Insights

Seven things to write down before you buy an AI phone receptionist, from building one

In September 2026 we built an AI phone receptionist for ourselves, made seven live test calls from a mobile, and changed something after every call. These are the seven things worth writing down before you buy one.

1. The first sentence

The greeting has to say two things: this is an AI, and the call is recorded. The first has no dedicated statute in Australia, but consumer law prohibits misleading conduct, and a system that presents itself as a person is exactly that risk. The second is a hard requirement in New South Wales, Western Australia, Tasmania, South Australia and the ACT, where recording needs everyone's consent; an 1800 number takes calls from every state, so you plan for the strictest. A caller who hears the notice and keeps talking has consented.

2. When it hands over to a person

Write it as a rule, not as a judgement call. Ours: if the caller says something is down, wants to complain, or speaks a language the system is not set up for, transfer to a named mobile, whisper the reason to the person picking up, and keep the caller on the line throughout. On our first live test the transfer target was the same phone making the call, which tests nothing; you need a second handset.

3. What it must refuse

Someone asks whether you fix printers. The right answer is a plain "we don't do that": no referral, no offer to transfer. It sounds cold, but a front desk that takes everything turns unrelated calls into your to-do list. We made it a hard rule and one of the test calls exists only to ask that question.

4. What happens on silence

Eight seconds of silence gets one prompt; a second silence gets a goodbye and a hang-up, recorded as ended by the agent. Without this rule a dead call stays open and billing.

5. How latency adds up

Each turn's delay is three parts added together: speech to text, the model's reply, text to speech. We measured 0.45 s, 0.6 s and 0.3 s; with network on top the median sat between 1.4 and 2.0 s. That number decides whether it feels like a conversation. A vendor quoting one of the three parts is telling you nothing; ask for the whole turn.

6. What seven calls changed

Speakerphone echo made the system hear its own greeting as the caller and hang up. A phone number read out with pauses was cut into twelve turns until the silence threshold went from 0.2 s to 0.5 s. Warming the model during the greeting took the first reply from two or three seconds to 0.6 s. None of that is on any product page; it only shows up on real calls.

7. Where its answers come from

Only from the material you supply, and it says so when the answer is not there. Ours uses fixed lookups for hours, location and service area, retrieval for the rest, and for something like price it says "per scoped case" rather than inventing a figure. When buying, ask for a live demonstration of a question the material does not cover.

Write these down, then go shopping

  • The greeting, word for wordAI disclosure and recording notice, planned for the strictest state.
  • The handover trigger and the target numberTest it with a second handset.
  • What it must refuseA plain "we don't do that"; no referral, no transfer.
  • The silence rulePrompt after how many seconds, hang up after how many prompts.
  • Median latency for a whole turnThree parts added together, not one of them.
  • Where answers come fromOnly your material; ask a question the material does not cover.
  • Where call records live and for how longThe notice is step one; storage and deletion need writing down too.

Tell us which step costs you the most

Contact us