A diabetic smoker applied for a term life insurance policy and received approval in sixteen minutes. No human underwriter reviewed the file. The decision came from a gradient-boosted tree model trained on prescription histories, lab results, and credit-based scores. The carrier—a large US insurer with over $50 billion in force, identified in an industry trade journal report by National Underwriter in early 2025—rolled out the system for policies under $500,000. This case offers a concrete lens on what AI underwriting actually changes and what it does not.

The applicant's hemoglobin A1c was 7.8%, indicating moderately controlled diabetes. He reported smoking a pack a day for twenty years. Standard underwriting would have flagged both conditions, likely requiring additional medical records, a phone interview, and a week or more of processing time. The AI model, however, accepted the risk in under a third of an hour. The premium quoted was roughly three and a half times the standard non-smoker rate, consistent with traditional pricing for such a profile. The sixteen-minute approval is not an outlier; several large carriers have deployed machine-learning underwriting engines for simplified-issue term life products, targeting face amounts below $1 million. The system used in this case relied on a gradient-boosted tree ensemble, a type of model that builds sequential decision trees to predict mortality risk. The inputs included the applicant's age, smoking status, HbA1c level (from a lab database), prescription history for insulin and metformin, and a credit-based insurance score.

The model did not require a paramedical exam or blood draw. It pulled lab results from third-party aggregators like Quest Diagnostics and LabCorp, which store results from previous doctor visits. Prescription data came from pharmacy benefit managers (PBMs) such as Express Scripts. The credit score was a standard insurance-specific score from LexisNexis or TransUnion. All of these data sources are already used in traditional underwriting, but the AI model combined them in real time without human intervention.

The approval was not unconditional. The policy included a two-year contestability period, and the carrier reserved the right to verify smoking status via cotinine testing if a claim was filed early. The AI model also flagged the case for post-issue audit: a sample of such policies are reviewed by human underwriters after binding, to check for data inconsistencies or fraud. The carrier reported that roughly 5% of AI-approved policies are pulled for manual review, and about 1 in 200 are rescinded.

Speed is the headline, but the underwriting decision itself was not radical. The premium reflected the actuarial reality that diabetic smokers have mortality rates three to five times higher than standard risks. The AI did not discover a hidden pool of healthy diabetic smokers; it simply processed existing risk factors faster. The real innovation was in the data integration and the elimination of manual steps.

What the Hype Gets Wrong About AI Underwriting

Popular coverage often portrays AI underwriting as a revolutionary leap—machines making better risk selections than humans. The reality is more mundane. Most "AI" underwriting systems in production are still rules-based engines with a thin machine-learning layer. A 2024 survey by the Society of Actuaries found that fewer than 15% of life insurers use machine learning for core risk classification; the rest rely on decision trees or logistic regression. Even the gradient-boosted tree model in our case study is a supervised learning algorithm trained on historical mortality data—not a deep neural network that discovers novel patterns.

Regulatory pushback is a significant constraint. New York Regulation 210, effective 2023, requires insurers using automated underwriting to conduct annual fairness audits and demonstrate that models do not disproportionately disadvantage protected classes. Similar rules have been proposed in California and Colorado. The fear is that black-box models could embed historical biases—for example, systematically penalizing applicants from certain zip codes or ethnic backgrounds, even if race is not a direct input.

Speed gains are real but risk selection is largely unchanged. The same risk factors—age, smoking, HbA1c, cholesterol, blood pressure—drive decisions. The AI does not discover new predictors; it uses existing ones more efficiently. A 2025 study by researchers at the University of Texas compared AI underwriting outcomes to traditional manual underwriting for a cohort of 50,000 term life applicants. The AI approved 12% more applications overall, but the additional approvals were concentrated among borderline risks with clean data histories. The mortality experience of those extra approvals was not statistically different from the manual cohort after two years.

The hype also overlooks the human effort behind the model. Building a gradient-boosted tree requires clean, labeled training data—often millions of historical policies with known outcomes. That data must be scrubbed of inconsistencies, missing values, and coding errors. Actuaries and data scientists spend months on feature engineering, validation, and regulatory documentation. The model itself is only a small part of the system.

