AI Insurance Underwriting: Can Insurers Use Algorithms to Assess Risk?
Quick Answer: Yes. Insurers can use artificial intelligence, machine learning and predictive analytics to support underwriting and risk assessment, subject to the laws and regulations applicable to the relevant insurance product and jurisdiction. The principal legal concerns include unfair discrimination, inappropriate use of data, proxy variables, explainability, privacy, actuarial justification and regulatory oversight.
Insurance has always been about prediction.
An insurer asks:
How likely is this risk to produce a claim?
Historically, insurers relied on:
- Actuarial tables.
- Historical claims.
- Demographic information.
- Property characteristics.
- Medical information.
- Driving history.
Artificial intelligence changes the scale.
An AI system can analyse millions of records.
It can identify relationships between variables.
It can continuously update predictions.
It can potentially assess an application in seconds.
This creates an obvious commercial attraction.
But insurance underwriting is not simply a technical prediction problem.
An insurer's risk model can affect:
- Whether a person receives coverage.
- What coverage is offered.
- How much the policy costs.
- What exclusions apply.
- What underwriting information is requested.
That means AI underwriting can directly affect consumers.
And once an algorithm makes a consumer-facing decision, a new question arises:
When does lawful risk classification become unlawful discrimination?
That question is complicated because insurance law historically permits risk classification in many circumstances.
Insurers are not expected to treat every risk identically.
The entire business model depends upon differentiating risks.
The legal issue is therefore not:
โCan insurers differentiate between risks?โ
It is:
โWhich distinctions are legally permissible, actuarially justified and appropriately governed?โ
Legal disclaimer: This article provides general educational information and is not legal, insurance, actuarial, financial, privacy or regulatory advice. Insurance regulation varies significantly by product, state and circumstances.
Key Takeaways
- AI can automate and enhance insurance underwriting.
- Machine learning can analyse substantially more variables than traditional manual underwriting.
- AI underwriting does not eliminate insurers' existing regulatory obligations.
- Risk classification and unlawful discrimination are not the same concept.
- Proxy variables can create discrimination concerns even when protected characteristics are not explicitly used.
- Alternative data can increase both predictive power and regulatory risk.
- Insurers should understand the purpose and relevance of each material underwriting variable.
- AI models should be validated before and after deployment.
- High-impact underwriting decisions require appropriate governance and oversight.
- Third-party underwriting models should be subject to vendor due diligence.
- Explainability becomes particularly important when AI affects eligibility or material policy terms.
- State insurance regulators remain central to U.S. insurance AI oversight.
What Is Insurance Underwriting?
Quick Answer: Insurance underwriting is the process through which an insurer evaluates a risk and determines whether and on what terms to provide insurance.
Depending on the insurance product, underwriting can influence:
- Eligibility.
- Premium.
- Coverage.
- Limits.
- Deductibles.
- Exclusions.
What Is AI Insurance Underwriting?
Quick Answer: AI insurance underwriting uses artificial intelligence, machine learning, predictive analytics or related automated systems to evaluate insurance risk.
A simplified process is:
Application Data โ AI Model โ Risk Assessment โ Underwriting Decision โ Policy Terms.
The model may analyse:
- Historical claims.
- Applicant characteristics.
- Property information.
- Driving behaviour.
- Medical information, where lawfully relevant.
- External datasets.
How Is AI Underwriting Different From Traditional Underwriting?
Quick Answer: Traditional underwriting often relies on predefined rules and actuarial models, while AI systems can identify complex relationships in large datasets and generate predictive scores.
Traditional system:
If X + Y + Z โ Risk Category A.
Machine-learning system:
Thousands of variables โ learned relationship โ predicted risk.
The second approach can be more flexible.
It can also be harder to explain.
What Is Automated Underwriting?
Quick Answer: Automated underwriting uses software to evaluate applications and make or support underwriting decisions without requiring every application to undergo manual assessment.
