The Proliferation of Predictive Models in Life Insurance Underwriting

Life insurance underwriting has always been about one core challenge: making good risk decisions with imperfect information. Traditionally, that meant lengthy applications, medical exams, lab work, and extensive manual review, often stretching days into weeks or months.

Today, predictive modeling is changing that experience for many applicants. By using historical outcomes and modern analytics, life insurers can triage cases faster, reduce friction for eligible applicants, and bring more consistency to decisions, while maintaining the rigor underwriting requires. The key is doing it in a way that supports accuracy, fairness, and regulatory compliance.

What predictive modeling means in life insurance underwriting

Predictive modeling uses historical data to estimate the probability of an outcome, such as mortality risk, expected claims experience, or the likelihood that additional evidence is needed for a confident underwriting decision. In practice, this can range from traditional statistical models to newer machine learning approaches.

In the insurance regulatory context, predictive models are commonly described as using historical data with algorithms and/or machine learning to identify patterns and predict outcomes that support decisions.

However, predictive models generally do not replace underwriting judgment. Instead, predictive models typically support underwriting decisions by helping to:

• prioritize which cases need deeper review,
• identify when traditional requirements can be reduced without increasing risk beyond appetite, and
• increase consistency in how evidence is gathered and interpreted.

Where predictive models show up in the underwriting process

Predictive modeling isn’t a single “black box” step. In many underwriting programs, it’s embedded across the workflow in targeted ways.

Accelerated and automated underwriting

Accelerated underwriting (“AU”) aims to reduce or eliminate the need for a paramed exam and fluids for eligible applicants by supplementing the application with external data sources and analytics. The National Association of Insurance Commissioners (“NAIC”) describes AU as using external data plus new analytics/modeling techniques to reduce timelines from weeks to hours for some applicants.

In AU programs, predictive models may help determine whether the available data is sufficient to make a decision, whether to waive an exam/labs for a given risk profile, or when to route the case to traditional underwriting due to uncertainty.

Evidence ordering and next best action

Even when insurers still order medical requirements, models can help make evidence ordering more precise, for example, targeting Attending Physician Statements (“APS”) only when the incremental value is high, rather than using broad age/amount grids.

Risk classification and pricing support

Models may contribute to more consistent risk classification by identifying patterns correlated with outcomes, while still requiring actuarial, underwriting, and compliance review to ensure the variables and relationships are appropriate and not unfairly discriminatory.

Quality control and fraud detection signals

Predictive tools can flag inconsistencies, anomalous patterns, or application elements that warrant review. Importantly, a “flag” should trigger a thoughtful process, not an automatic adverse decision.

Why external data is a concern

The performance of any predictive model depends on the data feeding it. In life underwriting, insurers commonly use a mix of:

Traditional underwriting inputs

• application disclosures (weight, height, tobacco use, medical history, occupation, avocation)
• labs and vitals (if collected)
• APS/physician records (when ordered)

External and third‑party data sources

Many AU and predictive programs incorporate third-party sources such as prescription drug history, motor vehicle records, and the Medical Information Bureau (“MIB”), among others.

Regulators have paid special attention to external consumer data and information sources (“ECDIS”), a category that often includes non-medical or non-traditional consumer data. For example, New York’s Department of Financial Services (“NYDFS”) issued Insurance Circular Letter No. 1 (2019) which focused on ECDIS used in life underwriting and expressed concerns about discrimination risk and transparency to consumers.

NYDFS also described “external data” as data not directly related to the medical condition of the applicant that is used to supplement or proxy traditional medical underwriting or establish “lifestyle indicators.”

This matters because external data can improve speed and convenience, but it can also increase the risk of inaccurate or low‑quality inputs, outcomes that create unfair discrimination (including through proxy variables), and limited explainability for customers and regulators.

Why predictive modeling is attractive in life insurance

When implemented thoughtfully, predictive modeling can create meaningful value for both insurers and applicants. In many cases, the most visible benefit is the customer experience: predictive models allow carriers to make faster decisions for eligible applicants, sometimes within hours instead of days or weeks. That speed matters because life insurance shopping is often time-sensitive, and long underwriting timelines can increase drop-off rates. Just as importantly, predictive modeling can reduce friction in the process by allowing insurers to waive or limit paramed exams, fluids, and other invasive requirements when sufficient data exists to support a confident decision.

