AI in Life Insurance Underwriting and the New Regulatory Test

Life insurance underwriting is where predictive modeling meets mortality. For decades, the industry built its risk selection around medical exams, lab tests, prescription history, and actuarial tables. The process worked, but it was slow. A traditional application could take weeks. Many consumers abandoned the process before a policy was issued. Accelerated underwriting was supposed to fix that. It uses external data and predictive models to shorten the timeline from weeks to hours, sometimes without a medical exam or fluids.

The trade-off is governance. Every input that replaces an exam or a question is a new potential source of unfair discrimination. Regulators have been waiting for this. The NAIC has long warned that the use of big data and external consumer data in underwriting can produce outcomes that correlate with race, income, geography, or disability even when no protected class is an explicit input. New York DFS Circular Letter 2024-7 turned that warning into a concrete requirement: insurers must be able to demonstrate that their external data sources and AI systems do not serve as proxies for protected classes ¹. Colorado has gone further with its own rules, making the issue national.

Life underwriting is now the test case for whether AI in insurance can be both fast and fair.

What accelerated underwriting actually is

Accelerated underwriting is the practice of issuing life insurance without the traditional medical exam, blood work, and urine sample, relying instead on data from external sources and predictive models. The NAIC describes it as a response to consumer demand for faster digital service, with insurers using prescription drug history, motor vehicle records, credit attributes, and other data to process applications in hours ².

The business case is straightforward. A shorter application process reduces the drop-off rate, increases sales, and lowers administrative costs. Consumers get coverage faster. Carriers get more policies on the books. The LIMRA AI Industry Today report noted that the industry is still early in AI adoption, with many carriers focused on reengineering business processes rather than simply layering technology on top of existing workflows ³. Accelerated underwriting is one of the clearest examples of that reengineering. The broader underwriting AI governance framework, including claims and P&C dimensions, is covered in AI in underwriting insurance.

But the same external data that makes the process fast also makes it opaque. An applicant may not know that a prescription history record, a traffic violation, or a third-party risk score influenced the decision. The underwriter may not know exactly how the model combined those inputs. The examiner’s question is whether the carrier can explain the decision and prove that the explanation is fair.

The data sources that matter

The inputs in accelerated underwriting fall into four categories. Each carries its own regulatory risk.

Prescription history is the most common and usually the least controversial. It is directly related to the risk being insured. But it can still embed disparities if the data is incomplete, outdated, or influenced by factors unrelated to mortality, such as geographic access to care.

Motor vehicle records are used because they correlate with risky behavior. A history of serious traffic violations is plausibly related to mortality. But the same record may also correlate with income, geography, or disability, depending on how it is scored.

Credit attributes and financial data are more sensitive. The relationship between credit and mortality is statistical, but the mechanism is disputed. Regulators have long questioned whether credit-based scoring is a proxy for race or income. New York and Colorado both require insurers to examine such proxies carefully.

Third-party risk scores and black-box models are the highest-risk category. A carrier may license a score that combines dozens of data sources and produces a single mortality risk indicator. The carrier may not know how the score was built or what data it uses. That is a problem when the regulator asks whether the score was tested for disparate impact.

The rule is the same across all four categories. If the data is not grounded in the actual risk being insured, it is a proxy discrimination risk waiting to happen.

The proxy test

The central regulatory requirement in modern life underwriting AI is the proxy assessment. New York DFS Circular Letter 2024-7 requires insurers to demonstrate that external consumer data and information sources do not serve as proxies for protected classes that may result in unfair or unlawful discrimination ¹. The circular applies to all insurers authorized in New York, which means it applies to most national carriers.

The test is not whether the model’s creators intended to discriminate. It is whether the outcome shows a disparate impact. A carrier must test its model outputs across protected classes and, if a disparity is found, either justify it with actuarial necessity or find a less discriminatory alternative. The documentation must show the testing was done, what was found, and what the carrier did about it.

Colorado’s SB 21-169 and subsequent regulations build on the same principle. The state requires insurers to demonstrate that their use of AI, algorithms, and external data does not result in unfair discrimination ⁴. The Colorado Division of Insurance has expanded the rule to multiple lines and is actively enforcing it. For carriers writing in both states, the stricter of the two sets of expectations becomes the de facto standard.