AI can make automated underwriting more sophisticated by allowing systems to process:
- Unstructured information.
- Historical claims.
- External datasets.
- Behavioural information.
What Is Predictive Underwriting?
Quick Answer: Predictive underwriting uses statistical or machine-learning techniques to estimate the probability or expected severity of future insurance losses.
For example:
Probability of claim = predicted risk.
But:
Predicted risk โ automatic legal entitlement to charge any price.
The prediction must operate within the applicable insurance-regulatory framework.
What Is Algorithmic Risk Assessment?
Quick Answer: Algorithmic risk assessment uses mathematical or computational models to classify or score an applicant's expected insurance risk.
A model might produce:
Risk Score: 782
The important question is:
What does 782 actually mean?
Is it:
- Expected claim probability?
- Expected loss?
- Relative ranking?
- Underwriting category?
Insurers should be able to define the meaning of the score.
Can AI Underwriting Be More Accurate?
Quick Answer: AI can improve predictive performance for particular underwriting tasks, but greater predictive accuracy does not automatically establish legal or actuarial acceptability.
A model may be extremely good at predicting:
Who will generate a claim.
That does not automatically answer:
Whether the variables used to make that prediction may lawfully be used.
What Is Alternative Data in Insurance Underwriting?
Quick Answer: Alternative data refers to information outside traditional insurance datasets that may be used to assess risk.
Potential examples include:
- Consumer behaviour.
- Digital activity.
- Property data.
- Telematics.
- Public records.
- Transaction information.
The availability of alternative data creates substantial opportunities and risks.
Can Insurers Use Social Media Data for Underwriting?
Quick Answer: Whether social-media information can lawfully be used depends on the insurance product, jurisdiction, data source, purpose and applicable law.
Insurers should not assume:
Publicly visible = unrestricted underwriting use.
Privacy, consumer-protection, discrimination and insurance regulations may still be relevant.
Can AI Use Location Data for Underwriting?
Quick Answer: Location information can be relevant to certain insurance risks, but it can also correlate with protected characteristics or socioeconomic circumstances.
Insurers should therefore ask:
- Why is location relevant?
- What risk does it measure?
- What other characteristics does it correlate with?
- What consumer outcomes does it produce?
What Is Proxy Discrimination in Insurance?
Quick Answer: Proxy discrimination can arise when a seemingly neutral variable indirectly captures information associated with a protected characteristic.
For example:
Protected characteristic โ geographic pattern โ model variable โ underwriting outcome.
The model may never explicitly receive the protected characteristic.
Yet the relationship can still create discrimination concerns.
Does Removing Race Eliminate AI Underwriting Bias?
Quick Answer: No.
Removing race from the dataset does not necessarily eliminate variables correlated with race.
Potential proxies can include:
- Geography.
- Consumer behaviour.
- Income-related information.
- Property characteristics.
- Healthcare utilisation.
Can AI Underwriting Discriminate?
Quick Answer: Potentially.
AI underwriting can produce discriminatory outcomes through:
- Biased training data.
- Proxy variables.
- Historical underwriting practices.
- Inappropriate target variables.
- Unequal error rates.
- Model design.
However, not every difference between groups constitutes unlawful discrimination.
The applicable insurance and civil-rights law must be examined.
What Is Historical Bias in Insurance Underwriting?
Quick Answer: Historical bias occurs when a model learns patterns from previous underwriting decisions that may reflect outdated assumptions, unequal treatment or structural differences.
The model effectively learns:
โThis is how insurers classified risk in the past.โ
That does not necessarily mean:
โThis is how risk should be classified today.โ
Can AI Reproduce Historical Underwriting Practices?
Quick Answer: Yes.
Consider:
Historical decisions โ Training data โ AI model โ New underwriting decisions.
If historical practices contained problematic patterns, AI can scale them.
What Is Actuarial Fairness?
