The landscape · Insurance
Insurance: what AI is really doing
Brief me in 2 minutes
AI in Insurance.
AI has moved past chatbots into the two functions that actually decide who pays what: claims triage and underwriting submission review, both now run by software that reads documents and photos and drafts a decision for a human to check. The bigger fight is regulatory: no single U.S. federal AI-insurance rule exists, so state insurance departments are writing their own bias-testing and governance rules on different clocks, and insurers are simultaneously excluding AI failures from old policies while a handful of startups sell narrow new ones to cover exactly that gap.
The game: we show you 6 things happening. You call each one real now, early or mostly hype. Then the receipts.
Shift 1 of 6
Claims move from adjuster-first to AI-first, humans supervise the exceptions
Real now
Lemonade's number is a reported financial metric, not a pilot result, and Corgi is running the same workflow for other carriers' claims now.
How it works. AI now reads the first notice of loss, photos and documents, scores severity, flags coverage issues and drafts a payout before a human ever opens the file. Lemonade's bots run this for its own book; Corgi Claims does the same as a third-party administrator for other carriers, backed by a network of licensed adjusters for anything the model flags.
Overstated: "AI settles claims" describes the easy half of the book. Every one of these systems is built to hand off severity or coverage-flagged cases to a human; the automation number is a mix, not a replacement rate.
Shift 2 of 6
Underwriting submissions get read, scored and routed by AI agents before an underwriter opens them
Early
Named carriers report real time savings, but each deployment is customer by customer with its own line of business, not a sector-wide standard yet.
How it works. Startups sell software that ingests broker emails, ACORD forms, loss runs and statements of value, extracts and validates the data, and produces a triaged, risk-scored submission, sometimes with a draft quote attached. Sixfold, Federato and Cytora each sell a version of this to named commercial P&C carriers.
Overstated: "End-to-end" underwriting is marketing language for a system that still needs a human in the loop for anything outside the appetite band; the vendors themselves describe the human role as supervising, not being replaced.
Shift 3 of 6
Fraud detection becomes an always-on layer across claims and underwriting, not a separate desk
Real now
A ten-year AXA-Shift relationship just got renewed for five more, which is a vendor being kept after a decade of use, not one being piloted.
How it works. Instead of a standalone fraud unit reviewing flagged files, insurers now run AI decisioning across the whole claims and underwriting pipeline, scoring every submission and claim against patterns pooled from many carriers' data, not just their own history.
Shift 4 of 6
AI-native companies become the risk-bearing entity, not just the vendor selling software to one
Early
Both launched or are launching in 2026; neither has a loss history long enough yet to say the underwriting actually works.
How it works. Corgi went from selling AI-underwritten policies as an MGA to holding an admitted insurance carrier license itself, and Ollive is building an MGA to sell AI-liability cover directly rather than licensing detection software to incumbents.
Overstated: Corgi's admitted carrier debut is aimed at dry cleaners, salons and small apartment buildings, not the AI-startup clientele that built its name; the AI-native pitch and the actual first product line are two different stories.
Shift 5 of 6
State regulators write AI rules one at a time while Washington argues over who gets to
Early
Colorado's own bias-testing deadline only just landed in mid-2026 and the multi-state examiner tool is still mid-pilot, so nobody has been enforced against under the new bar yet.
How it works. There is no single federal rule for AI in insurance decisions. Instead, the NAIC's Model Bulletin on AI governance has been picked up by roughly two dozen states, a twelve-state pilot is testing a shared AI Systems Evaluation Tool for examiners, and Colorado's own regulation requires insurers to bias-test the algorithms behind auto, health and life decisions and document it.
Overstated: Anyone expecting a unified federal AI-insurance framework soon is reading the room wrong; the NAIC's own position is to keep this at the state level.
Shift 6 of 6
Insurers start selling narrow cover for AI's own mistakes, while quietly excluding them from everything else
Early
Only a handful of named products exist, each capped at a fixed dollar limit, while AI exclusions are already rolling into standard policies faster than affirmative cover is being written.
How it works. New products from Armilla AI and Munich Re with Mosaic now price specific AI failure modes: hallucinated outputs, algorithmic discrimination, copyright claims from generative output, and regulatory fines. Munich Re's aiSure pays out on a pre-agreed accuracy benchmark rather than adjudicating each claim, which is what lets it settle in weeks instead of years.
