Three angles this week

Observation

State Farm is mandating AI use for 19,000 agents whose contracts expire in 2027, after Progressive dethroned it as America's largest auto insurer by selling more than half its policies direct, AI-driven, with no agent involved.

Angle

State Farm is solving the wrong problem. Progressive didn't win by giving agents better tools — it won by removing them from the transaction entirely. Mandating AI onto a distribution model that AI makes obsolete isn't transformation. It's an expensive way to delay an unavoidable structural decision about what the agent role is actually for in P&C.

Implication for P&C carriers

For insurance technology leaders, State Farm's move is a cautionary signal about how not to frame an AI mandate. The question isn't how to equip existing roles with AI — it's which roles belong in a world where the policy closes inside a single digital conversation. In P&C specifically, the distribution architecture is the competitive moat, not the tools sitting on top of it. Executives who conflate AI adoption with AI transformation will spend heavily and still cede ground to carriers whose product, pricing, and distribution loop is closed end-to-end without a human in it.

Observation

AI-native startups operate with 25% fewer staff at equivalent scale, with more engineers and fewer managers. Simultaneously, AI cited as reason for nearly 40% of US job cuts in May, and the top 1% of firms spend 650x more on AI per employee than the average firm.

Angle

The AI productivity story is real but wildly unevenly distributed. Most organizations are not becoming AI-native — they're grafting AI tools onto existing org designs and calling it transformation. The 650x spending gap between AI-leading firms and the median isn't just a budget difference; it reflects two completely different theories of what AI is for. One group is redesigning the work. Everyone else is augmenting the org chart.

Implication for P&C carriers

For a technology executive in an established enterprise, the benchmark that matters isn't the average firm's AI spend — it's the top decile. The research on AI-native startups is unambiguous: real structural gains come from re-engineering the product and the process around AI, not equipping existing teams with assistants. The org design question is the AI question. Firms that treat headcount reduction as the metric are missing it; the actual signal is whether AI closes the feedback loop inside the product itself, eliminating the human handoff rather than supporting it. Architecture decisions made now — about where AI sits in the workflow — will define the cost structure for the next five years.

Observation

Experiments show that multi-model AI councils — where several models deliberate and a summarizer synthesizes — drop roughly 75% of the high-value ideas that appeared in only one model's answer, even when blind judges rated those ideas as non-obvious and worth keeping.

Angle

The consensus problem in AI councils is identical to the one in human committees: the process rewards ideas that multiple participants already agree on and filters out the genuinely novel ones. Enterprises building multi-agent workflows are importing this failure mode at scale without knowing it. Smoother output is not the same as better output — and in most enterprise AI deployments, nobody is checking the difference.

Implication for P&C carriers

For technology leaders deploying multi-agent or agentic architectures, the design of how agents synthesize matters as much as the models themselves. The standard pattern — parallel generation, then one model summarizes — systematically discards the most distinctive thinking while producing something that reads well. In insurance specifically, where the value is often in the edge case (the unusual claim pattern, the overlooked exclusion, the non-consensus risk signal), this failure mode is expensive. The practical fix requires explicitly extracting and preserving minority views before synthesis, not after. This is an architecture and prompt-design problem, not a model selection problem, and it belongs in the engineering specification for any agentic workflow that makes decisions.