AI in Insurance Underwriting: How Generative and Agentic AI Are Transforming Risk Assessment

A commercial property submission can arrive with broker emails, ACORD applications, schedules of values, COPE details, inspection reports, loss runs, photographs, and supporting documents. Before an underwriter can determine whether the account fits the insurer’s appetite, the submission must be triaged, its data validated, exposures assessed, and the risk compared against underwriting guidelines and delegated authority. This process is central to sound risk selection, but it can be slow when information is incomplete, inconsistent, or distributed across multiple sources.

Because these activities involve different types of work, insurers are applying different forms of AI to specific stages of the underwriting workflow. Predictive models help identify patterns, score risks, and support pricing and risk-selection decisions. Generative AI extracts information from documents, summarizes loss experience, explains findings, and produces underwriting content. Agentic AI extends these capabilities by coordinating approved tasks, such as checking submission completeness, enriching exposure data, screening an account against appetite, preparing referrals, and routing submissions for quote, decline, or approval.

These technologies serve different purposes. Predictive models support analysis and decision-making. Generative AI helps underwriters interpret information and produce decision-ready content. Agentic AI connects approved workflow steps and determines what should happen next within defined rules, authority limits, and human-approval requirements.

The market is moving in this direction. Accenture’s survey^[1] of 430 senior insurance underwriting executives found that insurers expect AI adoption in underwriting to increase from 14% to 70% over the next three years. It also found that 81% of underwriting executives expect AI and generative AI to create new roles while improving efficiency in operations, risk selection, and decision-making. These findings suggest that insurers are viewing AI as more than a tool for automating individual tasks. They are considering how it can reshape the underwriting operating model.

The article examines how different technologies support specific stages of the underwriting process, from submission intake and data extraction to exposure analysis, appetite screening, pricing preparation, referral management, and documentation. Underwriters retain responsibility for judgment, exceptions, broker relationships, coverage interpretation, consequential decisions, and accountability.

What Is AI in Insurance Underwriting and How Does It Work?

AI helps insurers turn large volumes of underwriting information into clearer risk selection and more consistent decisions. Its value is not limited to automating individual tasks; it can connect data, analytical models, and workflow controls across the underwriting process.

What AI in Underwriting Means

AI in insurance underwriting refers to the use of algorithms, machine learning models, and related technologies to evaluate risk, identify patterns, and support underwriting activities. These AI-driven underwriting systems can process structured and unstructured information, including applications, loss runs, property details, inspection reports, images, and external risk data. Depending on the line of business, AI can extract and validate information, identify adverse underwriting indicators, compare submissions against underwriting appetite, and generate recommendations for review.

Where AI Fits Across the Underwriting Process

AI can support several stages of underwriting, from submission intake to policy issuance. Its role depends on the line of business, the quality of available data, the complexity of the risk, and the level of automation permitted.

Here is how AI supports different stages of underwriting:

1. Data Collection and Enrichment

AI extracts information from applications, emails, PDFs, inspection reports, financial statements, and other documents. It can also enrich submissions using approved internal and external data sources, such as property records, geospatial data, catastrophe information, and industry databases.

This gives underwriters a more complete view of the risk without requiring them to search manually across multiple systems.

2. Data Validation

AI identifies missing information, inconsistent fields, duplicate records, and contradictory data. For example, it can flag differences between an application, a schedule of values, and an inspection report.

Automated validation improves data quality before the underwriter evaluates the submission.

3. Risk Assessment and Scoring

Machine learning models analyze historical and current data to identify patterns, correlations, and emerging red flags, adverse underwriting indicators, exposure characteristics, hazard indicators, or loss drivers. Depending on the line of business, these may include property characteristics, claims history, occupancy, construction type, location, or financial indicators.

The resulting insights can support risk scoring and help underwriters prioritize submissions for closer review.

