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.
Everyone in insurance is asking the same question right now: which AI model should we use?
It's the wrong question.
This week, Satya Nadella articulated something that should land differently for P&C carriers than it does for most industries. His framing: every company in the AI era will have human capital and token capital. The token capital — the models trained on your specific decisions, your specific data, your specific outcomes — is the moat. Not access to a frontier model anyone can call via API.
P&C carriers have been sitting on one of the most valuable proprietary training datasets in any regulated industry. Years of underwriting decisions. Claims outcomes against those decisions. Loss patterns tied to specific risk characteristics. Adjuster behavior across hundreds of thousands of files.
Right now, most of that sits in Guidewire as an operational record. It gets queried for reporting. It almost never gets treated as a strategic AI asset.
Microsoft is building infrastructure so enterprises can take general-purpose models and fine-tune them using reinforcement learning environments built from their own workflows. The pitch is: the resulting model is yours, not shared, not feeding a frontier lab's next training run.
For P&C, this distinction matters more than in most industries. An underwriting model trained on your book, your appetite, your loss experience performs differently than a generic model. An AI claims tool trained on how your best adjusters resolve specific coverage disputes is different from one that learned from the internet.
The carriers asking 'Claude or GPT?' are thinking about the next six months. The ones asking 'what do we need to build so our data becomes our model?' are thinking about 2028.
The question isn't which model. It's whether you treat your historical decisions as raw material worth owning.
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.