How algorithmic underwriting works and where fairness questions arise

Algorithmic underwriting means a computer system, rather than a person reading a file, makes or heavily shapes the decision on a life insurance application. The system takes your answers and data pulled from outside sources, applies fixed rules or a statistical model, and returns one of a few outcomes: approve at a rate class, ask for more evidence, or send the file to a human underwriter. When it works, applicants get answers in minutes. When it goes wrong, the reasons can be hard to see, and the questions that follow are about accuracy, fairness and what you are entitled to know.

This explainer lays out the mechanism step by step, then turns to where fairness questions arise and what regulators generally expect. It uses no current industry statistics, and the example numbers are invented. For the rules in particular states, see the guides on AI and data in life insurance underwriting and AI regulation in insurance: Colorado, the NAIC bulletin and New York.

What underwriting is trying to do

Life insurance pricing rests on an estimate of how likely it is that the insured person will die during the coverage period, compared with other people the insurer covers. Underwriting sorts applicants into groups of similar expected mortality, called rate classes, so that each person pays a price that roughly reflects their group's risk. The sorting has always relied on information: the application, an exam and labs, medical records, and checks of outside records. The traditional version has a human underwriter reading all of it.

The algorithmic version keeps the goal and changes the method. Instead of an underwriter reviewing each page, a system reads structured data and applies logic. The insurer keeps human underwriters for files the system cannot decide, and the proportion of files the system handles alone is a business choice.

Step 1: The data that goes in

An automated decision starts with inputs. Common sources, described in more detail in accelerated underwriting: approval without an exam:

• Your application answers. Age, height and weight, tobacco use, health history, occupation, hobbies and finances. These are the foundation. A misstatement here is the cause of many later problems.
• The MIB. A membership organization of insurers keeps coded summaries of prior applications' medical findings. See The MIB report: what insurers know about your history.
• Prescription histories. A database check lists drugs dispensed to you over past years, which can point to conditions you did not mention. See how insurers check your prescription history.
• Motor vehicle records. Violations, suspensions and DUIs.
• Identity and fraud checks. Whether the person and details are real.
• Other external data. Some insurers or vendors use consumer-level attributes, credit-based information or marketing data in scores. See do life insurers check your credit?. Some programs also accept data from wearables with consent, covered in wearables and wellness programs in life insurance.
• Medical data pulled with your authorization. Electronic health records and lab results, where the applicant signs a release.

The data are only as good as the records. A wrong prescription entry or a mislabeled driving record goes straight into the decision. Your right to see and correct those records is a recurring theme in the rules.

Step 2: Rules, the simplest kind of algorithm

The earliest and still most common automated systems are rule-based. Underwriters and actuaries write if-then statements, and the system runs them in order. Some rules decide on their own. Others add up points.

Say a hypothetical insurer uses a simple scorecard for a term product. Each applicant starts at zero:

Under the rules, a score of 0 to 2 is approved instantly in the best available class, 3 to 5 is approved in a standard class or routed for an exam, and 6 or more goes to a human underwriter. An applicant with a clean build, no concerning drugs, no MIB code and one minor violation scores 1 and is approved instantly. A second applicant with a slightly outside build (2 points) and a prescription tied to a managed condition (3 points) scores 5 and is sent to an exam. Every point can be traced, and a person can explain it. The scorecard, its thresholds and the invented points are illustrations only.

Rules are transparent to the people who wrote them, and they are easy to audit. Their weakness is that they treat factors one by one and may miss combinations. They also embed the assumptions of the people who set them.

Step 3: Predictive models, a different kind of logic

A predictive model, which may use machine learning, works differently. Instead of a human writing the rules, the model is trained on historical data. Given many past applicants with known outcomes, such as who died during the coverage period or what rate class a full underwriting process gave them, the model learns the combination of inputs that best predicts the outcome. Its output is typically a score or probability, which the insurer converts into a decision using thresholds.

Models can find patterns that rules miss, and they can use many weak signals together. They can also be trained to imitate the decisions of human underwriters, a common approach when real mortality outcomes take decades to observe. In that case, the model learns the underwriters' habits, including any bias in them.

What changes for the applicant is mostly visibility. A model with hundreds of inputs may not produce a short list of reasons. The insurer may be able to say which factors pushed a score up or down, using explanation techniques, but those explanations are approximations. A team inside the insurer, usually with actuaries and data scientists, builds, validates and monitors the model.

Step 4: The decision and the fallback

The system's output becomes a business decision, usually one of:

1. Approve at a stated rate class, with an offer.
2. Request evidence: an exam, labs, or medical records, often described as a fallback to traditional underwriting.
3. Refer to a human underwriter for a decision.
4. Decline in clear cases, though many insurers prefer to send those to a person.

Insurers often review samples of automated approvals, and some require exams on a random subset to test whether the model's results hold up. Accuracy checks after the fact are a normal part of managing these systems. For the applicant, the practical point is that a "no instant approval" result is rarely a decline; it generally means the file needs more evidence.

Where fairness questions arise

The concern is not that insurers use data. Insurance depends on classifying risk, and differences in price for real differences in risk are the point. The concern is that a model or rule can treat people differently in ways that are not justified by risk, or that track protected characteristics. The recurring issues:

Proxies. A factor that looks neutral can correlate with a protected characteristic such as race, and so stand in for it. A consumer-marketing attribute, a neighborhood-level measure or a purchasing pattern can do this even though the model never sees race. Regulators call the result unfair discrimination when it occurs without a sound underwriting basis.

Biased training data. A model trained on past decisions learns from them. If past practices disadvantaged a group, whether through underwriting rules or unequal access to health care that shows up in records, the model may reproduce the gap.