Quick Answer: Actuarial fairness generally concerns whether risk classification and pricing appropriately reflect expected losses and comply with applicable insurance principles and law.
Actuarial fairness is not necessarily identical to:
Equality of outcomes.
Insurance inherently differentiates among risks.
The legal question is whether the differentiation is permitted and appropriately justified.
Why Is Insurance Different From Ordinary Consumer Pricing?
Quick Answer: Insurance is fundamentally based on risk pooling and classification.
An insurer may need to distinguish:
- Higher-risk drivers.
- Lower-risk drivers.
- Higher-risk properties.
- Lower-risk properties.
Therefore, simply observing different premiums does not establish discrimination.
The relevant question is:
Why are the risks being classified differently, and is that distinction legally permissible?
Can AI Underwriting Use Protected Characteristics?
Quick Answer: The answer depends heavily on the type of insurance and applicable law.
Some characteristics may be legally restricted in particular insurance contexts.
Others may be relevant under specific regulatory frameworks.
Insurers should therefore avoid treating all insurance products as legally identical.
Why Does State Law Matter?
Quick Answer: Insurance regulation in the United States is heavily state-based.
Different states can impose different restrictions on:
- Underwriting.
- Rating.
- Data use.
- Discrimination.
- Consumer disclosures.
Therefore:
AI underwriting compliance is not simply a federal question.
What Is the NAIC AI Model Bulletin?
Quick Answer: The National Association of Insurance Commissioners issued a model bulletin concerning the use of artificial intelligence systems by insurers.
The bulletin addresses governance and risk-management expectations surrounding AI systems used by insurers.
It emphasises that insurers should maintain a governance framework appropriate to the risks associated with AI use.
The NAIC framework is particularly important because state insurance regulators can use it as a reference point in supervising insurer AI practices.
Does the NAIC Regulate Insurers Directly?
Quick Answer: The NAIC is a standard-setting and coordinating organisation rather than a federal insurance regulator.
Actual regulatory authority generally rests with state insurance regulators.
Therefore, the legal effect of NAIC guidance depends on how individual states adopt, reference or implement relevant principles.
What Is AI Governance in Insurance?
Quick Answer: AI governance is the organisational framework used to control how AI systems are developed, deployed, monitored and retired.
It should cover:
- Model approval.
- Data governance.
- Risk assessment.
- Validation.
- Bias testing.
- Human oversight.
- Vendor management.
- Incident response.
Should Insurers Have an AI Inventory?
Quick Answer: Yes, maintaining an inventory of material AI systems is an important governance practice.
The inventory should identify:
- Model owner.
- Business purpose.
- Data used.
- Consumer impact.
- Vendor.
- Validation date.
- Risk classification.
What Is Model Validation?
Quick Answer: Model validation evaluates whether a model performs as intended for its stated purpose.
Validation can assess:
- Predictive accuracy.
- Stability.
- Error rates.
- Data quality.
- Population performance.
- Model assumptions.
Why Is Model Drift Important?
Quick Answer: Model drift occurs when changes in the underlying environment reduce the accuracy or appropriateness of a model.
Insurance markets change.
Consumer behaviour changes.
Claims patterns change.
Climate conditions change.
Medical practices change.
A model that worked well five years ago may not perform identically today.
What Is Explainable Underwriting?
Quick Answer: Explainable underwriting means that the insurer can identify and communicate the principal factors that materially influenced an AI-supported underwriting outcome.
Consider two outputs.
Output A:
โRisk score: 811.โ
Output B:
โThe model classified the application as high risk primarily because of factors A, B and C.โ
Output B provides significantly more information for human review.
Why Does Explainability Matter?
Quick Answer: Explainability can help insurers identify model errors, support internal review and respond to regulatory or consumer inquiries.
It can also help determine whether a seemingly neutral variable is producing an unexpected outcome.
Can AI Underwriting Be a Black Box?
Quick Answer: Complex machine-learning models can be difficult to interpret, but complexity does not eliminate the need for governance.
