AI Underwriting and Claims Automation: Past the Pilot

Claims and underwriting automation are in production at some insurers, and some of those deployments predate the current wave of generative AI: Lemonade said in December 2016 that its claims bot had paid a claim in three seconds. Companies Gallagher Re classes as AI-focused took the large majority of insurtech funding in the first three quarters of 2026. In practice, AI in insurance underwriting means models that read a submission, pull out the data, and price the risk or route it to an underwriter.

The named deployments offer concrete operating results. In August 2024, Hiscox's London Market business launched a Google Gemini model for US and Canadian sabotage and terrorism renewals, at first excluding the New York and Chicago metropolitan areas. It reads the broker's submission and drafts a priced quote for an underwriter to review. Hiscox's 2023 proof of concept described pulling the data out of those emails by hand as a job that could take up to three days; Hiscox says the quote now goes back within minutes. Lemonade says it automates more than half of its claims, and puts its loss-adjustment expense ratio at about 5%, against roughly 9% at many large incumbents.

Adoption beyond the leaders is harder to pin down. In a survey Celent published in May 2025, 22% of participating insurers said they planned to have an agentic AI system in place by the end of 2026. That survey measures plans for agentic AI as reported in 2025; it doesn't show how much of the market already runs other kinds of AI or automation. My read is that the leaders named here are past the pilot.

These deployments show faster quoting and cheaper claims handling. I think they point to something bigger than productivity. If most decisions can go through without a person touching them, the cost of writing a policy could fall far enough that risks previously too small to cover profitably become viable, and that would eventually change what insurance products exist.

Key takeaways

• Hiscox's London Market business launched a Gemini-assisted underwriting system in August 2024 for US and Canadian sabotage and terrorism renewals, at first excluding the New York and Chicago metropolitan areas. Hiscox reports quotes within minutes; its proof of concept described manual data extraction taking up to three days.
• In a Celent survey published in May 2025, 22% of participating insurers said they planned to have an agentic AI system in place by the end of 2026. That measures plans for one kind of AI, not what is already in production.
• As of November 3, 2026, 27 states, Washington, D.C. and Puerto Rico had adopted the NAIC's model bulletin on insurers' use of AI, and New York expects insurers to give specific reasons for adverse underwriting or pricing decisions without citing a vendor's proprietary algorithm.

Why is AI in insurance underwriting hard to sell?

AI underwriting vendors are hard to tell apart because so many of them cite similar numbers, and carriers know it.

The buyer is an insurance carrier: regulated, actuarially conservative, and institutionally allergic to unexplained decisions. Three specific objections come up in nearly every evaluation, and vendors who haven't prepared for them lose slowly.

Explainability. When a model helps decline a policy or a claim, the insurer still has to explain the decision under the rules that apply, and that explanation may have to survive a regulator, a complaint or a court. "The model scored it" is not an answer, so vendors need to show how their system supports the insurer's explanation. New York's insurance regulator spelled this out in July 2024: insurers that use AI or outside consumer data in underwriting and pricing are expected to give applicants the specific reasons for an adverse decision, and they can't cite a vendor's proprietary algorithm as a reason not to. Vendors who treat this as a compliance checkbox rather than a core product requirement get filtered out.

Adverse selection risk. If a faster model systematically underprices certain risks, you don't get a little bit of bad business. You attract the risks competitors correctly priced higher. Carriers understand this deeply, and it makes them appropriately paranoid about accuracy claims.

Bias and fair-pricing exposure. Underwriting and claims decisions are among the most regulated in the sector. As of November 3, 2026, the NAIC lists 27 states, Washington, D.C. and Puerto Rico as having adopted its model bulletin on insurers' use of AI. The bulletin is supervisory guidance rather than a law: it expects insurers using AI in decisions that affect consumers to keep a written, risk-based AI program and to oversee the systems they buy from vendors. Colorado goes further: life, private auto and health insurers that use outside consumer data and predictive models there must run a governance and risk-management framework built to detect unfair discrimination.

