For most of the twentieth century, insurance underwriting was a craft passed down through apprenticeship. Senior underwriters taught junior ones how to read between the lines of an application, how to weigh a flood zone against a reinforced foundation, how to price the unquantifiable human tendency toward optimism about one's own health. The job required judgment, and judgment required experience.
That model is dissolving. Across property, casualty, life, and health insurance, machine learning systems now ingest thousands of variables per applicant — credit behavior, social determinants of health, satellite imagery of rooftops, telematics from vehicles — and produce risk scores in seconds. What once took a trained professional days of deliberation now happens before the customer finishes their coffee.
The efficiency is undeniable
Insurers have embraced algorithmic underwriting for reasons that are difficult to argue with. Processing costs drop dramatically when humans review only edge cases flagged by models. Pricing becomes more granular, theoretically allowing lower premiums for genuinely lower-risk customers. Fraud detection improves when patterns invisible to human reviewers become obvious to systems trained on millions of claims.
The technology also addresses a demographic problem the industry rarely discusses publicly: experienced underwriters are aging out faster than new ones are being trained. Automating routine decisions preserves institutional knowledge in algorithmic form, even as the humans who generated that knowledge retire.
The opacity problem
But efficiency has a shadow. Traditional underwriting decisions could be explained: a human could tell you why your premium increased or your application was declined. Algorithmic decisions often cannot be explained in any meaningful sense, even by the engineers who built the systems. A model might weight a combination of hundreds of variables in ways that produce accurate predictions but defy intuitive justification.
This creates regulatory headaches. Insurance commissioners in multiple jurisdictions now require that automated decisions be explainable to consumers, but the industry's definition of "explainable" remains contested. Telling a customer that their premium reflects "elevated risk factors in your profile" satisfies no one.
More troubling is the question of proxy discrimination. Models trained on historical data can encode historical biases. A system might never see an applicant's race but effectively infer it from zip code, purchasing patterns, and other correlated variables. Proving such discrimination is difficult; the algorithm itself cannot articulate its reasoning.
Our take
The transformation of underwriting is neither salvation nor catastrophe — it is simply the latest chapter in insurance's long history of using whatever tools exist to price uncertainty. The question is not whether algorithms will dominate the industry; they already do. The question is whether regulators, insurers, and consumers can negotiate a framework that captures the efficiency gains without abandoning the principle that consequential decisions about people's lives should be defensible in human terms. That negotiation has barely begun, and the algorithms are not waiting.