4. Pricing Recommendations

AI analyzes risk characteristics alongside historical performance, market conditions, customer behavior, and other relevant variables to support pricing recommendations.

When supported by reliable data, appropriate models, and effective governance, this can make pricing more responsive than approaches based entirely on static rules and historical averages.

5. Fraud Detection

AI identifies unusual patterns and inconsistencies across applications, claims histories, financial records, and related data. It can flag duplicate submissions, suspicious changes in information, or discrepancies between declared and externally verified details.

Potentially suspicious cases can then be routed for investigation.

6. Compliance Checks

AI can perform compliance checks against underwriting guidelines, delegated authority limits, and regulatory requirements. These may include GDPR and other data-protection requirements, sanctions screening under OFAC and the UN sanctions lists, Anti-Money Laundering (AML) and Know Your Customer (KYC) requirements, Solvency II, the EU AI Act, and applicable state or national insurance regulations. It can also verify broker licensing, required forms and endorsements, coverage limits, geographic restrictions, and internal risk policies.

AI can identify missing documentation, flag exceptions that require approval, and maintain records of the rules and information considered during the underwriting process.

7. Decision Support and Policy Issuance

Once the information has been collected and assessed, AI can generate an underwriting brief, summarize key exposures, highlight exceptions, and present a risk score or recommendation. In suitable, low-complexity cases, AI can support straight-through processing (STP) by moving a submission through predefined checks and approval rules without manual intervention.

The business impact can be significant. BCG’s research^[2] with US and UK commercial P&C insurers suggests that AI can improve efficiency in complex lines of business by up to 36%, primarily by augmenting manual underwriting. BCG also estimates that better use of data could improve loss ratios by up to three percentage points.

What Is Generative AI in Insurance Underwriting?

Traditional AI is primarily used to predict outcomes, identify patterns, classify risks, and calculate scores. Generative AI adds the ability to interpret and produce content, particularly natural language. Its value in underwriting is not simply that it processes more information. It can work with unfamiliar or inconsistently formatted material and produce new, reviewable prose, rather than only returning predefined fields, scores, or rule-based alerts.

That distinction matters because commercial submissions often contain hundreds of pages of broker correspondence, engineering reports, financial statements, loss runs, policy documents, and other supporting material. Generative AI can interpret these materials, connect information across documents, and create an underwriting narrative that explains what the information means for the risk.

Underwriters do not necessarily need more data. They need unstructured information converted into underwriting intelligence that is relevant, traceable, and ready for evaluation.

I. Generating Risk Narratives

Generative AI combines information from multiple documents and explains the major characteristics of a risk. For example, it may connect a company’s operations, locations, claims history, financial condition, and engineering findings to highlight exposures, loss drivers, unusual conditions, and issues requiring further investigation.

The generative contribution is the creation of an explanatory narrative, not simply the extraction of values into predefined fields.

II. Summarizing Submissions

GenAI summarizes lengthy submissions, loss runs, engineering reports, emails, and supporting documents into a decision-oriented brief. Because it can adapt to different formats and terminology, it can identify and explain potentially material details without relying entirely on a fixed template.

III. Interpreting Unfamiliar Documents

Document extraction is not inherently a generative AI use case. OCR, intelligent document processing, machine learning, and rules engines can already capture information from known document types.

GenAI becomes useful when documents are unfamiliar, inconsistently structured, or contain important information in narrative form. It can interpret the content, propose field values, and explain the associated exposure for underwriter review.

IV. Drafting Controlled Underwriting Language

Within controlled workflows, GenAI drafts policy language, endorsements, broker correspondence, and other documentation using approved wording and underwriting guidelines. It can adapt that language to the facts of a particular risk, while final wording remains subject to version control and human approval.

V. Explaining and Contextualizing Recommendations

Generating a recommendation is not uniquely generative. Predictive models and underwriting rules can already calculate risk scores and apply eligibility criteria.

