Generative AI in Insurance Underwriting: The 2026 Guide

Submission intake, document extraction, risk summaries, and where it fits, plus limits and governance

Generative AI (gen AI) in insurance underwriting is the use of large language models to read, summarise, and structure the unstructured information in a submission, from applications and loss runs to inspection reports and emails, so underwriters can assess risk faster and more consistently. Its core value is not making the risk decision; it is removing the manual grind around it. The move from proof-of-concept to production is now measurable: Celent’s third annual survey of US property and casualty insurers found generative AI in production rising from 8% in 2023 to 28% in 2024 and 44% in 2025, and 22% of insurers expect an agentic AI solution in place by the end of 2026. In underwriting that increasingly means embedded workflows, often paired with AI agents that triage submissions against appetite before an underwriter opens them.

This guide covers where generative AI fits in the underwriting workflow, what it does well and what it should not do, the benefits, how it works, the risks, the regulatory landscape, and how to implement it.

What Is Gen AI in Insurance Underwriting?

Underwriting AI splits into two complementary roles, and conflating them is the most common source of confusion.

  • Predictive AI (machine learning) scores and prices risk from structured data. This is the decision maths:  loss propensity, severity, and rating 
  • Gene AI  handles language and documents:  extracting data, summarising long submissions, drafting text, and answering natural-language questions 

These two roles work together. Generative AI prepares the inputs and augments the underwriter; predictive models and the underwriter still make the call. An insurer that expects generative AI to decide accept, decline, or price has misunderstood the tool.

Where Gen AI Fits in the Underwriting Workflow

The insurance submission is one of the most document-heavy transactions in any industry. A single large commercial property submission can run to hundreds of pages: ACORD applications, three to five years of loss runs, inspection reports, statements of value, and policy wordings. Gen AI inverts the ratio of reading to deciding.

Workflow step What gen AI does
Submission intake Extracts key data (insured, coverage, limits, deductibles, loss history, building characteristics) from mixed document formats
Data structuring Turns hundreds of pages of unstructured text into structured, comparable fields
Appetite check and triage Checks the submission against guidelines, flags clear declines, and surfaces only in-appetite risks
Risk summarisation Produces a concise risk summary, cross-referencing loss runs and inspection findings
Underwriter support Answers natural-language questions such as “analyse this property for wildfire risk” using submission and reference data
Drafting Drafts quotes, referral notes, and compliance documentation for the underwriter to review

The pattern across every step is the same: gen AI does the reading and preparation so the underwriter spends time on judgement, not rekeying information.

What Gen AI Does Well, and What It Doesn’t

The fastest way to waste an AI underwriting programme is to point it at the wrong tasks. Gen AI is strong at language and documents and weak at deterministic, high-stakes decisions.

Strong fit Poor fit (needs predictive models plus a human)
Extracting data from varied documents The final accept, decline, or price decision
Summarising long, dense submissions Deterministic, auditable rate calculations
Drafting quotes, notes, and correspondence High-stakes decisions without human review
Natural-language search over guidelines and wordings Any task where an unguarded hallucination is unacceptable
Classifying and routing submissions Bias-sensitive decisions without testing and oversight

Keep gen AI in the left column, keep predictive models and underwriters on the right, and most of the risk in an AI underwriting programme disappears.

Benefits: Speed, Capacity, and Fewer Missed Controls

The value concentrates in underwriter capacity. Accenture surveyed almost 500 underwriters and found that 40% of the average underwriter’s time goes on administrative work and a further 30% on negotiation and sales support, leaving roughly 30% for underwriting itself. Accenture puts the resulting efficiency loss at $85 billion to $160 billion over five years. Administrative work is precisely what generative AI absorbs, which is why automating intake and summarisation lifts submission-to-quote rates and reduces premium leakage from missed underwriting controls. 

Translated into operational terms, the benefits are:

  • Speed: Faster submission-to-quote, because intake and triage happen in seconds
  • Capacity: Underwriters handle more submissions without proportional headcount
  • Consistency: Guideline and appetite checks are applied uniformly, reducing missed controls and leakage
  • Focus: Underwriters spend time on in-appetite, well-prepared risks instead of triage

How It Works: RAG and AI Underwriting Agents

Two concepts explain most modern gen AI deployments.

Retrieval-augmented generation (RAG)
Instead of relying on the model’s general training, RAG retrieves relevant, current, proprietary information like your appetite guidelines, prior policies and reference data, and feeds it to the model as context. The answer is grounded in your data, which sharply reduces hallucinations and keeps outputs aligned to your rules.

AI underwriting agents
An agent receives a submission, extracts the key data, checks it against appetite, flags immediate declinations, and pre-populates the file. The underwriter then sees only qualified submissions with the relevant data already in place, rather than a stack of raw documents.

Risks and Limits

The barriers are as much about trust and compliance as technology:

  • Bias and unfair discrimination: Models can inherit bias from historical data; underwriting is exactly where regulators look 
  • Hallucination: Ungrounded generative models can produce confident, wrong output, which is unacceptable in a bind or decline 
  • Black-box explainability: Underwriting decisions must be justifiable; opaque reasoning is hard to defend in an exam 
  • Data quality: Extraction is only as good as the source documents and the reference data behind it 
  • Over-automation: Automating the decision itself, rather than the preparation, invites both regulatory and reputational risk 

Governance and Regulation for AI in Underwriting

Underwriting is the most regulated place to deploy AI, so governance has to be designed in. One point to settle first: these rules govern AI systems in general, not gen AI as a separate category. The NAIC adopted its Model Bulletin, “Use of Artificial Intelligence Systems by Insurers,” on December 4, 2023.  As of the NAIC’s Spring 2026 National Meeting, 24 states and the District of Columbia had adopted it, and four further states had issued their own AI-specific insurance guidance. The bulletin expects a written AI Systems Program covering governance, documentation, third-party oversight, and testing for unfair discrimination, with information available for market-conduct exams.

