Insurance executives are being pushed to decide where AI belongs in underwriting, pricing, claims, distribution and risk prevention. The temptation is to start with the tool: what can it automate, what can it predict, how quickly can it lower expense? Birny Birnbaum argues that is the wrong starting point.
Birnbaum, executive director of the Center for Economic Justice and a longtime consumer representative at the National Association of Insurance Commissioners, sees genuine promise in AI and big data. His warning is that insurers must decide what outcome the technology is meant to serve before building it into the business.
“Before we choose our tools and our techniques, we must first choose our dreams and our values,” Birnbaum said. “Some tools will serve them, while others will make them impossible to achieve.”
The distinction matters because insurance AI investment is accelerating. BCG’s 2026 analysis found AI spending as a share of P&C revenue is expected to triple this year, while only 38% of carriers are generating value at scale from AI in core workflows.
The usual answer is efficiency: faster sales, faster underwriting, faster claims. Birnbaum does not object. “AI technologies hold tremendous promise for increasing the availability and the affordability of insurance,” he said. “They hold great promise for making the sale of insurance and the claim settlement process more efficient.”
But efficiency is not neutral. An AI system that identifies vulnerable building materials could help direct mitigation dollars to homes most likely to suffer losses. The same insight could also be used to further segment the book, avoid certain risks and leave consumers with fewer affordable options. The data point is the same. The outcome is not.
“So the issue isn’t what technologies are good or bad,” Birnbaum said. “It’s can we use the technology in a way that promotes the public policy goals of insurance, which are strengthening the risk pool, broadening the risk pool, creating more availability and affordability, and promoting loss prevention and risk mitigation.”
For C-suite leaders, that reframes AI governance as strategy. The first decision is not whether a model performs. It is whether the model advances the carrier’s stated purpose.
A claims tool that gathers documents for an adjuster may improve service without transferring judgment to the model. A property analytics system that identifies where roof, drainage or vegetation improvements would reduce losses can support underwriting and resilience. A pricing model designed only to identify the most profitable customers may improve targeting while worsening the broader availability problem.
That availability problem is no longer theoretical. The US Treasury’s Federal Insurance Office said in January 2025 that homeowners insurance is becoming more costly and harder to procure for millions of Americans as climate-related events intensify. Birnbaum’s view is that AI and big data could help address that problem if directed toward loss prevention and mitigation, not merely risk selection.
“The use of AI and big data can be instrumental in helping identify opportunities for investments in resilience,” he said, pointing to climate vulnerabilities in property insurance and accident likelihood in auto as examples where analytics could guide prevention.
The practical test is simple: when an AI use case reaches the executive table, ask who benefits if it works.
If the answer is only the carrier, through better segmentation, lower expense or selective appetite, the project may still be lawful and attractive. But it is not the same as a project that improves service, directs mitigation, expands availability or helps insureds reduce loss.
That does not mean every AI initiative must become a public-policy program. It does mean executives should be honest about whether they are using technology to solve an insurance problem or simply move the problem off their balance sheet.
Birnbaum argues that insurers have tended to respond to climate risk by cutting coverage, narrowing availability and shifting risk onto consumers and public programs. AI can reinforce that playbook or challenge it. The choice happens before implementation.
The most important AI question for insurers, then, is not what the model can predict. It is what the company is trying to make possible. A tool built to find profitable risks will do that. A tool built to prevent losses, improve affordability and preserve insurability must be designed, measured and governed differently from the start.
