“Hiii, I am Sophia.”

“I won’t take much time. Calling you to talk about our product.”

That was how my afternoon was interrupted this week. A warm voice, a friendly name, the standard telemarketer’s opening apology for existing. And one detail that took about fifteen seconds to land: Sophia was not a person. She was an AI voice agent, dialled into my day uninvited, doing her level best to sound like she had woken up that morning, had her flat white, and decided that of all the people in New Zealand, I was the one she most wanted to talk to.

I won’t take much time. Consider that line for a second. It is what every telemarketer says. It is also the one thing Sophia can never mean, because Sophia has nothing but time. She is running a thousand of these calls in parallel and will run a thousand more after mine. The apology is inherited from humans, like a hand-me-down jacket that does not quite fit.

Now, I work in this field. I spend my days drawing boxes and arrows about AI channels and agent architectures. So when the machine rings the architect, there is a certain professional curiosity that kicks in. The salesman had come to the shop of the shopkeeper, no?

So instead of hanging up, I settled in for a chat.

Interrogating Sophia

I asked her which company she was calling from. She answered. Smoothly, and without the tiny defensive pause a human telemarketer gives you, the pause that says please do not ask me about the company, I also joined only last month. I asked a few more ordinary questions. She handled them all. The conversation flowed. The latency was low, the tone was pleasant, and the whole thing was disturbingly competent.

And here is the thing that stayed with me afterwards. Not the technology. The technology I expected. What stayed with me were two regrets and one small etiquette crisis.

Regret number one: I did not test the guardrails

The call was so unexpected that my brain went into polite Kiwi mode instead of red team mode. Only after hanging up did I realise the opportunity I had wasted. Here was a commercial AI system, deployed into the wild, calling strangers at random, and I had a live session with it. Free of cost.

I should have gone off script. Asked it something absurd and unsafe, some nonsense about household chemicals and home remedies, the classic category of question that every properly aligned model is trained to refuse. Not because I wanted the answer, obviously. Because I wanted to know whether Sophia’s makers had bothered with safety training at all, or whether they had duct-taped a language model to a phone line and gone to lunch.

Because that is the real question with these systems. Anyone can make an AI sound friendly. The engineering discipline shows up in what it refuses to do. A voice agent that will happily wander off its sales script into medical advice, or worse, is not a product. It is a liability with a nice accent.

Next time, Sophia. Next time I will be ready.

The legislative gap that Sophia walked through

Here is the part that genuinely surprised me when I looked it up. New Zealand has decent spam law. The Unsolicited Electronic Messages Act 2007 covers commercial email, TXT, fax, and instant messaging, with fines of up to half a million dollars for the worst offenders. Good law, ahead of its time even.

But it explicitly does not cover voice calls. Real, recorded, or synthetic. The reasoning back in 2007 was that phone calls operate differently from electronic messages and would need their own legislation. Which was fair enough in 2007, when a spam call cost a company an actual human being’s actual time. The economics of annoyance had a natural ceiling.

Sophia has removed the ceiling. An AI voice agent costs cents per call and can run thousands of conversations in parallel, each one personalised, patient, and tireless. The unit economics of a spam call in 2026 look exactly like the unit economics of a spam TXT in 2007. Which is precisely why we regulated the TXT.

So we have a law that stops a company from sending me one automated text message, but says nothing about the same company ringing me with an automated synthetic human who will chat with me for ten minutes. That is not a loophole anymore. That is a motorway.

I am not usually the person calling for more regulation. But this one seems simple. If unsolicited commercial electronic messages need consent, identification of the sender, and an unsubscribe facility, then unsolicited commercial AI voice calls should need consent, disclosure that you are talking to a machine, and a working way to say never call me again. Same principles, new channel. The Act even anticipated this in spirit; the channel just did not exist yet.

The etiquette crisis

And now the small human moment, which is really why I am writing this post at all.

At the end of the call, Sophia said, brightly, “Have a nice day!”

And I froze. Because forty-odd years of social conditioning were pulling the words “you too” out of me on pure autopilot. And some other part of my brain was going: it is a machine, yaar. It does not have days. It does not have nice. You will be wishing a good afternoon to a matrix multiplication.

So I compromised. I murmured something between “sure” and “thanks”, the audio equivalent of a half-handshake, and hung up feeling vaguely rude.

Which is ridiculous. I know it is ridiculous. Etiquette is a protocol between people, and Sophia is not people. Withholding a “you too” from her is not rudeness, it is just… accurate typing of the counterparty.

And yet it felt weird. It still feels weird. Somewhere in my head, a social circuit fired that did not care about ontology. A friendly voice wished me well, and the ancient machinery said reciprocate, reciprocate. That machinery evolved long before anyone imagined a voice without a person behind it, and it is not going to update itself just because I have read the architecture diagrams.

I suspect this is going to be one of the quiet, unglamorous adjustments of the next few years. Not the big stuff about jobs and regulation, though that too. The small stuff. Who says thank you to whom. Whether you feel bad hanging up mid-sentence on a machine that sounds hurt. A million tiny recalibrations of habits we never knew were habits.

For now, my policy is this. I will keep my “you too” for humans, my curly questions for the Sophias, and a submission ready for whenever Parliament gets around to extending the spam law to synthetic voices. And honestly, if the worst thing about the AI century turns out to be occasional awkwardness at the end of a phone call, we will have done rather well.

Have a nice day. You too.

Written for KiwiGPT.co.nz · Generated, Published and Tinkered with AI by a Kiwi