, each with its own scenarios and its own pass rate.

Nobody is watching the night line live. During the day a supervisor hears a bad call within minutes. At 11pm the first person to know is whoever reads the transcripts the next morning, so the pass rate on the scenarios is the only number your team can trust before that.

Evals also decide when the pilot ends. A query type that passes 99 per cent of graded calls goes live and stays live. One that does not stays on voicemail, and nobody has to argue about it, because the number is on the screen.

The night line is the smaller of the two benefits. The larger one is elasticity. The agent that finishes a query at 9pm finishes it at 10am, so at the Monday morning peak it can take the calls that would otherwise queue past your wait target. That is the overflow pattern, and you get it almost free, because the work of proving what the agent can finish is already done.

At that point the roster covers the calls that need a person, and you size it for those instead of for the worst hour of the week. That is where the saving comes from, and it is a different saving from replacing the night shift.

Time zones decide the hours. A national line that follows Sydney closes mid-afternoon in Perth, so the out-of-hours set has to be scoped in the caller's time, and the schedule the agent runs on has to be the caller's too. The line, the closed message and the transcript timestamps all resolve against one zone you choose.

The queries that finish are industry-specific. Clinics ring about appointment times, and trades ring about a booking. Each is a query with an end, and each is the kind of thing the agent takes first.

We build custom AI solutions for Australian operators from Perth, and the build starts with your own call recordings rather than a generic script. We listen to what people ring about after five, pick the query types with an end, build the agent for exactly those, put the scenarios around it, and leave the voicemail for the rest until the numbers move. The call data that persists is stored in Sydney on Australian infrastructure, and the security page lists where each part of a call runs.