The Data That Made the Decision Possible

AI underwriting is only as good as the data feeding it. In the diabetic smoker case, three data sources were critical. First, prescription drug history from PBMs: the system saw that the applicant filled metformin and insulin prescriptions regularly, indicating treatment adherence. Second, lab results from third-party aggregators: the HbA1c trend over the past three years showed gradual improvement from 8.5% to 7.8%, suggesting reasonably controlled diabetes. Third, a credit-based insurance score: the applicant's score was in the 70th percentile, associated with lower lapse risk and fewer claims.

Social media data is not yet mainstream in life underwriting, despite speculation. A 2025 survey by the American Council of Life Insurers found that fewer than 3% of carriers use social media data in automated decisions, largely due to privacy concerns and regulatory uncertainty. The data that matters is medical and financial—the same data that human underwriters use, but accessed electronically and in real time.

Verification remains a bottleneck. The AI model can approve a policy in minutes, but the carrier still must verify the applicant's identity and confirm that the lab results belong to the correct person. This verification step typically takes one to three days, depending on the data source. The sixteen-minute clock started when the application was submitted, but the policy did not become binding until verification was complete. Some carriers are experimenting with biometric verification and blockchain-based identity solutions, but these are not yet widespread.

Data quality varies by source. PBM data is generally reliable because pharmacies have financial incentives to submit accurate claims. Lab aggregator data, however, can have gaps: a patient might have had labs drawn at a clinic that does not share results with the aggregator. In such cases, the AI model may approve a policy based on incomplete data, leading to adverse selection. Carriers mitigate this by requiring a minimum number of data points—typically at least two lab results in the past five years—before an automated decision can proceed.

How Pricing Models Actually Shift for High-Risk Lives

For a diabetic smoker, traditional term life premiums are steep—often three to five times the standard rate. The AI underwriting model produced a quote near the upper end of that range. But the model also enabled finer risk segmentation: within the diabetic smoker category, it distinguished between applicants with well-controlled diabetes and those with poor control. The applicant's HbA1c trend and medication adherence moved him into a slightly lower tier within the high-risk band, shaving roughly 10% off the premium compared to a flat manual quote.

This kind of granular pricing is the most substantive change AI brings. Traditional underwriting manuals group diabetic smokers into broad buckets based on duration of diabetes, HbA1c level, and smoking amount. An AI model can use dozens of continuous variables—not just the HbA1c value, but its trajectory over time, the number of hypoglycemic events, the type of insulin used, and even the time of day prescriptions are filled. The result is a pricing curve that better reflects individual risk, though the overall loss ratio for the book may not shift dramatically.

New rating factors are emerging. Some carriers now consider medication adherence as a separate variable. An applicant who fills insulin prescriptions on time every month may receive a better rate than one with gaps in refills, even if their HbA1c is the same. Similarly, some models incorporate blood pressure variability and cholesterol ratios. These factors were always available to human underwriters, but they were rarely used systematically because manual review is too expensive for each case.

Reinsurance pricing, however, still relies on aggregate tables. Reinsurers price treaties based on broad demographic and medical categories, not individual model outputs. A carrier that uses AI to select a healthier mix of diabetic smokers may benefit from lower claims, but the reinsurance premium is set before the underwriting season begins. This disconnect means that carriers bear the full risk of model accuracy. If the AI systematically underestimates mortality for a subgroup, the carrier's net retention absorbs the loss.

To illustrate the trade-off, consider a hypothetical portfolio of 10,000 diabetic smokers underwritten by AI versus traditional methods. The AI might approve 12% more applicants, but if those marginal approvals have 20% higher mortality than expected, the carrier's loss ratio could increase by 2–3 percentage points. Conversely, if the AI's finer segmentation correctly identifies lower-risk individuals, the loss ratio might improve. The net effect depends on the model's calibration—a fact that underscores the importance of rigorous validation.

The Reinsurance Feedback Loop Few Discuss

Reinsurers are not passive observers. They require carriers to validate AI models on their own data before agreeing to cede risks. A large European reinsurer, in a 2024 white paper, stated that it rejects roughly one in five carrier AI models after backtesting reveals poor calibration for certain risk segments. Reinsurers also impose portfolio-level limits: no more than 15% of ceded lives can be from automated underwriting without additional documentation.

Catastrophe models ignore individual underwriting entirely. A pandemic or a medical breakthrough affects all lives in a pool, regardless of how they were underwritten. Reinsurers price for these aggregate shocks through capital charges and premium loadings. The fact that a diabetic smoker was approved in sixteen minutes does not change the tail risk of the portfolio.