Two further points bear on underwriting specifically:

  1. Colorado. Under SB 21-169 and Regulation 10-1-1, life insurers using external consumer data and information sources must govern and quantitatively test the algorithms and predictive models built on them, and file an annual attestation; the second round was filed on December 1, 2025. The broader Colorado AI Act (SB 24-205) no longer applies as drafted. SB 26-189, signed on May 14, 2026, repealed and replaced it with a narrower disclosure and transparency regime for automated decision-making technologies, effective January 1, 2027.
  2. EU AI Act. Classifies AI used for risk assessment and pricing in life and health insurance as high-risk. The original August 2, 2026, deadline has moved. Under the Digital Omnibus on AI, agreed in May 2026, obligations for stand-alone high-risk systems in the Annex III use cases now apply from December 2, 2027, and from August 2, 2028, for AI embedded in regulated products. The Act’s transparency duties, including disclosure that a person is interacting with an AI system, took effect on schedule on August 2, 2026. Carriers with EU exposure should plan against the new dates rather than treat the delay as a reprieve, even though the high-risk pricing classification centres on life and health.

The practical requirement is per-decision audit records, explainability, and demonstrable bias testing, all of which is far easier when gen AI prepares inputs and a documented model plus a human makes the decision.

How to Implement Gen AI in Underwriting

  1. Start with intake and extraction. Highest volume, lowest risk, fastest payback, and no automated decisions.
  2. Ground models with RAG. Feed your guidelines and reference data so outputs are accurate and on-policy.
  3. Keep humans on the decision. Automate preparation fully; keep underwriters on accept, decline, and price. 
  4. Embed in the underwriting workbench. Gen AI in the core workflow beats a separate tool for adoption and auditability.
  5. Build governance from day one. Documentation, bias testing, and audit trails are cheaper to design in than to retrofit.
  6. Plan change management. Underwriters need to trust and adopt the workflow, or the capacity gains never materialise.

Common Mistakes

  • Expecting AI to make the risk decision: Gen AI prepares and augments; it does not underwrite
  • Running ungrounded models: Without RAG on your data, outputs drift and hallucinate
  • Bolting on instead of embedding: Standalone tools fragment the workflow and stall adoption
  • Skipping bias and audit controls: Treating governance as a later phase, then failing an audit
  • Ignoring underwriter change management: The technology works but the adoption is where programmes can fail

How Sapiens Fits

The value of gen AI in insurance underwriting depends on where it runs. Embedded in the underwriting workflow of a modern core system, it prepares submissions, structures data, and supports decisions with the traceability that regulators expect. Bolted on separately, it fragments the process and is harder to govern.

At Sapiens for example, Sapiens Life Underwriting  includes gen AI-powered summarisation of documents to support the underwriting risk assessment. For P&C and global specialty, Sapiens Global Specialty Underwriting turns broker emails and attachments into submissions and quotes, then triages them against appetite. 

Above these core capabilities sits the Sapiens Autonomous Insurance Platform (SapiensAIP), launched in September 2026, which enables  agentic tools across underwriting, policy, billing, claims and customer engagement. Its foundation layer is built on Sapiens’ insurance ontology, drawn from more than 40 years of experience, and that is what gives each agent the context to reason like an insurer rather than run on generic prompts.

The payoff comes from embedding gen AI with governance, not from piloting it at the edges. An AI-native core is what makes AI dependable, explainable, and measurable.

FAQ

What is generative AI in insurance underwriting?

It is the use of large language models to read, extract, summarise, and structure the unstructured information in a submission from applications and loss runs to reports, so underwriters can assess risk faster. It prepares and augments underwriting rather than making the risk decision.

How is generative AI different from predictive AI in underwriting?

Predictive AI (machine learning) scores and prices risk from structured data: the decision maths. GenAI handles language and documents: extracting, summarising, and drafting. They are complementary: generative AI prepares the inputs, predictive models and underwriters make the call.

What underwriting tasks can generative AI do?

Submission intake and data extraction, structuring unstructured documents, appetite and guideline checks, risk summarisation, natural-language querying of submissions and guidelines, and drafting quotes, referral notes, and compliance documentation.

Does generative AI replace underwriters?

No. It removes administrative and document work so underwriters handle more submissions and focus on judgement. The accept, decline, and price decision stays with the underwriter, supported by predictive models.

What is RAG in underwriting?

Retrieval-augmented generation(RAG) feeds the model relevant, current, proprietary data, such as appetite guidelines and prior policies, as context, so answers are grounded in your data and rules. This sharply reduces hallucinations.

Is generative AI underwriting regulated?

Yes, though not by gen AI-specific rules. Gen AI is regulated as one class of AI system, so the general AI obligations apply to it in full. The NAIC Model Bulletin on the Use of AI Systems by Insurers, adopted in December 2023 and now in force in 24 states and the District of Columbia, requires governance, documentation, and bias testing. Colorado’s Regulation 10-1-1 and the EU AI Act add further requirements for underwriting and pricing. 

How do insurers start with generative AI in underwriting?

Begin with high-volume, low-risk tasks like submission intake and document extraction, ground models with RAG on your guidelines, keep humans on the decision, embed AI in the underwriting workbench, and build governance from the start.

Explore More