Overstated: "Insurable AI risk" sounds like actuaries have a proven loss model; a parametric trigger is really a bet on where you set the accuracy benchmark, not a claims history, and none of these products has run long enough to prove the benchmark was set right.
The scary part
What happens to jobs in Insurance
What AI is already taking:
- First notice of loss and routine claims settlement Investing.com ↗
- Claims first-notification review: assessing severity and flagging coverage issues before a human adjuster opens the file fintech.global ↗
- Underwriting submission intake and data entry: reading broker emails, ACORD forms and loss runs and executing the workflow Sixfold ↗
- Fraud scoring across claims and underwriting, run continuously instead of by a dedicated investigations desk Shift Technology ↗
What stays human:
The judgment calls stay human: deciding what a carrier's actual risk appetite is, negotiating a disputed large claim, and signing off on the bias-testing report a regulator can subpoena. Nobody has built an AI system yet that a state insurance examiner will accept as the accountable party when a decision gets challenged.
Get better · 15 minutes
The ability to audit an algorithmic underwriting or claims decision for bias and document that testing the way Colorado's regulation now requires; that is a specific, defensible skill, not a vague "AI literacy" line on a resume.
Try this this week:
Pull up Colorado's Amended Regulation 10-1-1 compliance report template on the Colorado Division of Insurance site, then run one of your own recent underwriting or claims decisions through its required disclosures (data sources used, testing performed, remediation steps) to see how far your current documentation is from what a regulator will ask for in July 2026.
Your move
Underwriting-workflow AI (Sixfold, Federato, Cytora, FurtherAI) is already crowded with well-funded players selling to the same short list of P&C carriers. The open space is narrower: products that hold the risk on AI failure itself, where only Armilla and a Munich Re joint venture exist, and compliance tooling built specifically to Colorado's bias-testing bar before the rest of the country catches up.
Watch whether a deal ties AI directly to a loss ratio, quote-to-bind ratio or claims-expense number a carrier actually reports, versus one that only cites internal pilot metrics; the NAIC's own state examiners are about to start asking carriers for exactly that documentation. Red flag: a pitch that claims to replace underwriters rather than one that names which specific manual step it removes.
If you write commercial P&C, pull your last 90 days of declined or delayed submissions and run them through Cytora's or Sixfold's public demo to see what a triage agent would have flagged before your team saw the file; measure only the time-to-first-decision, not the bind rate.
Learn to audit an AI underwriting or claims model for bias, specifically the quantitative testing Colorado's Regulation 10-1-1 now requires; start by reading the actual regulation text and Colorado's compliance report template before the July 2026 deadline forces every carrier to hire for exactly this.
What's actually changing
- 1
Claims move from adjuster-first to AI-first, humans supervise the exceptions
How it works. AI now reads the first notice of loss, photos and documents, scores severity, flags coverage issues and drafts a payout before a human ever opens the file. Lemonade's bots run this for its own book; Corgi Claims does the same as a third-party administrator for other carriers, backed by a network of licensed adjusters for anything the model flags.
Why it matters. The payoff shows up directly in the loss ratio insurers report to shareholders, not just in a demo: Lemonade's claims-handling expense is now roughly half the industry norm.
What's overstated"AI settles claims" describes the easy half of the book. Every one of these systems is built to hand off severity or coverage-flagged cases to a human; the automation number is a mix, not a replacement rate.
LemonadeCorgi
2 receipts
- Lemonade's Q2 2026 results show more than half its claims fully automated and a record-low claims-expense ratio.
The company now automates more than 50% of claims, resulting in a loss adjustment expense ratio of just 5%
Investing.com, 29 July 2026 ↗ - Corgi's AI-native claims TPA screens every claim at first notice before a human adjuster sees it.
The platform will review every claim at the first notification, using AI to assess severity, flag potential coverage issues and identify missing information before an adjuster begins reviewing the file
fintech.global, 30 June 2026 ↗
- Lemonade's Q2 2026 results show more than half its claims fully automated and a record-low claims-expense ratio.
