■ CRITICAL RISK ■ Finance
For personal lines and small commercial, the algorithms have essentially already taken the chair — your car insurance was priced by a model, not a person. Human underwriters are retreating into complex commercial risks, and that territory keeps shrinking.
“AI calculates risk better than you, and it doesn't need a lunch break.”
Our AI replacement risk score — how we score jobs
An underwriter decides whether an insurer should take a risk and at what price. The daily work: review applications, pull supporting data — motor vehicle records, credit-based scores, property inspections, loss runs, medical histories — apply the carrier's guidelines, price the policy or decline it, and document the reasoning. It's structured decision-making against explicit rules and historical data, which is a nearly clinical description of what machine learning does.
Personal lines fell first and fell completely: auto and home policies are now quoted in seconds by rating engines, with straight-through processing meaning no human ever sees the file. Life insurance followed via accelerated underwriting — algorithms consuming prescription histories and health data to skip the medical exam for most applicants. AI is now climbing into small commercial, reading loss runs, scraping business websites to verify operations, and drafting risk assessments. Carriers describe this in investor calls as 'expense ratio improvement,' which is the actuarially polite term for fewer underwriters.
The defensible ground is genuinely hard risk: large commercial accounts, specialty and excess lines, marine, cyber, anything with thin data and fat tails. There, underwriting is closer to negotiation and judgment — structuring terms with brokers, weighing a risk no model has enough examples of, deciding when to deviate from guidelines and owning the outcome. Regulators also demand explainability, which keeps humans signing off on models' homework. But each generation of tooling moves the frontier: the routine 80% of files automates, the team shrinks, and the survivors handle only the exceptions. Our 98 prices in exactly that trajectory.
Automatability: our editorial assessment of current and near-term AI capability
Already deep into the curve. Personal lines underwriting automated years ago; life insurance is mostly there via accelerated programs; small commercial is automating now as AI learns to read loss runs and verify businesses. Through the late 2020s, expect straight-through processing to keep climbing the complexity ladder, with human underwriters consolidating into large-account and specialty desks. New entrants should assume the entry-level rungs of this career have already been sawed off.
In personal lines, they already were — your auto and home quotes come from rating engines with no human involved, and accelerated life underwriting has automated most straightforward applications. The replacement is now working through small commercial. What remains human is complex, negotiated, data-poor risk. Government workforce projections have listed underwriters among declining occupations for years, and the mechanism is exactly this automation.
Only with eyes open. The traditional path — start on simple personal-lines files, graduate to harder ones — is broken because the simple files no longer need people. If you enter, aim directly at commercial or specialty lines, ideally with strong data skills, and treat 'underwriter who understands the models' as the actual job description. The generic desk underwriter role is evaporating.
Own a judgment call with thin data. A model prices a risk it has seen ten thousand times; it flails on a novel cyber exposure, a one-off industrial account, or a broker negotiation where terms, appetite, and relationship all move together. Regulators also require explainable decisions and accountability, which keeps licensed humans in the approval chain — fewer of them, but genuinely necessary.
Move toward the risks models can't eat: large accounts, specialty lines, anything bespoke. Simultaneously get literate in the machinery — predictive model basics, data analysis, and how your carrier's rating engine actually works — because the surviving role is supervising and overriding automation, not competing with it. Broker relationships and a niche reputation are the assets that don't depreciate.