As a founder you wear a lot of hats: CTO, CFO, COO and CEO, sometimes all before morning tea. Most days they get along. Occasionally they have a proper stoush in your own head.
This is the story of one of those arguments.
The CTO speaks first
The CTO has been listening to call recordings again.
A caller rings a plumber's AI agent and asks for a booking “at ten.” The transcript says team. Another caller is in Paraparaumu, and the transcript has a creative go at spelling it. Someone else says they'll “pop round after the dairy,” and the agent has to work out whether that's a shop, a farm or a food group.
None of these calls failed. The agent recovered, asked again, and got there. But the CTO knows the difference between works and works properly. The Kiwi short “e” drifts towards an “i” (ten becomes tin, pen becomes pin, yes becomes yis), and speech models trained mostly on American and British audio don't always expect it. Add te reo Māori placenames, local slang and a noisy ute on State Highway 1, and there's a real gap to measure.
So the CTO wants a benchmark. That means a proper test set of New Zealand voices saying New Zealand things: times, numbers, addresses, Whangārei and Whakatāne and Ōtāhuhu, “sweet as,” “chur,” “heaps,” “bach,” “chilly bin.” Run it through every speech model and setting we might use, score the results, and repeat whenever anything changes.
“And I'm going to need a lot of AI tokens.”
The CFO is unimpressed
“Why? It works already.”
That's a fair point, and the CFO isn't being a villain. Customers aren't complaining in droves. The agent recovers from mishearings. Every dollar spent on benchmarking is a dollar not spent on sales, and running thousands of test clips through multiple models isn't free. The CFO's job is to ask whether we're polishing something nobody has noticed is dull.
The COO has a quiet word
“Actually, we've had some feedback on placenames.”
Not a flood, just a few partners mentioning that the agent stumbled over a suburb, or read an address back slightly wrong. The COO also knows something the CFO's spreadsheet doesn't show: the calls that get a complaint are a fraction of the calls where a caller noticed. Most people don't phone back to say “your robot said Taupō funny.” They just trust the business a little less.
And our partners' customers are tradies, clinics, accountants and property managers. Their callers give addresses, times and names on almost every call. That's exactly the vocabulary that's hardest to get right.
New Zealand isn't one accent
It would be easy to make this all about the classic Kiwi vowel. But the voices ringing a New Zealand business are much more varied than that.
A plumber in South Auckland takes calls from Samoan, Tongan and Cook Islands customers. An accountant in Christchurch has clients with South African, British and Filipino accents. A clinic in Hamilton hears Punjabi, Hindi and Mandarin speakers every day, often switching between English and their first language mid-sentence. Te reo Māori speakers say placenames the way they're meant to be said, which is sometimes not how a speech model expects to hear them. Then there are older rural callers, fast-talking teenagers, and people on a bad line from the back of a paddock.
That's New Zealand. So the benchmark isn't a Kiwi accent test. It's a New Zealand callers test. If our agent only works well for one kind of voice, it doesn't work for the businesses we serve.
The CEO?
This is where it's meant to be a tough call, balancing cost against quality and sentiment against spreadsheets.
Honestly, it isn't.
dareena.ai exists because we think New Zealand businesses deserve a voice agent that sounds like it's from here and listens like it's from here. That's the whole point. Plenty of global platforms can answer a phone. Very few will ever care whether “ten” comes through as “team,” because New Zealand is a rounding error on their roadmap.
For us, it's the roadmap. If we won't sweat how New Zealanders actually talk, nobody will.
So the CTO gets the tokens. The CFO gets a benchmark with a clear scope and a cost cap. The COO gets placename feedback turned into test cases, not anecdotes. And the CEO gets to sleep at night.
What the benchmark actually does
Without getting too far under the hood, the benchmark tests five things:
- Minimal pairs: words the New Zealand accent pushes together, such as ten and tin, pen and pin, and six said the Kiwi way. We check the model gets the right one in context, especially for times and numbers where a wrong guess books the wrong slot.
- Placenames: te reo Māori and English placenames across the country, spoken the way locals say them rather than the way a map app does.
- Everyday Kiwi phrasing: the slang and shorthand real callers use, so the agent understands the meaning and not just the words.
- Many voices, one country: callers from across New Zealand's communities, including Pasifika, Asian, South African and British accents and te reo Māori speakers, not just one “standard” Kiwi voice.
- Real-world audio: mobile calls, background noise and fast talkers, because nobody rings a tradie from a recording studio.
Every time we change a speech model or setting, the benchmark runs again. If a change makes New Zealand callers harder to understand, it doesn't ship.
We won't always get it right
People don't ring a business at their best. Their pipe has burst, their kid is sick, or they're running late and stressed about it. They talk fast, they mumble, they get cut off, they change their mind halfway through a sentence. Human receptionists mishear people too.
Our agent won't understand every caller perfectly every time. We'd be lying if we said otherwise. What we can promise is that each call makes it a little better. Every mishearing a partner flags becomes a test case. Every test case becomes something the benchmark checks forever.
That's the quiet advantage of owning our own platform. When we find a problem, we don't file a ticket with an overseas vendor and hope New Zealand makes their priority list. We fix it ourselves, test it against real New Zealand voices, and ship it. Some of that improvement happens in weeks, and some happens overnight.
Why this matters to partners
If you resell dareena.ai, your name is on the call. When a caller in Te Awamutu gives their address and hears it read back correctly the first time, that's your brand getting a quiet win. When it's wrong, that's on you too.
We'd rather spend the tokens than have you spend the goodwill.
Sometimes the hats argue. But when the question is “should we make it work for the way New Zealand actually talks?”, there's only ever one answer.
We improve every day, because we can.