I spent yesterday at Rutherford House in Wellington for Gloves Off: Getting Serious with AI — a day put on by Serge van Dam and Dascha Cooper with NZTE and Taiawa, aimed squarely at tech companies actually shipping AI, not theorising about it. Chatham House rules in the room, but I'm allowed to share my own talk and the broad strokes of what stayed with me.
I spoke in the morning on three lessons from building AI products at Blackpearl — the same three I've been chewing on for months and have written about elsewhere. Briefly: meet your customers where they are with AI fluency; price for value but understand that your costs aren't fixed the way they used to be; and now that the cost of delight has dropped, go on and build the magical feature. That talk has its own write-up. What I want to do here is the bit that's harder to capture in a deck — what I learned from sitting in the audience for the rest of the day.
I came out with a notebook full of scribbles and a slightly humbled sense of how much is happening across NZ tech right now. A few things I'm still thinking about.
The knowledge in people's heads
Loic Estier from RossOps talked about getting the knowledge that lives in key people's heads — the senior engineer, the long-tenured operator, the person everyone walks over to ask — and making it available to anyone on the factory floor.
That landed for me. Every business has these people. The ones whose tacit knowledge is the actual moat, except no one realises it until they go on leave. AI gives us, for the first time, a credible way to extract and encode that knowledge without it feeling like a soul-crushing documentation exercise. The interesting question isn't can you do it. It's whether you're willing to invest in capturing it now, while the people are still in the building.
When the data doesn't exist
Jess Venning-Bryan from Factor was a lovely time machine back to my electricity-market days. But her sharper point was this: AI struggles where the public data doesn't exist. Where there's no shared body of training material, no commonly indexed corpus, no Reddit thread of people debating the answer, the models flounder.
It's a useful counterweight to the dominant narrative that AI can do anything. It can do a lot — in domains where humans have already done a lot, and that work is online. Niche, regulated, infrastructure-heavy spaces are where the leverage looks very different. That's a moat for some businesses. A drag for others. Worth being honest about which side you're on.
Who owns what you build
Andrew Butel from Telos posed a great provocation: when your team builds Claude Skills, or any AI-native artefact, who owns them? They're likely core IP. Are you treating them that way?
He also dropped a number that's been rattling around in my head: 10% of revenue going to AI costs. Not unusual. Not the worst case. Just the new baseline of what this kind of business looks like. If your margin model hasn't accounted for that yet, it should.
Setup is leverage
Ben Gleisner from Cogo showed a virtual GTM agent team built on Claude's Cowork. The demo was impressive, but the framing was what stayed with me: your leverage with AI is only as good as your setup.
We're past the phase where typing prompts into a chat is the impressive thing. The teams pulling ahead are the ones investing in the surrounding scaffolding — the workflows, the integrations, the data pipes, the skills library. Cheap to dismiss as plumbing. Quietly, the difference between teams that compound and teams that don't.
Eight people, a million customers
Carl Olsen and Ruby Picton from Sharesies told us their support team is eight people supporting a million customers, with around 90% of questions answered by Fin. I had to write that ratio down twice to make sure I had it right.
But what stuck with me wasn't the headline number. It was Ruby's description of the internal culture they've built — a regular drumbeat of sharing AI progress across the company. Not a Slack channel where people post links to interesting articles. A genuine, structural rhythm of showing each other what they've tried, what worked, what didn't. The compounding effect of that, over months, is enormous.
I think a lot of the AI fluency gap in companies isn't about access to tools. It's about whether anyone has built the social infrastructure for the team to learn from each other. Sharesies has.
The duo, not the trio
Cody Bunea from Re-Leased framed something I've felt for a while but hadn't found the words for: every product person is now a builder. The classic trio of PM, designer, and engineer is becoming the duo.
That's a big claim. It won't be true everywhere. But the direction of travel is unmistakable. The PMs accelerating right now are the ones who can prototype, build, and ship a thin version of an idea themselves — even if the production version still needs the full team. The PMs who can't are slower. The gap is widening.
It's a slightly uncomfortable conversation in product circles, but it's the right one to be having.
The problem still comes first
Nicole Retter from PAM gave one of the talks I keep coming back to. AI is the tool, she said, but the problem comes first. PAM is using AI to lighten the genuinely heavy mental load that mums carry — the invisible coordination work that runs households.
It's a beautifully concrete example of what AI can do when it's pointed at a real, well-understood human problem. And it cuts against the "build the AI thing first, find the problem later" pattern that's been everywhere for two years. The best AI products I'm seeing right now start somewhere very human and bring the AI to it.
The question of risk
A few people came up between sessions to ask about risk and liability. What happens when AI gets it wrong? It's a fair question — I get asked it often, especially in board contexts.
The honest answer, I think, is that the bigger risk right now is moving too slowly. The cost of a wrong build has never been lower. The cost of not building has never been higher. That doesn't mean recklessness. It means choosing experiments where the downside of failure is recoverable, and the upside of learning is real. Most of the AI risk conversation I see in the wild is framed in a way that quietly defends inaction. That's a luxury fewer and fewer businesses can afford.
What I left with
A few different threads, but one through-line: this stuff is moving fast, but the operators in the room weren't the ones theorising. They were the ones in the weeds, sharing margin numbers, showing the demos that broke, naming the things that didn't work. That's the conversation worth having.
Huge thank you to Serge and Dascha for the day. The best events make you go home and immediately want to change three things about how you work. This one delivered on that.