- 2
Underwriting submissions get read, scored and routed by AI agents before an underwriter opens them
How it works. Startups sell software that ingests broker emails, ACORD forms, loss runs and statements of value, extracts and validates the data, and produces a triaged, risk-scored submission, sometimes with a draft quote attached. Sixfold, Federato and Cytora each sell a version of this to named commercial P&C carriers.
Why it matters. Underwriters stop doing data entry and start doing the judgment calls the job was supposed to be about, which is also how carriers quote faster without adding headcount.
What's overstated"End-to-end" underwriting is marketing language for a system that still needs a human in the loop for anything outside the appetite band; the vendors themselves describe the human role as supervising, not being replaced.
SixfoldFederatoCytoraZurich North AmericaSkyward Specialty
3 receipts
- Sixfold's pitch is that its AI agents run the underwriting workflow while people handle judgment and strategy.
AI agents execute the work end-to-end. Humans focus on judgment, portfolio performance, and market opportunities.
Sixfold, 29 Jan 2026 ↗ - Federato customers report large gains in quote speed and how much of the business they write is on-appetite.
Federato customers regularly see 90% improvement in quote speed, 3x improvement in the proportion of good business bound, and 50-90% reduction in systems used by underwriters.
Federato, 20 Nov 2024 ↗ - Cytora's Autopilot runs risk workflows without a human triggering each step, reacting as new data arrives.
The platform allows workflows to execute automatically without human intervention, responding to available data and adapting as new information arrives
fintech.global, 18 Mar 2026 ↗
- Sixfold's pitch is that its AI agents run the underwriting workflow while people handle judgment and strategy.
- 3
Fraud detection becomes an always-on layer across claims and underwriting, not a separate desk
How it works. Instead of a standalone fraud unit reviewing flagged files, insurers now run AI decisioning across the whole claims and underwriting pipeline, scoring every submission and claim against patterns pooled from many carriers' data, not just their own history.
Why it matters. Pooled, cross-carrier data catches fraud rings that a single insurer's own claims history would never surface, and it does it before payout instead of after.
Shift TechnologyAXAGradient AI
2 receipts
- Shift Technology and AXA renewed their partnership for another five years, running AI decisioning across claims, fraud and underwriting in 15 countries.
reinforcing a long-term collaboration that now spans 15 countries across Europe, Asia, and Latin America
Shift Technology, 5 Mar 2026 ↗ - Gradient AI prices and screens risk using a data lake built from many carriers' policies and claims, not just one insurer's own book.
Its SaaS platform draws on a proprietary data lake spanning tens of millions of policies and claims, layered with economic, health, geographic, and demographic signals.
AI News, 9 Mar 2026 ↗
- Shift Technology and AXA renewed their partnership for another five years, running AI decisioning across claims, fraud and underwriting in 15 countries.
- 4
AI-native companies become the risk-bearing entity, not just the vendor selling software to one
How it works. Corgi went from selling AI-underwritten policies as an MGA to holding an admitted insurance carrier license itself, and Ollive is building an MGA to sell AI-liability cover directly rather than licensing detection software to incumbents.
Why it matters. Owning the balance sheet captures more of the premium than selling software ever could, but it also means these companies now carry the loss-ratio risk they used to just help someone else price.
What's overstatedCorgi's admitted carrier debut is aimed at dry cleaners, salons and small apartment buildings, not the AI-startup clientele that built its name; the AI-native pitch and the actual first product line are two different stories.
CorgiOllive
2 receipts
- Corgi launched its own admitted insurance carrier, adding to the MGAs, captives and reinsurers it already runs.
Corgi Insurance, the AI financial infrastructure company, today announced Corgi Insurance Company, Inc., its admitted insurance carrier.
PR Newswire, 26 Aug 2026 ↗ - Ollive is closing a seed round and targeting a summer 2026 launch for vendor-focused AI liability coverage.
The company is also reportedly close to completing a $5 million seed funding round.
InsNerds / The Insurer, 17 June 2026 ↗
- Corgi launched its own admitted insurance carrier, adding to the MGAs, captives and reinsurers it already runs.
- 5
State regulators write AI rules one at a time while Washington argues over who gets to
How it works. There is no single federal rule for AI in insurance decisions. Instead, the NAIC's Model Bulletin on AI governance has been picked up by roughly two dozen states, a twelve-state pilot is testing a shared AI Systems Evaluation Tool for examiners, and Colorado's own regulation requires insurers to bias-test the algorithms behind auto, health and life decisions and document it.