From the carrier’s perspective, predictive modeling also supports operational efficiency and consistency. Underwriting teams are constantly balancing speed, risk control, and staffing constraints, and models can help standardize triage decisions and evidence ordering. Rather than relying heavily on broad grids (e.g., age and face amount) to determine requirements, predictive models can route cases more intelligently—sending straightforward cases through accelerated pathways while ensuring more complex or uncertain cases receive the right level of human review. Over time, this can reduce administrative burden and underwriting expense, especially when predictive tools are paired with accelerated underwriting programs and well-integrated data sources.

Predictive modeling can also strengthen risk selection when used as decision support rather than as a replacement for underwriting judgment. Underwriters frequently have to synthesize many signals at once, such as application disclosures, prescription history, MVR data, lab results, and more. Models can help interpret combinations of variables at scale and identify patterns that might be difficult to detect consistently in manual review. This doesn’t eliminate the need for underwriting expertise, but it can help ensure that underwriter attention is focused where it adds the most value, on edge cases, exceptions, and higher-uncertainty risks.

How to responsibly use predictive modeling

Predictive modeling can create real underwriting value, but it also introduces risks that can be easy to underestimate, especially because model failures don’t always look like “failures” at first. A model can appear to perform well operationally (faster decisions, fewer requirements, improved throughput) while quietly creating compliance, fairness, or consumer harm issues in the background. For that reason, the success of predictive modeling in life underwriting depends as much on governance and controls as it does on technical performance.

Unfair discrimination and proxy bias

Even when a model does not use protected class variables directly, it can still produce discriminatory outcomes if other inputs serve as proxies or if historical data reflects unequal patterns. Regulators have been explicit on this point. For example, NYDFS highlighted that external data sources, algorithms, and predictive models can negatively impact protected classes, particularly when correlations are misunderstood or when “non-traditional” variables are used as indirect indicators of risk.

Data quality and accuracy

Predictive models can generate outputs that feel precise and authoritative, but the results are only as reliable as the data feeding them. External data can contain errors, mismatched identities, incomplete histories, or outdated records. When low-quality inputs are introduced into underwriting, the downstream impact is not just operational inefficiency, it can lead to inconsistent consumer outcomes, improper evidence ordering, or inaccurate risk classification that is difficult to detect without strong monitoring and controls.

Explainability and transparency

If a consumer is adversely impacted, whether through a decline, a higher premium class, or additional requirements, the insurer must be able to support the decision with defensible reasoning and appropriate disclosures. This is not only a customer trust issue, but also a regulatory issue: insurers need to demonstrate that their underwriting process is grounded in permissible factors, is not unfairly discriminatory, and can be explained in a way that withstands scrutiny.

Model drift

Even a well-designed model can degrade over time as applicant behavior changes, medical treatment patterns shift, data sources evolve, or macro trends affect mortality and claims experience. A model that was accurate at launch can become less reliable months or years later if performance is not continuously monitored and recalibrated. Drift is particularly dangerous because it often occurs gradually, which can allow problems to persist undetected until consumer harm or compliance concerns emerge.

Third-party and vendor dependency

While partnerships can accelerate innovation, they also expand operational and compliance risk. Insurers remain accountable for outcomes even when external parties provide the underlying tools or datasets. That means vendor reliance must be paired with strong due diligence, contractual controls, documentation standards, and ongoing oversight to ensure the carrier can validate the model’s behavior and defend its use to regulators.

How the U.S. regulatory landscape is evolving

Life insurers in the U.S. operate under a state-based regulatory system, and expectations around AI and predictive analytics have become more explicit in recent years.

NAIC Model Bulletin on AI Systems (04 December 2023)

The NAIC adopted Model Bulletin: Use of Artificial Intelligence Systems by Insurers to remind insurers that decisions affecting consumers, when made or supported by AI systems (“AIS”), must comply with applicable insurance laws, including unfair trade practices and unfair discrimination.

The bulletin outlines expectations that insurers maintain a written program for responsible AI use (an “AIS Program”) with governance, risk management controls, and internal audit functions, scaled to the risk and potential harm of the use case.

As of January 6, 2026, the NAIC published a reference list identifying multiple states that have adopted versions of the model bulletin, including adoption dates and bulletin identifiers. As of that date, twenty-five states had adopted the NAIC Model Bulletin, and four states had adopted insurance-specific regulation or guidance.

California Bulletin 2022-5 (30 June 2022)

California Department of Insurance’s Bulletin 2022-5 addresses allegations of racial bias and unfair discrimination in insurance marketing, rating, underwriting, and claims practices, reminding insurers that the use of artificial intelligence and “Big Data” must comply with California’s non-discrimination laws and must not produce disparate impacts on protected classes. The bulletin emphasizes that insurers should avoid proxy discrimination, conduct due diligence before deploying complex data and algorithmic tools, and provide consumers with specific reasons for adverse actions such as declinations or rate changes when algorithms are involved.