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers adds the governance layer. It expects insurers to maintain a written AI Systems Program, validate models, test for unfair discrimination, and monitor outcomes over time ⁵. The broader program requirements are covered in AI governance in insurance. The bulletin is principle-based, but the state rules give it teeth.

Where the risk is highest

Accelerated underwriting creates risk in three predictable places.

The model that no one owns. A predictive model licensed from a vendor may sit between the application and the underwriter without a clear internal owner. If the model’s accuracy drifts, or if its inputs change, the carrier may not notice until a regulator points it out. The NAIC Model Bulletin expects the insurer, not the vendor, to be accountable for the AI systems it uses ⁵.

The override that is not real. A model may be labeled as decision-support, but if the underwriter accepts its recommendation almost every time, the human is not really reviewing. The regulator will look at the override rate, the reasons given, and whether the underwriter had enough information to disagree. A system that is theoretically overridable but practically determinative is the highest-risk configuration.

The use case that expands. A model built for accelerated underwriting of healthy applicants may be reused for broader risk classes without revalidation. A model trained on one distribution of applicants may fail when applied to a different age band, geography, or product. Monitoring must catch this before the regulator does.

What a defensible program looks like

A defensible life underwriting AI program has four components.

Disparate-impact testing. The carrier must test the model’s outcomes across protected classes. The testing should be done at deployment and repeated periodically, especially when the model is updated or applied to a new population. The results should be documented, including any disparities found and the actuarial justification for the inputs that caused them.

Less-discriminatory alternatives analysis. When a disparity is found, the carrier must consider whether the same predictive power could be achieved with a less discriminatory variable. This is not an academic exercise. Regulators will ask to see the alternatives that were tested and why the chosen input was kept.

Human override with audit trail. The underwriter must be able to override the model and must be able to explain why. The system should record the model’s recommendation, the underwriter’s decision, and the reason for any difference. This is the evidence that the AI is decision-support, not decision-making.

Vendor diligence and audit rights. For third-party models, the carrier must have contractual rights to inspect the model, understand its inputs, and receive notice of changes. If the vendor cannot provide this, the carrier should treat the model as high-risk and increase its own monitoring. The NAIC Model Bulletin makes clear that outsourcing the model does not outsource the duty to govern ⁵.

What to do before the next regulatory request

The examiner’s request will come with a small set of sharp questions. Can you list every AI system or external data source used in underwriting? Can you show the last time it was tested for disparate impact? Can you produce the override rate and the audit trail? Can you explain why a variable that correlates with a protected class is actuarially necessary? The first step to answering those questions is the AI inventory by line of business playbook.

The second step is to run the tests. The third step is to close the gaps. A variable that cannot be justified should be removed or replaced. A model that cannot be explained should not be in production. An override that never happens should be redesigned so that the human review is genuine.

Speed and fairness are not inherently in conflict. The carriers that will survive the regulatory test are the ones that built fairness testing into the accelerated underwriting process from the start, not the ones that added it after the first examiner’s letter.

Where this fits in the broader governance map

Life underwriting is the longest-standing AI use case in insurance, and it is where fairness scrutiny is most developed. The same principles apply to P&C and health, but life has the clearest precedents and the most specific state requirements. A carrier that can pass the New York and Colorado tests for life underwriting is likely to pass them elsewhere. For a business-line view of where AI risk is highest across the company, see AI use cases in insurance by business line.

Footnotes

1. New York Department of Financial Services, “Insurance Circular Letter No. 7 (2024): Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing,” July 11, 2024: https://www.dfs.ny.gov/industry-guidance/circular-letters/cl2024-07 ↩ ↩²

2. NAIC, “Accelerated Underwriting,” Insurance Topics, updated April 2026: https://content.naic.org/insurance-topics/accelerated-underwriting ↩

3. LIMRA, “The AI Industry Today: Understanding the Current State of Play,” 2024: https://www.limra.com/globalassets/limra-loma/trending-topics/ai-governance-group/the-ai-industry-today---understanding-the-current-state-of-play.pdf ↩

4. Colorado Division of Insurance, “SB21-169: Protecting Consumers from Unfair Discrimination in Insurance Practices,” 2025: https://doi.colorado.gov/for-consumers/sb21-169-protecting-consumers-from-unfair-discrimination-in-insurance-practices ↩

5. NAIC, “Use of Artificial Intelligence Systems by Insurers,” Model Bulletin adopted December 4, 2023: https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf ↩ ↩² ↩³