An insurer should understand enough about a material model to determine:
- What it does.
- What data it uses.
- What its limitations are.
- What risks it creates.
Can Third-Party Vendors Perform AI Underwriting?
Quick Answer: Insurers can use third-party technology providers for underwriting-related functions, subject to applicable legal, contractual and regulatory requirements.
Vendor due diligence should examine:
- Model methodology.
- Training data.
- Validation.
- Security.
- Bias testing.
- Change management.
- Audit rights.
Can Insurers Blame AI Vendors for Bad Underwriting?
Quick Answer: An insurer should not assume that outsourcing a model eliminates its own regulatory responsibilities.
The insurer should understand the model sufficiently to exercise appropriate oversight.
Contracts should also address:
- Data ownership.
- Permitted data use.
- Model changes.
- Regulatory cooperation.
- Audit rights.
Can AI Use Medical Data for Insurance Underwriting?
Quick Answer: Whether medical information may be used depends heavily on the insurance product, jurisdiction, consent requirements, privacy law and insurance regulations.
Insurers should not assume that access to health information automatically authorises any underwriting use.
Can AI Underwriting Use Consumer Data?
Quick Answer: Potentially, but the insurer must assess whether the data source and use comply with applicable insurance, privacy and consumer-protection requirements.
Questions include:
- Where did the information come from?
- Was it accurate?
- Was the consumer aware of its use?
- Is the data relevant to the risk?
- Is the use legally permitted?
What Is Data Relevance in Underwriting?
Quick Answer: Data relevance concerns whether a variable has a legitimate relationship to the risk the insurer is attempting to assess.
This is particularly important with alternative data.
The fact that a variable improves prediction does not automatically answer:
Should the insurer use it?
Can a More Accurate AI Model Be Less Fair?
Quick Answer: Yes.
A model can improve overall predictive accuracy while producing worse outcomes for a particular population.
This is why insurers should evaluate:
- Overall accuracy.
- Group-level performance.
- Error rates.
- Consumer outcomes.
AI Insurance Underwriting Risk Matrix
| Risk | Example | Safeguard |
|---|---|---|
| Proxy discrimination | Geography indirectly reflects protected characteristics | Proxy-variable testing |
| Historical bias | Past underwriting patterns reproduced | Training-data review |
| Black-box model | Reason for classification unclear | Explainability |
| Model drift | Risk patterns change | Continuous validation |
| Bad data | Incorrect consumer information | Data-quality controls |
| Vendor risk | External model inadequately governed | Vendor due diligence |
| Privacy risk | Unnecessary personal data used | Data minimisation |
| Automation bias | Underwriter blindly accepts AI output | Human oversight |
AI Insurance Underwriting Compliance Checklist
- Create an inventory of AI underwriting systems.
- Identify the insurance product involved.
- Identify the applicable state regulatory framework.
- Document the model's purpose.
- Document every material input variable.
- Assess the relevance of each variable.
- Identify potential proxy variables.
- Review training-data quality.
- Validate predictive performance.
- Test group-level outcomes where appropriate.
- Review consumer-protection implications.
- Review privacy requirements.
- Establish human-review procedures.
- Document model changes.
- Monitor model drift.
- Conduct vendor due diligence.
- Maintain audit records.
- Establish complaint-handling procedures.
- Review regulatory developments.
- Revalidate material models periodically.
Frequently Asked Questions
Can insurers use AI for underwriting?
Yes. AI can support or automate aspects of insurance underwriting, subject to applicable insurance, consumer-protection, privacy and anti-discrimination requirements.
What is AI insurance underwriting?
It is the use of artificial intelligence, machine learning or predictive analytics to evaluate insurance risk and support underwriting decisions.
Can AI decide whether someone gets insurance?
AI can potentially support eligibility decisions, but the legal permissibility and required oversight depend on the insurance product and jurisdiction.
Can AI underwriting discriminate?