Unequal data quality. Records are not equally complete for everyone. Applicants with sparse prescription or medical histories may be scored differently from people with rich records, for reasons unrelated to health.

Feature interactions. In a model with many inputs, a combination of harmless-looking factors can act as a proxy even if no single factor does.

Feedback loops. If a model routes certain applicants to slower, more intrusive processes, fewer of them complete the purchase, and the data the insurer later sees reflects that routing.

Lack of explanation. An applicant who cannot learn why a result occurred cannot spot an error, and a regulator may struggle to test the result.

A simple fairness check, in words and numbers

Insurers and regulators often start with a plain comparison. Say a hypothetical model approves 700 of 1,000 applicants in Group A instantly, and 560 of 1,000 in Group B. The approval rates are 70% and 56%. Dividing the lower rate by the higher gives 56 ÷ 70 = 0.80. A gap does not prove unfairness, because the groups may differ in real risk factors, and the invented figures here are not a legal test. It is a prompt to ask why. The follow-up is to find which inputs drive the gap and whether they have a sound relationship to mortality. Colorado's rule for life insurers, discussed below, requires insurers to test for unfair discrimination and describe how.

What regulators generally expect

There is no single federal law on how life insurers may use algorithms. The rules come from state insurance regulators and from general law. The main developments, with their dates as stated in the guides:

• Colorado enacted Senate Bill 21-169 in 2021, and its Division of Insurance adopted Regulation 10-1-1 for life insurers, effective November 14, 2023. The regulation requires a governance and risk management framework for use of external consumer data, algorithms and predictive models, with board oversight, written policies, a model inventory, vendor oversight and documented testing for unfair discrimination.
• The NAIC adopted a Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023. It is guidance, not a statute, and each state decides whether to issue it. Insurers are expected to comply with existing law and to maintain a written program covering governance, risk management and oversight of vendors. The NAIC publishes an adoption map.
• New York's Department of Financial Services issued Insurance Circular Letter No. 7 on July 11, 2024. It expects insurers to show that outside data and AI systems are not unfairly discriminatory, and to give applicants specific reasons for an adverse decision.

The common thread is accountability, not prohibition. Regulators expect an insurer to know what its models do, to test them for discrimination, to oversee outside vendors and to produce documentation on request, for example during an exam. The insurer cannot excuse a result by saying a vendor built the tool. How these rules are enforced is covered in from market conduct exam to consent order, an enforcement timeline. Colorado's general AI statute is a separate law; the guide to AI regulation in insurance: Colorado, the NAIC bulletin and New York explains the interaction, and it recommends confirming how it applies to insurers.

Humans are not a neutral baseline

It is tempting to treat the human underwriter as the fair alternative to the machine. The comparison is less clean. People apply judgment unevenly: two underwriters can rate the same file differently, and a tired or rushed one may lean on rough rules of thumb. Human decisions are also harder to audit, because the reasoning lives in someone's head and in brief file notes. A well-run algorithm applies the same logic to every file and leaves a record of what it did, which is easier to test than a thousand individual judgments.

The trade-off runs both ways. A person can notice that an unusual medical history has a benign explanation, or that a database entry is an obvious error, and a rigid system may not. The practical safeguard is the combination: automated processing for the straightforward files, human review for the unusual ones, and regular testing of both for consistent and justified results. That is also the structure behind most regulators' expectations, which focus on governance and evidence of testing rather than on a particular technology.

For an applicant, the useful habit is to treat the first answer as an answer from one system on one day. If the decision looks wrong and the facts support a better outcome, correct any bad data, ask for a review and consider a different insurer or a fully underwritten application.

What this means for you as an applicant

• You may not be told a model was used. Disclosure rules vary.
• Adverse decisions still carry notice rights. If a consumer report contributed to a decline or higher price, federal notice and dispute rights apply, as described in the guide to how life insurance underwriting works. Your rights under state insurance privacy laws are in privacy and your life insurance data: who sees it and your rights.
• Check what the model read. Request your MIB and prescription reports and dispute errors. The steps are in How to request your own MIB and prescription reports step by step.
• Ask for a person. If an automated process declines you or sets a rate that seems high, ask whether an underwriter can review it, and whether a fully underwritten application with an exam would be considered.
• Compare offers. A different insurer's model, rules and appetite may reach a different result on the same facts.
• Complain when it seems unfair. Your state insurance department takes complaints, and a pattern of them is part of the data regulators use.

When a decision seems wrong, four scenarios where an automated underwriting decision goes wrong walks through four scenarios, and your rights when an insurer takes an adverse underwriting action explains what you are entitled to ask for.

What no one can promise

Algorithms can be faster and more consistent than people, and they can also be wrong in ways that are hard to see. Supporters point to faster decisions and less invasive testing for healthy applicants. Critics point to opacity and the risk of proxies. Regulators are building oversight, and the rules are changing. Check your state department's current position rather than relying on a summary written earlier. A fee-only planner or an agent who works with several insurers can tell you which companies tend to be flexible with which profiles.

Key takeaways

• Algorithmic underwriting combines application answers and outside data with rules or predictive models to approve, request evidence or refer a file to a person.
• Rule-based scorecards can be traced point by point; machine-learning models can be harder to explain and may learn from past decisions.
• Fairness questions center on proxies for protected traits, biased training data, uneven record quality and lack of explanation.
• Regulators generally require governance, testing for unfair discrimination, vendor oversight and documentation rather than banning the tools; Colorado's Regulation 10-1-1 took effect November 14, 2023.
• Applicants can request their reports, ask for human review and complain to the state insurance department.