How we got to 100,000 calls
Nobody was watching when it happened. We noticed a few days later, in March, when we pulled the monthly numbers and the running total had gone past 100,000.
That number is cumulative. It counts every inbound call that a Vozi receptionist answered on a customer's phone number, across every business on the platform, from the day we launched. Internal test calls from our own team are not in there.
When we wrote about whether to tell callers they're speaking to an AI, the sample was around 10,000 calls. It's ten times that now. The interesting thing is that the picture hasn't really changed. It has just got sharper, and I trust it more.
How callers felt about talking to an AI
Across the 100,000 calls, 96.9 per cent were scored as neutral or positive for caller sentiment. 3.1 per cent were scored negative. Put another way, fewer than one caller in thirty ended the call unhappy.
A note on how that was measured, because it matters. Nobody was surveyed. Every call transcript is scored by a language model that reads the conversation and classifies the caller's tone as positive, neutral or negative. It's the same process for every call, which makes the number consistent. But it is a model's reading of a transcript, not a customer's own rating. I'd treat it as a reliable thermometer rather than a precise one.
The negative 3.1 per cent is the part I find most useful. When we listen back to a sample of those calls, a fair share were never going to go well: wrong numbers, people chasing a specific staff member who wasn't in, callers who wanted to argue about an invoice. Some were genuinely our fault. The AI misheard a street name, or asked for a detail it had already been given. We go through a batch of the negative calls every month and that feeds straight into what gets fixed next.
What surprised me is how rarely "it's an AI" was the complaint. People are mostly fine with it, as long as it does the job quickly.
How often a person still had to step in
88.7 per cent of calls were handled entirely by Vozi. No transfer, no "let me get someone", no message left for a person to return. The remaining 11.3 per cent were passed to a human at some point during the call.
A transfer isn't a failure. Most of our customers set the rules for it on purpose. An existing client with an urgent job goes straight to the owner. A caller who asks for someone by name gets put through. Anything the AI is unsure about gets handed over rather than guessed at. I would much rather see a clean handover than an AI bluffing its way through a question about a quote it can't see.
So the 88.7 figure is really a measure of how much phone traffic a small business can take off its own plate. For a two-person trades outfit, roughly nine calls in ten never needing to reach someone on the tools is the whole point. If you want to see what that looks like on a busy day, we wrote about handling a spike of 1,000 calls earlier this year.
How fast it responds, and why we report the median
The median response time was 1.34 seconds. That's the gap between a caller finishing a sentence and Vozi starting to speak.
Human conversation has a rhythm to it. Between turns, people usually pause for well under a second, and once the gap stretches past a couple of seconds it starts to feel like the other person has stopped listening. 1.34 seconds sits inside the range where a call still feels like a conversation. It isn't instant, and I don't think it should be. A reply with no pause at all sounds like a machine reading a script.
We report the median rather than the average on purpose. Averages get dragged around by a handful of slow moments (a caller on a patchy mobile signal, or the AI checking a calendar before it answers). The median tells you what the typical caller experienced. Half of all responses were faster than 1.34 seconds; half were slower.
Do people call back?
More than 37 per cent of callers rang the same business back at least once.
We track this because of a worry we hear from almost every business owner before they sign up: "If people work out it's an AI, will they stop calling?" Over a hundred thousand calls, the answer looks like no. More than a third of callers came back, which is not the behaviour of people who were put off.
Two caveats. Some repeat calls are people ringing back because the first call didn't finish the job, so it isn't a pure satisfaction signal. And we can only match callers by phone number. Someone who rings from the office landline and then from their mobile counts as two people.
Why the average call is 58 seconds
The average call lasted 58 seconds.
That's short, and it's short for a good reason. Most calls to a small business are simple. Book me in. Are you open Saturday? How much is a call-out? Can someone ring me back? A good receptionist gets the details and gets off the phone. A bad one, human or AI, talks too much.
The under-a-minute average also puts a ceiling on how much "AI-ness" a caller is exposed to. If the whole exchange is three questions and a confirmation, there isn't much room for the conversation to go sideways. It also helps that Vozi is trained on Australian speech, so it isn't asking people in Frankston to repeat their suburb.
How we measured this
- Sample
- Every inbound call answered by Vozi on a customer's number, from launch until the 100,000 mark in March 2026. Internal test calls excluded.
- Sentiment
- Scored per call by a language model reading the transcript and classifying the caller as positive, neutral or negative. Not a caller survey.
- Handled by AI
- A call with no transfer to a person at any point.
- Response time
- Median gap between the end of the caller's speech and the start of Vozi's reply, measured on the live call.
- Repeat callers
- The same phone number calling the same business more than once in the period.
- Call length
- Mean duration of connected calls.
What's missing: we haven't broken these numbers down by industry yet. Trades, legal and health practices behave differently on the phone and the averages hide that. That's the next piece of work.
What these numbers don't tell you
A few things I'd want to know if I were reading this as a business owner rather than writing it.
These are platform-wide averages. Your business won't get 88.7 per cent handled by AI on day one. That number depends on how your call flows are set up, how many of your calls are genuinely simple, and how often you want the AI to hand over. Some customers deliberately transfer far more than average, and that's the right call for them.
Sentiment doesn't measure outcome. A caller can be perfectly pleasant and still not book. We track bookings and captured leads separately and they're not in this piece.
And 100,000 is a big number for us, but it isn't a big number for the phone industry. A single large call centre would get through that in a week. What makes it useful is that the calls are real, they're Australian, and they're the kind of calls a small business actually gets: the after-hours enquiries, the quick questions, the bookings.
What's next
Two things. First, the industry breakdown, because "how does this go for a dental clinic" is a better question than "how does this go on average". Second, getting that negative 3.1 per cent down, and the only honest way to do that is to keep listening to those calls.
If you spot something in these numbers that doesn't add up, tell me. My LinkedIn is just below and I read every message.
Hear it rather than read about it.
Vozi will ring your mobile and have the conversation with you. Or book a 15-minute setup call and we'll have it answering your phone by tomorrow morning.
Nick Davenport
Founder of Vozi, an Australian voice AI receptionist used by small businesses in trades, legal, health and property. Nick works directly with customers on their call flows and reviews call data every month.
Nick on LinkedIn