GenAI adds value by explaining and contextualizing those outputs. It can combine model results, underwriting guidelines, and supporting evidence into a preliminary recommendation that identifies the submission triage, rationale, information gaps, and possible conditions or approvals.

VI. Drafting Context-Specific Communications

GenAI drafts context-specific requests for information, coverage explanations, follow-ups, and other communications. It can explain why additional financial statements, inspections, endorsements, or approvals are required while remaining within filing-approved wording.

VII. Synthesizing Context for Unfamiliar Risks

Generative AI helps underwriters synthesize information about unfamiliar industries, emerging exposures, operational changes, and other risk contexts. Its role is not simply to retrieve background information. It can connect relevant evidence, explain how different factors relate to the risk, identify areas requiring validation, and prepare a structured brief for underwriter evaluation.

Any externally sourced information should remain traceable, current, and subject to the insurer’s research and governance standards. Generative output should support underwriter research, not replace source validation or professional judgment.

McKinsey’s 2025 research^[3] describes this progression clearly: traditional analytical AI identifies patterns, GenAI expands the ability to work with unstructured information, and agentic AI adds greater automation to complex workflows.

The specific value of generative AI in underwriting is therefore its ability to interpret unfamiliar information, connect evidence across sources, and produce context-specific explanations and drafts. It complements, rather than replaces, predictive models, rules engines, templates, and underwriter judgment.

What Is Agentic AI in Insurance Underwriting?

Agentic AI coordinates underwriting activities by assessing the current state of a case, selecting the next permitted action, using approved tools, evaluating the result, and escalating when the case falls outside defined parameters.

Its value is not simply automating intake, data extraction, risk scoring, or policy issuance. Those capabilities are already supported by rules engines, predictive AI, GenAI, and straight-through processing. Agentic AI adds a layer of adaptive orchestration that determines which combination of data sources, underwriting capabilities, controls, and human reviews a particular submission requires.

Where Agentic AI Fits Across the Underwriting Process

In commercial and specialty underwriting, an agentic system may:

• Determine the appropriate underwriting path: Assess the line of business, risk characteristics, jurisdiction, complexity, and authority requirements to select the relevant workflow.
• Coordinate evidence gathering: Identify missing or conflicting information and initiate the appropriate next action, such as requesting documents, retrieving approved data, or seeking broker clarification.
• Sequence underwriting capabilities: Invoke document intelligence, exposure models, pricing tools, appetite rules, sanctions screening, and compliance checks according to the needs of the case.
• Reassess the submission: Respond when new information changes the risk profile, a model produces an unexpected result, or an underwriting rule is triggered.
• Manage referrals: Assemble the relevant evidence, rationale, exceptions, and required approvals before routing the case to the appropriate underwriter.
• Control downstream actions: Initiate permitted documentation or issuance activities only after required conditions and approvals have been satisfied.

Unlike fixed automation, an agentic system does not follow the same sequence for every submission. It evaluates the case state and selects the next permitted step. A complete, straightforward submission may proceed through approved automated checks. Conflicting loss information may trigger further evidence gathering, while a case outside delegated authority may be referred with the relevant rationale and supporting documentation.

Bounded Autonomy and Human Oversight

In underwriting, autonomy must be constrained by authority. An agent may be permitted to request information, retrieve approved data, run models, prepare recommendations, or initiate workflow transitions. Binding coverage, approving material exceptions, or making consequential decisions should remain subject to explicitly defined permissions and human approval.

Key controls include:

• Underwriting appetite and authority boundaries
• Approved tools, data sources, and model versions
• Confidence and completeness thresholds
• Mandatory compliance and sanctions checks
• Escalation rules for unusual or high-severity exposures
• Human approval for material decisions and exceptions
• Audit trails covering inputs, actions, outputs, and approvals

The Strategic Opportunity

For insurers, the opportunity is not just faster document processing. It is a more responsive underwriting operating model in which agents orchestrate submission and eligible workflow activities, while underwriters focus on complex risk selection, broker negotiation, portfolio management, and decisions that require judgment.

