Short-form writing on Claude implementation, insurance AI, governance and the ANZ market. Written for executives and technical leaders making practical decisions.
Anthropic, Microsoft and OpenAI each shipped a second mode beside the chat box this year. Chat answers; a work mode does the job in your files and systems and comes back with the finished item. What changes when you flip the switch, what each of the three products actually is, and the two things it means for a regulated firm.
A ranking with the bias declared up front: we're on it, and we're first. Six firms doing AI work with NZ insurers, brokers, lenders and advice businesses, and what each one is genuinely good at. The only firm we're hard on is our own.
Part 2 of the T-shaped thesis, written from the comment thread. A plain-language Skill is copyable, and that's fine: ownership and defensibility are different tests. Two insurers starting from the same playbook end up running different systems, and the defensible part is the evidence that only accumulates.
Every vertical AI app tells you a horizontal tool can't go deep enough for your industry. We told VCs that exact story, then threw the codebase away. Why you're buying the horizontal tool anyway, and why the deep part is now a skill you own rather than a product you rent.
Anthropic just launched Claude Tag: Claude as a shared teammate inside Slack, and soon Teams. It's genuinely good, and it's the cleanest example yet of adopting AI in a way that leaves the part that compounds, and the bill, owned by someone else. The ownership test, the cost nobody's pricing, and what to keep in-house.
Most firms aren't adopting AI badly - they're not adopting it at all. No pilots, no copilots, just a vague intention to look into it someday. Change is a login away; the agency to act on it isn't. Why owner-operators are pulling ahead, why regulated firms feel it most, and how to manufacture a pocket of agency without restructuring.
Satya Nadella says a firm's durable IP is the learning loop it owns, not the model it picks. The risk for regulated firms is adopting AI in a way that leaves the loop owned by someone else - the ownership test that decides it, and what to build instead.
New York is wall-to-wall AI, and the Insurtech Insights floor had at least double the AI-native tools we expected. When capability is that commoditised, the funded giants buy the crowd - and the only question left worth answering is whether you can prove what your AI did.
A deliberately provocative line for a real thesis: one of the leading AI assistants is on track to occupy the position the web browser did over the last two decades. Why that shift matters for regulated firms in NZ and AU.
We pitched the destination. Most firms needed a doorway. Why the gap between what frontier AI can do and what regulated firms can absorb is widening, and what to sell instead.
When MCP first shipped, the spec ran ahead of the implementations. Claude Code made me more sceptical. Claude Cowork changed my mind - and made remote MCP the right layer to build on for regulated businesses.
Capture and curate every conversation and decision in your business, and you end up with something approaching a company brain. For regulated firms, that destination is downstream of plumbing most haven't built yet.
Six months building an AI workflow product with a commercial insurance design partner taught us that messy unstructured data was the right target - but the application layer was about to be absorbed by the platforms. Why we changed direction.
One workflow, scoped on a free call, built for a fixed fee.