Week 24 · 8–14 Jun 2026

Three angles this week

3 angles · 17 items reviewed · generated Wed 10 Jun

P&C carriers are treating AI spend as a software line item.

Observation

AI token pricing is shifting from flat subscriptions to metered consumption, with Uber capping developer spend and Anthropic's Fable 5 priced out of most enterprise budgets.

Angle

P&C carriers are treating AI spend as a software line item. It isn't. Agentic workloads — claims triage, underwriting queues, document extraction — scale token consumption in ways that make flat-fee thinking dangerous. The budget surprise isn't a future risk; it's already happening at companies like Uber.

Implication for P&C carriers

Carriers deploying agents against Guidewire workflows need token cost governance now, not at renewal. Without it, the first autonomous claims run will produce a CFO conversation nobody wants.

Everyone in P&C is asking which AI vendor to pick.

Observation

Microsoft's enterprise AI pitch is now built around private model fine-tuning — companies owning their own 'hill-climbing machines' trained on proprietary workflows, not shared frontier models.

Angle

Everyone in P&C is asking which AI vendor to pick. That's the wrong question. The carriers who will have a durable advantage are those who figure out what their proprietary data — loss history, underwriting decisions, claims outcomes — is actually worth as training signal, and build toward owning that. The model is a commodity; the trained-on-your-data version isn't.

Implication for P&C carriers

Guidewire implementations sit on years of structured decision data. Carriers that treat that as an AI training asset rather than an operational record will be materially harder to compete with in three years.

2 sources · Stratechery +1 more

P&C carriers assume their AI bottleneck is model capability.

Observation

Anthropic's biology research found that fixing data infrastructure — not upgrading models — was what closed the capability gap for AI agents operating in complex environments.

Angle

P&C carriers assume their AI bottleneck is model capability. It almost certainly isn't. The real blocker is that core systems — policy admin, claims, billing — were built for human navigation, not agent traversal. A better model won't fix an API that returns inconsistent data or a workflow that requires a screen click to advance.

Implication for P&C carriers

Before carriers invest in frontier model access, they need an honest audit of whether their Guidewire and downstream systems can actually be traversed by an agent reliably. That's the unglamorous work that determines whether AI deployments produce results or expensive demos.

2 sources · AI Secret +1 more