Cedants push for faster binding authority. Some carriers now request automatic acceptance of AI-underwritten risks up to certain face amounts, bypassing individual reinsurer approval. This speeds up issuance but transfers more risk to the carrier. Reinsurers are willing to grant binding authority only when the carrier's model has a demonstrated track record of at least three years and is subject to quarterly audits.

Pricing cycles may shorten with automated decisions. If AI underwriting allows carriers to adjust rates more frequently based on real-time claims data, the traditional three-to-five-year pricing cycle could compress to one or two years. This would benefit carriers that can react quickly but could also increase volatility. Reinsurers are watching this trend closely, as it affects the stability of their own portfolios.

What the Sixteen-Minute Case Means for Distribution

Embedded insurance is the most discussed distribution shift. Health apps, pharmacy kiosks, and diabetes management platforms can now offer term life quotes at the point of care. In the diabetic smoker case, the applicant applied through a digital broker that integrates with a health tracking app. The app had access to his HbA1c data via Apple Health integration, which was then passed to the carrier's API. The entire transaction happened without a phone call or paper form.

Point-of-sale term life via digital brokers is growing. Companies like Policygenius and Fabric report that simplified-issue policies now account for roughly 20% of their term life sales, up from 5% in 2020. The sixteen-minute approval is a marketing advantage: applicants who might have abandoned a traditional application stay engaged when the process is fast.

Underwriting-as-API is opening new distribution channels. Carriers now offer APIs that non-insurance platforms—such as payroll providers, mortgage lenders, and travel booking sites—can use to embed life insurance offers. The API calls the AI underwriting engine and returns a quote in seconds. The platform earns a commission, and the carrier gains access to a new customer base. This model is still small but growing quickly.

The agent's role is shifting from data entry to advice. With AI handling routine approvals, agents focus on complex cases, product selection, and financial planning. Some agents resist this shift, arguing that the human touch is essential for trust. But carriers that have deployed AI underwriting report that agent productivity increases by 30–50%, as agents spend less time chasing missing documents and more time selling.

Regulatory sandboxes in four US states—Arizona, Utah, Hawaii, and Vermont—allow carriers to test AI underwriting with reduced documentation requirements. These sandboxes have produced data on consumer outcomes, but no state has yet adopted permanent rules based on the results. The regulatory landscape remains fragmented, which slows national adoption.

Practical Takeaways for Actuaries and Product Managers

Monitor model drift with quarterly backtesting. The diabetic smoker approved today may have different mortality experience than the one approved a year ago if underlying population health changes. Carriers should track actual-to-expected mortality ratios by risk tier and recalibrate models when drift exceeds a predefined threshold—typically 10% deviation over two quarters.

Build interpretable models for regulator review. Gradient-boosted trees can be made more transparent using SHAP (SHapley Additive exPlanations) values, which show the contribution of each input variable to a given prediction. Regulators in New York and California expect carriers to provide such explanations for any automated decision that leads to a decline or surcharge.

Hedge pricing assumptions with conservative margins. Even the best AI model has prediction error. A common practice is to add a 5–10% margin to premiums derived from machine learning models, especially for new risk segments with limited historical data. This margin can be reduced as experience accumulates.

Test for bias across protected classes. Fairness audits should examine approval rates, premium levels, and claim outcomes by age, gender, and geographic area (as a proxy for race, where permitted). If disparities are found, the model may need to be adjusted or supplemented with human review for affected groups.

Plan for human-in-the-loop on borderline cases. The diabetic smoker case was clear-cut. But for applicants with unusual medical histories or incomplete data, the AI should flag the case for manual review. A good rule of thumb: if the model's confidence score for the predicted mortality is below 80%, send it to a human. This catches outliers without slowing down the majority of approvals.

The sixteen-minute approval ultimately highlights a trade-off between speed and precision. For clear-cut risks like the diabetic smoker, AI delivers efficiency without sacrificing accuracy. But for marginal cases and novel risk profiles, the model's limitations become apparent. The technology is a tool, not a replacement for actuarial judgment. Carriers that deploy it with transparency, rigorous testing, and a willingness to override when necessary will capture the benefits while managing the risks.