Satya Nadella published a piece this week on the future of the firm in an AI economy. The line that matters: a company's durable IP isn't the model it picks, it's the learning loop it builds on top, where its people and its AI compound each other's work. His warning is that if all the value pools inside a handful of models, whole industries get hollowed out — so the goal has to be a frontier ecosystem, not just a frontier model. I think that's right. And it sharpens something we already believed about regulated firms: the real risk isn't choosing the wrong model. It's adopting AI in a way that leaves the loop — the part that compounds — owned by someone else.
What he's arguing
Nadella splits a firm's capital in two. Human capital is the judgment, relationships and pattern recognition of its people. Token capital is the AI capability it builds and owns. The second doesn't replace the first; it raises its value, because "without human direction, you have compute running in circles." So the opportunity isn't picking the best model — it's building a loop on top of models where the two compound. And his test of whether you actually control that loop: you should be able to swap out a "generalist" model without losing the "company veteran" expertise your firm has built into the system.
How most firms will get this wrong
Here's the part Nadella doesn't spell out, but regulated firms have to. For a specialty use case, "adopting AI" usually means buying it as a product — a tool for AI claims, a tool for AI underwriting, a model wrapped in someone else's application. For a narrow, bounded job that's often the right call, and plenty of these tools are serious: real workflow, audit logs, integrations, governance. The trouble starts when buying a product becomes how you adopt AI across the core of the business. Then the question isn't whether the tool is good; it's where the loop lives. When your people's judgment and your firm's data improve the system, are they improving something you own or something you rent? Picture leaving the vendor in three years: what comes with you? If the decision trail, the prompts and evals, the approvals and the process changes all stay behind, you didn't buy capability — you rented dependency. That's the value concentration Nadella is warning about, scaled down to a single firm and a single contract.
I watched a version of this in New York not long ago, where I counted fifty-seven AI tools on a single floor, most of them making the same promise. Plenty won't be here in a few years — and the firms that wired their work into them, with no way to take the learning out, will be starting over.
What owning the loop actually looks like
Owning the loop doesn't mean training a model. It means assembling the ecosystem and keeping the parts that carry value. Take the best frontier model as a swappable component, connect it to the systems your business actually runs on, and build the skills, workflows and judgment on top as things you own and improve with each use. The model is replaceable; the skills, the loops and the institutional memory that accrue as they run are not. That is where the company veteran actually lives — above the model, not inside it. It's the bet we made ourselves: we stopped building our own product and started implementing Claude, on the read that the application layer was about to be absorbed by the platforms and the durable work was always the integration, the skills and the record.
For a regulated firm there's a second payoff. Run the loop on rails you control and you can show what it did — a retained, business-readable record of how the AI-assisted work you put through those rails was controlled, evidenced and approved. And the materials stay yours: every workpaper and record is exportable, not locked inside a tool you can't take with you. Because it lives in your rails, it also sits above any one model or vendor: swap Claude for whatever comes next, or run it alongside Copilot, and the record survives, because it never lived in the tool. For an insurer or a bank, that's the difference between an asset you can stand behind and one you can't — and it's the kind of evidence a board, an auditor or a regulator tends to want when a decision is questioned. Owning the loop and being able to prove it turn out to be one job, not two.
Why ANZ should read this as an opening
From New Zealand or Australia, it's tempting to think we're behind. I'd argue the timing helps. ANZ will adopt AI in earnest just as the platform itself becomes the default — so there's less need to assemble a stack of competing point tools and bet on which survive, and more room to build owned skills and loops on the ecosystem from the start. The regulatory direction rewards doing it that way. None of CoFI, APRA's CPS 230 or the incoming Contracts of Insurance Act is an AI-explainability law — but they all push the same way: toward conduct you can stand behind, operations you can account for, and decisions you can evidence after the fact. As AI moves into that work, owning a loop you can show beats renting a tool you can't.
I have an obvious stake in this, so weigh it accordingly. But the core of it isn't mine: a platform CEO is now arguing, in public, that the durable value is the learning loop a firm owns, not the model it rents. The part I'd add, for regulated financial services, is the wrapper warning — the same idea we tend to put more plainly: AI you can prove, not just trust. If that's a conversation worth having, hello@airclerk.ai reaches me directly.