The Ownership Line

A carrier CEO takes a pitch this quarter from one of the new build-and-operate firms. The deck is polished. Engineers embed inside the shop, the system stands up in months, and the vendor runs it going forward so the carrier's team never has to touch the plumbing. There's a dashboard, a rollout plan, and a reference client on slide fourteen. It's a genuinely good pitch, and most of it is true. The question that decides whether it's a good deal never makes the agenda: when the engagement ends, what stays?

Carriers are right to rent much of the stack. Nobody builds their own cloud. Even the model layer, where the ownership debate got loud this week, is swappable by design: rent the frontier today, run open models when your volume makes the math work. Owning your inference is a real decision. It just isn't the first ownership decision that matters.

The agentic workforce built around a carrier's book and its underwriters' judgment is the first one. It's the one place carrier-specific advantage compounds: the underwriter's second question, the adjuster's reason for holding a claim, the exceptions the best people make without ever writing them down. Rent that layer and the learning compounds on someone else's ledger.

Rent the commodity. Own the compounding.

We learned this the slow way

We've spent nearly a decade running production AI inside a live carrier, and we didn't get the ownership question right the first time either. Early on, it was tempting to rent whatever moved fastest and sort out ownership later. What changed our approach was watching where the real learning sat: the exception patterns, the judgment calls that made the tool worth using at all, accrued to whoever controlled the workflow, not whoever sold the model underneath it. Owning that layer is what let the workforce keep improving years after the first version shipped, instead of plateauing at whatever the roadmap allowed.

A product can't be built for your book

For a decade, the responsible move was to buy vertical SaaS, because building software was hard and slow. Agentic tooling and shared model APIs flipped that math. A general-purpose underwriting or claims product still has to be built for the median carrier, because that's who's paying for it. It encodes median guidelines, median exceptions, median judgment. Your advantage was never in the median case. It was in the calls your best underwriter makes on the file that doesn't fit the template: how a hurricane-season endorsement gets handled two weeks before landfall, which venues produce claims that don't match the loss run, which declinations get a second look because the broker relationship is worth it. A platform built for everyone can't hold any of that. Score platforms on whose workflow you end up running once the contract is signed, not on feature count.

The compounding is the part worth sitting with

Every agentic workflow gets better as it runs, correcting against your claims, your declinations, your loss history. That correction loop is an asset, and it lands on somebody's ledger. Own the workflow and the asset compounds on your side: eighteen months in, your workforce knows things about your book that no general product does, and it keeps knowing more. Rent the workflow and the learning pools on the platform's side, by design. The same improvements that come from your book are what make the product better for every carrier that buys it, including the one down the street.

The capital markets just put a number on this layer. Anthropic's enterprise AI joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs launched May 4 with roughly $1.5 billion in committed capital; OpenAI's Deployment Company followed May 12 with more than $4 billion, led by TPG. Five and a half billion dollars, two announcements, eight days apart. The smartest money in AI just agreed the workforce layer is where the value sits. The open question is whose ledger the learning lands on, and that one is still each carrier's to decide. The ones who decide to own it keep the compounding for themselves.

Sources: Anthropic, "Building a new enterprise AI services company," May 4, 2026; Blackstone press release, May 4, 2026; OpenAI, "OpenAI launches the OpenAI Deployment Company," May 12, 2026.

Operators beat engineers

Building software has genuinely never been easier, and that part of every pitch is true. The part that gets misread is what became scarce instead. It was never the code. It's knowing which lever actually moves a combined ratio, which exception pattern is signal and which is noise, which claim needs a human and which doesn't. That judgment lives in the people who've run the operation, and a workforce is only worth owning if the judgment encoded in it is real. A good engineering team can ship a fast, well-built system. It takes operators, yours or a partner who has actually run a carrier, to tell whether the system is asking the right question or just a fast one.

Three questions before you sign

Has the partner actually operated inside a carrier, not just sold to one? Is the first deliverable a workflow your team can run, or a platform your team has to learn? And when the engagement ends, who owns the workforce? The first two questions sort quality. The third sorts business models, and it's the one to ask before the pilot, because the answer lives in the contract, not the demo.

The lesson from our decade: owning and renting are two different decisions, and treating them as one is how a carrier ends up owning the wrong things and renting the one thing it can't afford to.

So what

None of this requires a moonshot. It requires picking one workflow, this month, that you run yourselves: a submission triage rule, a claims note pattern, a coverage question your underwriters always ask. Own that one thing before you rent the next one, and put the three questions to whatever gets pitched to you after that. The third answer is in the contract. Read it before the demo wins the room.

If you're weighing one of these pitches right now and want to talk through what a carrier-owned version looks like, reply or reach out. That conversation is the whole point of writing this down.

Kyle Nakatsuji is the founder of Dearborn Labs and CEO of Clearcover, where the team has built and run production AI for nearly a decade.