The courts are involved too. A lawsuit filed in federal court in Illinois in 2022 alleges that State Farm's algorithms for screening homeowners' claims subject Black policyholders to extra scrutiny. State Farm said at the time that the suit "does not reflect the values we hold." In September 2023 the court let one Fair Housing Act claim proceed, and a case tracker reported discovery still under way as of February 2026; there has been no ruling on the merits. Demonstrating that a model doesn't produce unfairly discriminatory outcomes is a substantial ongoing burden.

To me, a vendor whose materials don't address these three is telling carriers it hasn't met a real carrier yet.

What persuades a carrier to buy AI underwriting software?

My view is that carriers buying AI underwriting software are persuaded more by evidence they can check for themselves than by accuracy claims.

Performance at a named carrier over a meaningful period. Not a benchmark. A production deployment with a loss-ratio outcome and a customer willing to be identified. This is the hardest asset to obtain in insurance because carriers are unusually private, and that's precisely why it differentiates.

Explainability demonstrated, not asserted. Showing the decision trail for a declined case is worth more than any accuracy statistic.

Regulatory posture stated plainly. Which jurisdictions, which filings, how model changes are governed and documented. Carriers evaluate this early and quietly, and the map keeps getting more complicated. The EU's AI Act classes AI used for risk assessment and pricing in life and health insurance as high-risk; an amendment that took effect this July moved the start of those obligations to December 2, 2027.

Honesty about where the model shouldn't be used. The most automated operations named here still keep people in the loop. Hiscox's model drafts the quote and an underwriter reviews it. Lemonade automates more than half its claims, which still leaves a large share for people. Vendors who describe which cases go to humans, and why, are trusted more than vendors implying full automation. In my view, a declined claim needs meaningful human review.

What is the risk of automating insurance claims?

Underwriting decisions carry reputational risk too, but I think claims automation carries the sharpest, because an automated claims decision determines whether a customer gets paid or gets care.

A claim is the moment a customer discovers whether their insurance was real. Automating it well is mostly invisible; automating it badly can bring complaints, lawsuits, bad press and regulatory scrutiny.

The best-known case is UnitedHealth. In 2023, families of Medicare Advantage patients sued, alleging the company used nH Predict, an algorithm from its naviHealth unit, to cut off post-acute care that patients' doctors said they needed. UnitedHealth says the tool is a guide and isn't used to make coverage decisions. In February 2025 a federal judge let two of the seven claims go forward, and the case is still in discovery. Separately, a Senate subcommittee staff report in October 2024 found that UnitedHealthcare's prior-authorization denial rate for post-acute care rose from 10.9% in 2020 to 22.7% in 2022, while the company was rolling out several efforts to automate that review. UnitedHealthcare said the report mischaracterized its clinical practices.

In my view that asymmetry makes the communications risk profile different: the upside of a claims automation story is modest and the downside is severe.

The vendors handling this well talk about speed for straightforward claims and human handling for complex or contested ones, and they say so publicly. Allianz's agentic claims system in Australia cut processing time on food-spoilage claims under AUD 500 by about 80%, from several days to a day or less. Allianz's description of it is just as clear about the limit: "By design, payout decisions are never automated." It's a less impressive claim and a safer one, as long as the human review is real: the same Senate report warned that reviewers may face pressure to rubber-stamp the recommendations of algorithms.

The story worth telling

The story I'd want an AI underwriting vendor to tell isn't accuracy percentages. It's what could become insurable if the cost of underwriting keeps falling.

Small policies, short-duration coverage, risks in markets that were never economical to serve. It's explainable to a general business audience, and I rarely see vendors telling it; most compete on the same few metrics. A vendor that can show those economics with a carrier's own numbers has a concrete story about which customers or risks that carrier can now serve profitably. Continuous risk monitoring is also changing policy terms, as in builders' risk on construction sites.