Potentially. Bias can arise through training data, proxy variables, historical practices and model design.
Does removing race from an AI model eliminate discrimination?
No. Other variables may act as proxies for race or other protected characteristics.
Can insurers use alternative data?
Potentially, but the legality and appropriateness of alternative data depend on the type of insurance, source, purpose and applicable law.
Can insurers use social-media data for underwriting?
Whether such information may be used depends on the relevant insurance product, jurisdiction, source and applicable legal requirements.
What is algorithmic underwriting?
Algorithmic underwriting uses computational models to classify or score insurance risks.
What is predictive underwriting?
Predictive underwriting uses statistical or machine-learning models to estimate future insurance risk or expected losses.
What is model drift?
Model drift occurs when changes in the underlying environment cause a model's performance or relevance to deteriorate.
Why does explainability matter in AI underwriting?
Explainability can help insurers understand, review and govern the factors influencing an automated or AI-supported underwriting decision.
Can a third-party vendor provide AI underwriting?
Yes, but insurers should conduct appropriate vendor due diligence and maintain oversight of material outsourced AI functions.
Is AI underwriting regulated by the federal government?
Some federal laws may apply depending on the insurance product and conduct, but insurance regulation in the United States is substantially state-based.
Does the NAIC regulate insurance AI?
The NAIC develops model laws, bulletins and regulatory guidance, while state insurance departments exercise actual regulatory authority.
Conclusion
Insurance underwriting has always been a prediction exercise.
AI simply makes that prediction more powerful.
An insurer can now potentially analyse millions of data points and identify relationships that would be difficult to detect through traditional underwriting processes.
That can create substantial benefits.
Applications can be processed faster.
Routine underwriting can become more efficient.
Risk assessment can become more granular.
Fraudulent or inaccurate information may be easier to detect.
But predictive power creates a regulatory challenge.
Insurance is not simply a data-science exercise.
It is a regulated mechanism for allocating risk.
The fact that an algorithm can identify a correlation does not automatically mean that an insurer should use that correlation to determine a consumer's insurance outcome.
The key questions are:
Is the variable relevant?
Is the data reliable?
Is the use lawful?
Could the variable operate as a proxy?
Does the model produce materially different outcomes across populations?
Can the insurer explain the result?
Can the insurer demonstrate appropriate governance?
These questions become particularly important when insurers move beyond traditional actuarial variables and begin using alternative data.
Alternative data can make an algorithm more predictive.
It can also make the model more difficult to regulate.
A variable may have a strong statistical relationship with claims without having a sufficiently clear legal or actuarial justification for use.
That is why:
Predictive relevance โ automatic regulatory permission.
AI underwriting should therefore be built around a governance framework.
The insurer should know:
- What the model does.
- What data it uses.
- Why the data matters.
- What risks the model creates.
- How the model was validated.
- How consumers are affected.
Third-party vendors should be subject to the same scrutiny.
An insurer should not be able to say:
โThe vendor's algorithm made the decision.โ
Outsourcing technology does not automatically outsource regulatory responsibility.
Human oversight also remains important.
Not every application needs a human to manually examine every data point.
But high-impact or unusual cases should have appropriate mechanisms for review, escalation and correction.
The ideal model is therefore not:
Human underwriting versus AI underwriting.
It is:
AI-assisted underwriting within an accountable governance framework.
The future of insurance underwriting will almost certainly involve more artificial intelligence.
The regulatory challenge will be ensuring that technological sophistication does not outpace legal accountability.
The fundamental principle is straightforward:
An insurer may use technology to predict risk, but the insurer remains responsible for how that prediction is converted into an insurance decision.
Legal Disclaimer
This article is provided for general educational and informational purposes only. It is not legal, insurance, actuarial, financial, privacy or regulatory advice and does not create an attorney-client relationship. Insurance underwriting rules vary according to the insurance product, jurisdiction and applicable federal and state law.