You can't walk a hundred metres through Manhattan right now without an AI billboard, and just about every ad on the subway is selling AI for something. That was the backdrop to the few days I spent on a kiosk at Insurtech Insights in New York — and the city turned out to be the whole story in miniature.
And I'll say it plainly: it was a fantastic few days. The energy — at the conference and out across a city riding the Knicks' run to the NBA Finals — was electric, the people were warm and easy to talk to, and if you want to stand where the cutting edge of this actually is, New York is plainly the place to do it. One caveat on everything that follows: I was at our stand for most of the event, so this is a read from the exhibition floor, not the sessions. The floor told a clear enough story on its own.
We thought we knew the field. It was double.
We'd done the homework before we went — plenty of searches, plenty of mapping, a reasonably confident picture of who else was building AI for insurance. I counted fifty-seven AI-native tools on that floor, every one of them pitching some flavour of "AI for underwriting," "AI for claims," "AI for broking." That was at least double what our research had turned up, and we'd looked hard. If you'd asked me the week before, I'd have named maybe half of them.
That gap — between what we expected and what was actually in the room — is the point. Fifty-seven tools all making the same promise is what a commoditised market looks like, and you could read the same thing off every billboard outside. When "the AI does the work" is written, in some form, on every booth and every subway ad, "the AI does the work" stops being a reason for anyone to stop walking. There's no one compelling line left that pulls a passer-by in, because they've already read a version of it forty times that day. Capability isn't the differentiator any more; it's the price of admission.
That's the inverse of something I wrote before we flew out — that the bigger the gap between what AI can do and what a firm can absorb, the smaller your pitch has to be. In the US that gap has narrowed — though mostly in one specific way. It's not necessarily that firms there have more AI running in production; it's that the baseline knowledge is so much higher. When AI is on every billboard and every subway ad you pass, a working understanding builds by sheer exposure. So the capability pitch falls flat for the opposite reason: not because it's too far ahead of the buyer, but because the buyer — and every booth beside you — has already heard it.
What cut through the crowd was investment
There was a clear exception on the floor. The very large, very well-funded players did have traction — busy stands, full demo stations, actual queues. It wasn't hard to see why. They're pointing serious venture dollars at marketing and conference spend, and in a commoditised market that's precisely what buys attention. If the product claim is a wash across fifty-seven booths, the remaining lever is reach, and reach is bought. For everyone outside that top tier — us included — the floor was quieter. I'm not going to pretend otherwise.
You could almost watch the cohort effect in the room. Twelve to eighteen months ago a wave of founders saw the same opportunity, raised on it, and started building. They've shipped in the last six months, and Insurtech Insights is where a lot of them turned up at once — looking around at each other and doing the same arithmetic in real time: this is already a saturated space. The funded few are buying their way above the noise. The rest are discovering how much noise there is.
The right people still stopped
Here's the part that mattered more than foot traffic. Part of why we took the kiosk was to test reaction to something we've been quietly building — AI Process Assurance: a retained, business-readable record of how an AI-assisted decision was controlled, evidenced and approved, over the AI tools a firm already uses.
The people who did stop were, almost to a person, the right people: already using Claude, wanting to use it better. None of them were out shopping for assurance — it wasn't on anyone's list. But shown it was an option, the reception was warm: pleasant surprise, then wanting to learn more, not less.
That's the signal I'd weigh over crowd size — and it's the question in the title of this post, so let me put it precisely. Walk the floor and everyone is answering the same one: can the AI do it? Yes, obviously — fifty-seven times over. And to be fair, most of those fifty-seven claimed deep auditing too, so it isn't that nobody had thought about proof. But in every case the proof came bundled with the product: buy the bespoke tool, move the work into it, and the audit trail comes with the walls. Nobody was answering it for the AI firms are actually gravitating towards — the general-purpose assistants their people already use every day. When a regulator, a board, or a disputed decision asks why a specific AI-assisted call was made, an audit trail inside one point solution doesn't cover the Claude conversation where the rest of the decision happened. Can you prove what your AI did — the AI you already use? That record — evidenced, controlled, approved, in the tools the work actually runs through — was the thing genuinely thin on the ground.
I have an obvious horse in this race — closing that gap is exactly what we built Airclerk to do, so read the rest with the disclosure made. But I didn't fly to New York to confirm our positioning. The floor confirmed it anyway: capability drew crowds for the funded giants and almost nobody else, while provability drew the conversations actually worth having. AI you can prove, not just trust.
We'd already made this bet — and the emptying floor sealed it
The clearest validation wasn't anything that happened at our own stand. From the booth you could tell which sessions mattered by how empty the floor went — and the floor went quietest for two of them: the keynote featuring Anthropic, and the technical session featuring Anthropic. When the rooms that pull the biggest crowds at an insurtech conference are both about the model you've already bet your business on, you take the signal.
Because we'd made that bet already. We'd booked the kiosk while still building our own AI workspace, then pivoted to implementing Claude before we flew out — having decided the application layer was about to be absorbed by the platforms, and that shipping another point solution was building on sand. We booked the stand as one business and turned up as another. We'd called the direction from the outside; what we hadn't called was the scale. That's why I'm glad we made the trip: from a desk in Christchurch we'd still be carrying a map of this market at half its true size. The call wasn't just right — it was more right than we knew, and it squared with a thesis I'd argued in the abstract: picking the assistant is the easy part; the durable work is the integrations, the permissions, and the audit trail.
Why being behind is good news for ANZ
For a firm in New Zealand or Australia watching all this, the reflex is: we're behind, we'd better catch up. Wrong reflex. The timing is the gift. ANZ is going to adopt AI in earnest at the precise moment the frontier models — and the harnesses now being built directly into them — become the dominant layer. You won't have to assemble a stack from fifty-seven competing point tools and bet on which of them survives; the platform itself will be the default, and the work that actually matters sits on top of it: the workflows, the governance, the evidence.
Being behind only costs you if you treat it as a reason to copy. Treat it as a reason to leapfrog — straight to building on the platform everyone's converging on anyway, with provability designed in from the start — and it's the most useful seat in the room.
What I'm not claiming
A few honest limits on all of this:
- This is a view from one booth. I was on the floor, not in the sessions, for most of the conference — a partial vantage by construction. Treat it as one exhibitor's read, not a market report.
- Ahead on adoption isn't ahead on assurance. If anything the saturation suggests the proof problem is still wide open — solved piecemeal inside individual point solutions, unsolved for the everyday AI the rest of the work runs through.
- The funded players' traction is real, and it may compound. Investment buys attention now, and attention buys data and logos later. A saturated market doesn't mean the giants lose — it means the undifferentiated middle does.
The one takeaway
So: New York, fifty-seven booths, a city wallpapered in AI, and one question almost nobody was answering for the AI firms already use — not "can the AI do it?" but "can you prove what it did?" The funded giants are busy buying the first question's attention. The second one is still open — here, and everywhere. New York is the place to feel the frontier, and I'd go back tomorrow; it isn't the only place you can build for it. That's a far better problem for a firm in Auckland or Sydney to be building for in 2026 than another way to generate a quote — and it's the one we flew home more certain about.