Common questions

How is AI used in insurance underwriting? To read submissions, pull out the data and draft a price or a decision. Hiscox uses a generative AI model to draft priced quotes on sabotage and terrorism renewals for underwriter review; it says manual data extraction could take up to three days and quotes now take minutes. Wider adoption is harder to measure: in a Celent survey published in May 2025, 22% of participating insurers planned to have an agentic AI system by the end of 2026.

Will AI replace insurance underwriters? These examples keep human review in specific workflows; they don't show what happens to underwriter employment. Hiscox's model drafts the quote and an underwriter reviews it. What moved to the model was pulling data out of broker emails. In claims, Allianz's agentic system in Australia checks coverage, weather and fraud on small food-spoilage claims, but a person makes the payment decision.

Are insurers actually deploying AI in claims? Yes, at least at some insurers. Lemonade says it automates more than half of its claims. Allianz's agentic system in Australia cut processing time on small food-spoilage claims by about 80%, with a person still approving every payout.

What should AI underwriting vendors prepare for a carrier's review? I'd prepare examples showing how the software supports decision explanations, pricing validation, testing for unfairly discriminatory outcomes and governance reviews. Which duties apply depends on the jurisdiction and the use case; the NAIC's AI model bulletin and New York's 2024 circular letter describe what supervisors expect.

What's the risk of automating claims decisions? A claim is where a customer learns whether their coverage was real. UnitedHealth faces a lawsuit, still in discovery, alleging an algorithm was used to cut off Medicare Advantage patients' post-acute care, which the company disputes. Credible deployments keep complex or contested claims, and often the final payout decision, with people.

What does faster underwriting actually enable? Less manual work and faster quotes, first. Over time, I'd argue it could make smaller and shorter-duration risks economical to write, which would change what insurance products exist rather than only how quickly they're issued. That case still has to be proven with carriers' own numbers.

Sources

• Hiscox, Hiscox and Google Cloud Collaborate on AI in lead underwriting for the London Market, December 12, 2023
• Tech Update: Hiscox Launches GenAI-Powered Lead Underwriting Platform, Insurance Journal, August 13, 2024
• Daniel Schreiber, Lemonade Sets a New World Record, Lemonade blog, December 23, 2016
• Lemonade, Investor Presentation, updated as of September 30, 2026
• Celent, Shedding Light on Agentic AI in Insurance, May 20, 2025
• Gallagher Re, Global InsurTech Report for Q3 2026, November 5, 2026
• New York State Department of Financial Services, Insurance Circular Letter No. 7: Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing, July 11, 2024
• NAIC, NAIC Members Approve Model Bulletin on Use of AI by Insurers, December 4, 2023
• NAIC, Implementation of NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers, status as of November 3, 2026
• Colorado Division of Insurance, Notice of Adoption: Amended Regulation 10-1-1, August 20, 2025
• Huskey v. State Farm Fire & Casualty Company, Civil Rights Litigation Clearinghouse
• Chad Hemenway, Lawsuit Against State Farm Alleges Racial Discrimination in Claims Handling, Insurance Journal, December 16, 2022
• Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the European Union, July 24, 2026
• EU AI Act, Annex III: High-Risk AI Systems Referred to in Article 6(2)
• Estate of Gene B. Lokken et al. v. UnitedHealth Group Inc. et al., Georgetown Law Health Care Litigation Tracker
• Lawsuit over UnitedHealthcare claim denials continues, Minnesota Lawyer, February 18, 2025
• U.S. Senate Permanent Subcommittee on Investigations, Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care, majority staff report, October 17, 2024
• Allianz, How Agentic AI Is Transforming Insurance Claims Processing, November 3, 2025

Salient PR works with insurance technology companies across underwriting, claims and risk monitoring. Regulatory descriptions are summaries of the rules as published, not legal advice; company figures are from each company's own statements.