Why it matters. A carrier operating nationally has to build compliance to the strictest state's bar, currently Colorado's bias-testing and reporting regime, rather than one national standard, and the NAIC is actively fighting to keep it that way.
What's overstatedAnyone expecting a unified federal AI-insurance framework soon is reading the room wrong; the NAIC's own position is to keep this at the state level.
NAICColorado Division of Insurance
3 receipts
- About 24 states have adopted the NAIC's Model Bulletin on insurers' use of AI.
That Model Bulletin has now been adopted in approximately 24 states.
Crowell & Moring, 25 Mar 2026 ↗ - Twelve states are piloting a shared NAIC tool for examining insurers' AI systems.
Twelve states are participating, including Colorado, Maryland, Louisiana, Virginia, Connecticut, Pennsylvania, Wisconsin, Florida, Rhode Island, Iowa, Vermont, and California.
Crowell & Moring, 25 Mar 2026 ↗ - Colorado's amended insurance regulation requires bias testing of algorithms and a compliance report by July 1, 2026.
Compliance with the Amended Regulation is required by July 1, 2026, and a compliance report is due to the Division by that date (and annually thereafter).
Faegre Drinker, 2 Sept 2025 ↗
- About 24 states have adopted the NAIC's Model Bulletin on insurers' use of AI.
- 6
Insurers start selling narrow cover for AI's own mistakes, while quietly excluding them from everything else
How it works. New products from Armilla AI and Munich Re with Mosaic now price specific AI failure modes: hallucinated outputs, algorithmic discrimination, copyright claims from generative output, and regulatory fines. Munich Re's aiSure pays out on a pre-agreed accuracy benchmark rather than adjudicating each claim, which is what lets it settle in weeks instead of years.
Why it matters. A company deploying an AI agent can no longer assume its existing E&O, cyber or general liability policy responds if that agent causes a loss; that gap is exactly what these new, capped products are built to fill, and buyers who don't check will find out the hard way.
What's overstated"Insurable AI risk" sounds like actuaries have a proven loss model; a parametric trigger is really a bet on where you set the accuracy benchmark, not a claims history, and none of these products has run long enough to prove the benchmark was set right.
Armilla AIMunich ReMosaic Insurance
2 receipts
- Armilla raised the limit on its standalone AI Liability Policy to $25 million per organization.
Armilla AI, a US-based InsurTech specialising in AI liability coverage, has raised its standalone AI Liability Policy to offer limits up to $25m to expand its insurance offerings amid growing demand for protection against generative AI risks.
fintech.global, 23 Jan 2026 ↗ - Munich Re's aiSure with Mosaic covers four named AI failure modes with a fixed per-claim limit.
LLM hallucination, algorithmic discrimination, copyright infringement, regulatory fines
actuary.info, 8 July 2026 ↗
- Armilla raised the limit on its standalone AI Liability Policy to $25 million per organization.
Who's doing it
11 companies to know, 5 with a round we can source. Newest first where dated.
Search these companies in the startup database ↗ · Explore all research tools
CorgiAI-native insurer and MGA infrastructure
Runs insurance for tech startups and, as of 2026, small businesses too, across MGA, captive, reinsurer and now admitted-carrier structures, plus an AI-native claims TPA.
Latest · 26 Aug 2026Launched its own admitted insurance carrier in August 2026, adding to the MGAs, captives and reinsurers it already operates. PR Newswire ↗
$108M funding round tied to regulatory approval, 12 Jan 2026 source ↗
LemonadeAI-native insurer, claims
Publicly traded insurer whose AI runs quoting (Maya) and claims (Jim) for renters, home, pet, life and car policies.
Latest · 29 July 2026Q2 2026: more than half of claims fully automated, pushing the claims-expense ratio to 5% against an industry norm near 9%. Investing.com ↗
Olliveinsuring AI risk, MGA
MGA building AI liability insurance for companies deploying AI agents and their enterprise customers, distributed through partner carriers.
