Homeowners underwriting has always been a document-heavy business. Applications, prior carrier loss runs, inspection reports, agent correspondence, roof condition assessments, contractor invoices. An underwriter's day is largely spent reading unstructured text and turning it into a decision. That happens to be exactly the work large language models do well, which is why LLM adoption in underwriting has moved faster than almost any prior wave of insurance technology. Conning's 2024 executive survey found 69% of respondents piloting LLMs in sales and underwriting, and its 2025 follow-up showed early or full LLM adoption climbing from 18% to 63% in a single year. The NAIC's Home Insurance AI/ML Survey tells the same story: 70% of the 194 responding homeowners insurers use, plan to use, or plan to explore AI/ML in their operations. For most carriers the question is no longer whether to deploy, but where the technology actually earns its keep.
Submission intake and triage
The clearest near-term win is at the front door. LLMs can read an incoming submission (ACORD forms, agent emails, attached PDFs), extract the relevant fields, normalize them, and pre-populate the policy administration system. They handle the messiness that defeated earlier rules-based tools: a handwritten note about a wood-burning stove, an agent's email mentioning a trampoline, a prior carrier's nonrenewal letter buried on page forty. And the value goes beyond rekeying. Models can route clean, in-appetite risks toward straight-through processing, flag borderline files for an underwriter, and surface obvious declines early. The payoff is faster quote turnaround, and underwriters who spend their time on judgment rather than data entry.
Synthesizing inspection and imagery data
Computer vision, not the LLM itself, scores the roof from aerial imagery. But the output of those tools (condition scores, hazard flags, detected pools or outbuildings) lands on an underwriter's desk alongside a 30-page inspection report and the insured's application. The synthesis step is where LLMs shine. They can reconcile what the imagery vendor flagged against what the inspector wrote and what the applicant disclosed, then produce a concise risk narrative with discrepancies highlighted. An undisclosed pool. A roof age that doesn't match the application. Deferred maintenance the inspector noted but the aerial view missed. These contradictions are where losses and disputes originate, and they are easy to overlook when the underwriter is reading under time pressure.
Guideline navigation and referral support
Underwriting manuals run hundreds of pages and change constantly. Retrieval-augmented LLM assistants let underwriters ask plain-language questions like "What's our position on knob-and-tube wiring in a partial renovation?" and get an answer grounded in the current manual, with the source section cited. The same capability supports newer staff handling referrals: the model can draft the referral memo, summarize the file, and identify which guideline provision is actually in question. Carriers report this shortens training curves and improves consistency across desks.
Renewal monitoring and correspondence
At renewal, LLMs can compare the current file against new imagery-derived data and claims activity, then draft the appropriate output: a risk-improvement letter asking the insured to address debris or roof wear, a conditional renewal notice, or a clean pass. Drafting policyholder and agent correspondence that is clear, jurisdiction-appropriate, and consistent in tone is an underrated use case. It saves meaningful time across a book.
Keeping the human in the loop
None of this works without governance, and regulators have been explicit on that point. The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, now adopted by roughly half the states, makes clear that existing unfair discrimination and trade practice laws apply fully to AI-supported decisions. Carriers should expect to document governance, testing, and third-party oversight in examinations. States are also moving on imagery-specific rules. California's pending AB 1559, for example, would require insurers to share aerial images relied on for adverse actions and give policyholders a chance to dispute or remediate.
The practical implication: LLMs should draft, summarize, extract, and flag, and underwriters should decide. Adverse actions need a documented, explainable basis that a human has reviewed. Models hallucinate, OCR misreads, and imagery goes stale. Workflows must assume imperfection and route exceptions to people.
Carriers getting this right are not replacing underwriters. They are removing the six thousand pages a week that stood between underwriters and underwriting. In a line of business where loss costs, catastrophe exposure, and reinsurance pressure leave little margin for slow or inconsistent risk selection, that is a competitive advantage worth building carefully.
References
- Conning, "Transformative AI Technology: Insights from Conning's Executive Survey" (2024), as reported by Risk & Insurance: https://riskandinsurance.com/insurance-industry-increasingly-adopting-ai-technologies-study-shows/
- Conning, "2025 AI in Insurance: The C-Suite Verdict" (3rd annual survey): https://www.conning.com/insurance-expertise/purchase-reports/article/2025%20AI%20in%20Insurance%20The%20CSuite%20Verdict%20%20How%20the%20Insurance%20Industry%20is%20Adopting%20AI/TIFS0625
- NAIC, Artificial Intelligence topic page (includes Home Insurance AI/ML Survey results, August 2023, and Model Bulletin background): https://content.naic.org/insurance-topics/artificial-intelligence
- NAIC, state adoption map for the Model Bulletin on the Use of Artificial Intelligence Systems by Insurers: https://content.naic.org/sites/default/files/cmte-h-big-data-artificial-intelligence-wg-map-ai-model-bulletin.pdf
- Troutman Pepper Locke, "Using Aerial Imagery in Insurance and Related AI: Emerging Regulatory Themes" (March 2026, discusses California AB 1559): https://www.troutman.com/insights/using-aerial-imagery-in-insurance-and-related-ai-emerging-regulatory-themes/