Colorado Regulation 10‑1‑1 (14 November 2023)

Colorado’s Division of Insurance adopted Regulation 10‑1‑1 establishing governance and risk management requirements for life insurers using ECDIS, as well as algorithms and predictive models using ECDIS.

Colorado later adopted an amended version expanding scope (including beyond life insurance) and set an effective date of 15 October 2025.

New York Circular Letter No. 1 (18 January 2019)

NYDFS’s Insurance Circular Letter No. 1 advised all life insurers licensed in New York of their statutory obligations when using ECDIS, including data used in algorithms and predictive models, in life insurance underwriting, emphasizing that such data must not result in unlawful discrimination and must be supported by sound actuarial rationale. The letter also stressed that insurers must provide adequate consumer disclosures explaining the specific reasons for any adverse underwriting decision based on external data and comply with existing anti-discrimination and transparency requirements under New York Insurance Law.

New York Circular Letter No. 7 (11 July 2024)

NYDFS’s Insurance Circular Letter No. 7 provides guidance for insurers on the responsible use of AIS and ECDIS in insurance underwriting and pricing, emphasizing governance, fairness, nondiscrimination, and compliance with existing law. The Circular Letter applies to all insurers licensed in New York and sets out expectations for comprehensive assessments, transparency, board oversight, and documentation to ensure that AIS and ECDIS do not result in unfair or unlawful discrimination.

Texas Commissioner’s Bulletin #B-0036-20 (30 September 2020)

The Texas Department of Insurance issued Commissioner’s Bulletin # B-0036-20 which reminds all regulated insurance entities, their agents, and representatives that they remain responsible for the accuracy of data used in rating, underwriting, and claims handling, even when that data is obtained from third parties, and encourages insurers to allow policyholders to review and correct such data while warning that the Texas Department of Insurance may pursue enforcement action if inaccurate data harms policyholders.

What a well-built predictive modeling program looks like

Whether a carrier builds predictive models in-house or partners with vendors, the strongest programs tend to share a consistent set of operational and governance characteristics. These programs treat predictive modeling not as a one-time technical deployment, but as an ongoing underwriting capability that requires clear controls, documentation, and monitoring over time.

Clear use cases and bounded automation

Well-built programs explicitly document what the model is permitted to do, and what it is not allowed to do. For example, a model may be authorized to support evidence ordering or triage decisions, but not to make final underwriting determinations. Programs typically establish thresholds for straight-through processing and define escalation paths so that cases involving uncertainty, complexity, or elevated risk are routed to human underwriting review.

Data governance and documentation

Well-built programs maintain strong data lineage (including where data originates, how it is refreshed, and how it is transformed), and they implement suitability checks to confirm that the data is reliable and appropriate for the underwriting use case. Retention standards and auditability are also critical, since insurers may need to reproduce model behavior and decisions during internal reviews, vendor oversight, or regulatory examinations.

Validation before and after deployment

Prior to implementation, models should be tested for performance across meaningful segments, with additional stress testing designed to identify weaknesses in edge cases or unusual variable combinations. After deployment, insurers conduct ongoing monitoring to detect drift, unexpected outcomes, or degradation in performance, ensuring the model remains aligned with underwriting intent and risk appetite over time.

Fairness and anti-discrimination controls

Effective programs embed fairness and anti-discrimination controls throughout the model lifecycle. This is not a one-time compliance step, but a recurring discipline that includes screening for prohibited variables, assessing proxy risk, and conducting disparate impact testing appropriate to the model’s use case. Mature programs also reevaluate these controls periodically, particularly when data sources change or underwriting strategies evolve.

Third‑party risk management

Insurers with best-in-class programs treat third-party risk management as a core requirement rather than an administrative formality. Many underwriting models and data sources involve vendors, and the NAIC model bulletin contemplates that regulators may request documentation related to third-party due diligence, oversight, and contracting when external models or data are used. As a result, strong programs maintain clear accountability, documented controls, and sufficient transparency to validate vendor-supported tools and defend their use in underwriting.

What this means for the future of underwriting

Predictive modeling has become a practical tool for modern life underwriting, especially as accelerated underwriting grows and insurers seek to meet digital-first customer expectations. But in the U.S. market, the winners won’t be the companies that deploy models the fastest. They’ll be the companies that can demonstrate, clearly and repeatedly, that their models are: accurate, explainable enough for the context, governed and monitored over time, and aligned with evolving state regulatory expectations.