The objective is controlled autonomy: greater underwriting capacity and consistency without forcing every submission through the same predetermined process.

How Does Agentic AI Take Generative AI Further?

Generative AI helps underwriters interpret information, summarize submissions, and prepare recommendations. Agentic AI takes this further by using those outputs to determine and coordinate the next permitted actions in the underwriting workflow.

For example, a generative AI tool may summarize a 200-page submission and identify key risk factors. An agentic underwriting system uses that assessment to decide whether the case has sufficient information to proceed, whether additional documents or broker clarification are required, which approved data sources or underwriting models should be consulted, and whether the results fall within the applicable authority and appetite guidelines.

If the submission contains a material information gap, the agent may initiate a request for clarification. If the risk falls outside the insurer’s appetite, it may route the case for referral. If the evidence is complete and the results are within defined thresholds, it may prepare the case for the next approved stage. The system therefore does more than generate content or execute a predetermined sequence. It evaluates the current state of the case, selects the next permitted step, and reassesses the workflow as new information becomes available.

Agentic AI does not replace underwriting judgment. It operates within approved tools, data sources, rules, authority limits, and governance controls, while escalating unusual, uncertain, or consequential decisions to an underwriter.

How Agentic AI Extends Generative AI

In short, GenAI provides interpretation and content generation. Agentic AI adds decision-making and controlled orchestration around those capabilities.

Why Do Human Underwriters Still Matter?

Artificial intelligence helps insurers process large volumes of data quickly and identify risk patterns more efficiently than manual methods. Hence, AI is well suited to low-complexity information-processing tasks. However, underwriting still requires contextual judgment, discretion, relationship management, and accountability. The role of the underwriter is therefore likely to change rather than disappear. As David Swaim, Chief Underwriter at CareScout, puts it, “The underwriter of the future is going to look very different.”

Agentic AI coordinates approved underwriting activities, evaluates intermediate results, and determines the next permitted workflow step. It does not, however, remove the need for human judgment in situations where the available evidence is incomplete, the risk falls outside established parameters, or the decision carries material commercial, regulatory, or reputational consequences.

The role of the underwriter is therefore not to coordinate every routine activity manually. It is to provide judgment where rules, models, and workflow logic cannot fully determine the appropriate outcome.

1. Interpreting Context Beyond the Available Data

Agentic AI operates on the information, rules, models, and tools made available to it. Underwriters assess the context behind that information, including the reliability of the sources, changes in the insured’s circumstances, and factors that may not be represented in historical data.

For example, a change in ownership, management, operations, supply-chain exposure, or risk controls may materially affect a submission even when the available data appears broadly consistent with previous records. Underwriters determine whether those circumstances alter the risk and whether additional investigation is necessary.

2. Exercising Judgment in Ambiguous Situations

Agentic AI may identify conflicting information, apply underwriting guidelines, and recommend a permitted next step. It cannot independently resolve every ambiguity, particularly where competing considerations must be weighed.

Underwriters may need to assess the quality of supporting evidence, interpret an unusual exposure, balance risk against commercial objectives, or determine whether an exception is justified within the insurer’s risk appetite. Their judgment is especially important when the appropriate decision is not directly specified by a rule, threshold, or model output.

3. Handling Complex and Unfamiliar Risks

Agentic workflows are most effective when the available data, decision boundaries, and escalation criteria are sufficiently defined. Complex or emerging risks may not fit those conditions. They may involve limited historical data, conflicting evidence, unusual exposures, new technologies, changing regulation, or circumstances outside the models’ training data.

In these cases, the agentic system should identify the uncertainty, preserve the relevant evidence, and refer the case to an underwriter. The underwriter may request further information, challenge the system’s recommendation, consult specialists, or determine an appropriate course of action.

