When you apply for life insurance, a human being probably does not read your application. In the time it takes to pour a cup of coffee, an algorithm has already parsed your prescription history, cross-referenced your motor vehicle records, scanned your credit behavior, and assigned you a risk score that will determine whether you pay modest premiums or get declined outright.
This is not the future; it is the present. Insurers across North America and Europe have spent the past several years quietly deploying machine learning systems to automate underwriting — the ancient art of deciding whom to insure and at what price. The efficiency gains are staggering. What once required weeks of medical exams and human review now happens in minutes for many applicants. But the opacity of these systems has created a new kind of asymmetry: the insurer knows exactly why it priced you the way it did, while you receive only a letter.
The data that shapes your premium
Traditional underwriting relied on a relatively narrow set of inputs: age, medical history, occupation, and perhaps a blood test. Algorithmic underwriting hoovers up far more. Electronic health records, pharmacy benefit manager data, consumer credit files, and public records all feed the models. Some insurers have experimented with social media analysis, though regulatory pushback has slowed that particular frontier.
The models themselves are often gradient-boosted decision trees or ensemble methods — not the large language models that dominate headlines, but older, well-understood techniques that excel at tabular data. Their virtue is predictive accuracy; their vice is that accuracy can encode historical biases. If past claims data reflects unequal access to healthcare, the model learns that inequality as signal.
Fairness in a black box
Regulators are playing catch-up. In the United States, insurance is regulated primarily at the state level, and only a handful of states have issued guidance on algorithmic underwriting. Colorado passed a law requiring insurers to test their models for unfair discrimination, but defining "unfair" in statistical terms is harder than it sounds. A model might never see your race or zip code directly, yet still produce outcomes that correlate tightly with both.
Europe's General Data Protection Regulation grants individuals the right to an explanation when subjected to automated decision-making, but what counts as an adequate explanation remains contested. Telling a declined applicant that their "risk score was elevated" satisfies no one.
The human underwriter, demoted but not extinct
Insurers insist that humans remain in the loop for complex cases — applicants with rare diseases, unusual occupations, or edge-case histories. In practice, the loop has narrowed. The algorithm handles the vast middle of the distribution, and human underwriters increasingly function as exception handlers, reviewing only the cases the model flags as uncertain. Their expertise atrophies; the model's influence grows.
For insurers, the calculus is simple. Faster decisions mean lower acquisition costs and happier brokers. More granular risk segmentation means better loss ratios. The competitive pressure to adopt these systems is intense, and carriers that cling to purely manual processes find themselves priced out of the market.
Our take
Algorithmic underwriting is neither villain nor savior. It has made insurance faster and, in many cases, more accessible — applicants who once faced weeks of uncertainty now receive instant decisions. But it has also concentrated epistemic power in the hands of insurers while leaving policyholders to guess at the logic behind their premiums. The regulatory response so far has been piecemeal and slow. Until there is genuine transparency about what these models measure and how they weigh it, the insurance industry's AI revolution will remain a quiet one — quiet, at least, for everyone except the people being scored.
