The insurance adjuster, that figure of mid-century Americana who knocked on doors and eyeballed flood damage, is becoming an anachronism. In his place sits something far less visible but considerably more consequential: machine learning models that ingest thousands of data points to determine, in milliseconds, whether you are a good risk.
This transformation has proceeded almost entirely outside public view. While legislators debate the existential implications of superintelligence and commentators fret about deepfakes, the insurance industry has quietly rebuilt its core business function around algorithmic decision-making. The shift matters because insurance is not merely a commercial product — it is the mechanism by which modern societies distribute and manage risk. When algorithms decide who qualifies for health coverage, flood protection, or auto insurance, they are making determinations that shape where people can live, how they can travel, and whether illness will mean financial ruin.
The data appetite
Traditional underwriting relied on a relatively narrow set of inputs: age, location, claims history, perhaps a medical exam or a driving record. Algorithmic underwriting is omnivorous. Auto insurers now incorporate telematics data from smartphone apps that track braking patterns, acceleration, and the hours you spend behind the wheel. Health insurers analyze prescription histories, fitness tracker outputs, and consumer purchasing behavior. Property insurers overlay satellite imagery with climate models to assess wildfire and flood exposure at the individual address level.
The results are, by the industry's own metrics, impressive. Models can identify risk gradients invisible to human underwriters. A driver who brakes smoothly but accelerates aggressively presents a different actuarial profile than one who does the reverse. A homeowner whose roof shows early satellite-detectable wear may face a claim before the policy year ends. Insurers argue this precision benefits low-risk customers, who pay less, while the industry as a whole prices more accurately.
The fairness problem
But precision and fairness are not synonyms. The same granularity that enables accurate pricing can reproduce and amplify existing inequalities. Residents of historically redlined neighborhoods often face higher premiums, not because underwriters intend discrimination but because the data itself encodes decades of disinvestment. Proxy variables — zip codes, credit scores, consumer behavior patterns — can correlate with protected characteristics even when those characteristics are formally excluded from models.
Regulators have struggled to keep pace. Insurance is primarily regulated at the state level in the United States, and most state insurance codes were written for an era of human judgment and actuarial tables. The opacity of machine learning models compounds the challenge. Even insurers themselves sometimes cannot fully explain why a particular applicant received a particular rate, a problem known in the industry as the "black box" dilemma.
The insurability question
Perhaps the most profound implication concerns the boundaries of insurability itself. Insurance functions by pooling risk across populations — the healthy subsidize the sick, the lucky subsidize the unlucky. When algorithmic precision allows insurers to identify and price individual risk with increasing accuracy, the cross-subsidization that makes insurance socially valuable begins to erode. The logical endpoint is a world where those who most need coverage are precisely those who cannot afford it, while those who can afford it barely need it.
Climate change accelerates this dynamic. As extreme weather becomes more frequent and predictable, insurers are withdrawing from high-risk markets entirely. Several major carriers have stopped writing new homeowner policies in wildfire-prone regions. Flood insurance, already heavily subsidized by government programs, faces mounting pressure. The algorithms are not wrong — the risks are real — but the social consequences of accurate pricing may be unacceptable.
Our take
The insurance industry's AI transformation deserves far more scrutiny than it receives. The technology is neither villain nor savior; it is a tool that reflects the values embedded in its design and deployment. The question is not whether to use algorithms — that ship has sailed — but how to govern them. Meaningful transparency requirements, robust testing for disparate impact, and clear boundaries on data collection would be reasonable starting points. The alternative is a future where the algorithmic assessment of your risk profile determines not just your premiums but your access to the basic infrastructure of modern life.