4. Managing Broker and Stakeholder Relationships

Underwriting decisions often require interaction with brokers, agents, policyholders, claims teams, risk engineers, and other stakeholders. These interactions involve clarification, negotiation, explanation, and trust.

Agentic AI may prepare a submission summary, identify information gaps, or draft a request for clarification. Human underwriters remain responsible for managing sensitive conversations, understanding commercial and customer circumstances, negotiating terms, and explaining why additional information, exclusions, conditions, or approvals are required.

Retaining Accountability for Consequential Decisions

Agentic AI may support decision-making, but accountability for consequential underwriting decisions remains with the insurer and its authorized personnel. Human underwriters must review material recommendations, approve exceptions, and ensure that decisions comply with underwriting policies, delegated authority, and applicable regulatory requirements.

They also have an oversight role in monitoring whether the system is using appropriate data, producing consistent recommendations, identifying potential bias, and operating within approved boundaries. Where the system’s behavior or output is unreliable, the underwriter must be able to challenge, override, or suspend the relevant workflow.

The most effective operating model is therefore not fully autonomous underwriting. It is agentic underwriting with human oversight, in which agents coordinate eligible activities and adapt the workflow within defined limits, while underwriters retain responsibility for context, judgment, complex risks, relationships, and accountability.

I. Contextual Awareness

Underwriters assess more than the data captured in an application. They consider the circumstances behind the information, the reliability of the sources, changes in the insured’s situation, and factors that may not be reflected in historical data. This context helps them interpret AI-generated findings and determine whether additional investigation is necessary.

II. Relationship Management

Underwriters work with brokers, agents, policyholders, claims teams, and other stakeholders to clarify information, explain decisions, negotiate terms, and maintain trust. AI can prepare summaries and draft communications, but human underwriters remain responsible for understanding individual circumstances and managing sensitive interactions.

This human role remains important even as customers become more comfortable using AI-enabled tools. A 2025 Geneva Association survey^[4] of 6,000 insurance customers found that more than 80% were either in favor of or neutral toward insurers using generative AI in customer interactions. As customers increasingly use such tools to compare insurance options, human underwriters will continue to play an important role in guiding decisions, explaining coverage, and communicating the value of insurance products.

III. Judgment and Discretion

AI can identify patterns, apply rules, and generate recommendations. Underwriters must decide how those findings should influence the final decision. They may weigh competing risks, assess the quality of supporting evidence, apply underwriting guidelines, or determine whether an exception is justified within the insurer’s risk appetite. Automated underwriting in insurance can then be trained along these lines to harness the capabilities of AI.

IV. Handling Complex Cases

Straightforward submissions are more suitable for automation. Complex or unusual cases may involve limited data, conflicting information, emerging risks, unusual exposures, or circumstances outside the model’s training data. These cases should be escalated to underwriters, who can request further information, challenge the recommendation, and determine the appropriate course of action.

V. Accountability and Oversight

Human underwriters remain responsible for reviewing consequential recommendations, approving exceptions, and ensuring that decisions comply with underwriting policies and regulatory requirements. They also oversee AI performance, identify potential bias or data-quality issues, and ensure that automated actions remain within approved boundaries.

The most effective model is therefore human-in-the-loop underwriting. AI handles suitable information-processing and workflow tasks, while underwriters retain control over judgment, exceptions, relationships, and accountability.

Conclusion

AI is changing insurance underwriting from a largely manual process into a more connected, data-driven operating model. Traditional AI supports data analysis, generative AI turns unstructured information into underwriting intelligence, and agentic AI in insurance underwriting coordinates approved actions across the workflow.

The strongest model remains human-governed and AI-enabled. Insurers can use AI to improve speed, consistency, scalability, and information processing while underwriters retain responsibility for judgment, complex risks, exceptions, relationships, and accountability.

References:

• 1. Accenture
• 2. BCG
• 3. McKinsey
• 4. Geneva Association