Latest · 17 June 2026Targeting a summer 2026 launch of its vendor-focused AI liability coverage. InsNerds / The Insurer ↗
$5M seed (reported, not yet closed), 17 June 2026 source ↗
ZestyAIunderwriting, property risk models
Property-level AI risk models used by P&C carriers, reinsurers and brokers to price climate and structural risk.
Latest · 24 Mar 2026Launched Z-SPARK, a model predicting non-weather fire risk at the individual property level. PR Newswire ↗
Cytoraunderwriting, digital risk processing
Autopilot product runs commercial insurance risk workflows from submission to decision, pulling data from emails, documents and calls as it arrives.
Latest · 18 Mar 2026Launched Cytora Autopilot, letting underwriting workflows execute automatically as new submission data arrives. fintech.global ↗
Gradient AIunderwriting, claims, fraud
Underwriting and claims prediction platform for insurers, pricing risk off a shared, cross-carrier data lake rather than one insurer's own history.
Latest · 9 Mar 2026Secured growth capital financing from CIBC Innovation Banking. AI News ↗
Shift Technologyfraud detection, claims decisioning
Fraud-detection and claims automation platform built for insurers, used across claims, underwriting and fraud workflows by carriers including AXA.
Latest · 5 Mar 2026Renewed its strategic partnership with AXA for another five years across multiple countries in Europe, Asia and Latin America. Shift Technology ↗
Sixfoldunderwriting
AI agents that read commercial P&C submissions and execute the underwriting workflow for carriers like Zurich and Skyward Specialty.
Latest · 29 Jan 2026Raised a $30M Series B to build out its AI underwriting agents. Sixfold ↗
$30M Series B, 29 Jan 2026 source ↗
Armilla AIinsuring AI risk
Lloyd's-backed MGA selling standalone AI liability insurance covering model error, harmful outputs, agent failures and AI-specific regulatory fines.
Latest · 23 Jan 2026Raised its standalone AI Liability Policy limit to $25 million per organization. fintech.global ↗
FurtherAIsubmission intake, brokers and MGAs
AI agents for carriers, MGAs and brokers that automate submission intake, policy comparison, claims processing and compliance review.
Latest · 7 Oct 2025Raised one of the largest Series A rounds in insurance AI, led by Andreessen Horowitz. FurtherAI ↗
$25M Series A, 7 Oct 2025 source ↗
Federatounderwriting, portfolio risk
RiskOps platform that has AI triage submissions against a carrier's target portfolio and feed underwriters real-time appetite guidance.
Latest · 20 Nov 2024Customers report large gains in quote speed and share of on-appetite business written. Federato ↗
$40M Series C, 20 Nov 2024 source ↗
Your move
Our read, built on the receipts above.
Underwriting-workflow AI (Sixfold, Federato, Cytora, FurtherAI) is already crowded with well-funded players selling to the same short list of P&C carriers. The open space is narrower: products that hold the risk on AI failure itself, where only Armilla and a Munich Re joint venture exist, and compliance tooling built specifically to Colorado's bias-testing bar before the rest of the country catches up.
Watch whether a deal ties AI directly to a loss ratio, quote-to-bind ratio or claims-expense number a carrier actually reports, versus one that only cites internal pilot metrics; the NAIC's own state examiners are about to start asking carriers for exactly that documentation. Red flag: a pitch that claims to replace underwriters rather than one that names which specific manual step it removes.
If you write commercial P&C, pull your last 90 days of declined or delayed submissions and run them through Cytora's or Sixfold's public demo to see what a triage agent would have flagged before your team saw the file; measure only the time-to-first-decision, not the bind rate.
Learn to audit an AI underwriting or claims model for bias, specifically the quantitative testing Colorado's Regulation 10-1-1 now requires; start by reading the actual regulation text and Colorado's compliance report template before the July 2026 deadline forces every carrier to hire for exactly this.
What's next
- 1 July 2026 Colorado insurers' first annual AI governance compliance report due to the Division of Insurance faegredrinker.com ↗
- Sept 2026 NAIC's 12-state AI Systems Evaluation Tool pilot concludes, ahead of expected adoption at the NAIC Fall National Meeting crowell.com ↗
Other industries
- FinanceBanks move agentic AI from chatbots to running back-office and dealmaking work
- Software engineeringSpaceX buys the leading AI coding startup outright, for $60 billion
- HealthcareAmbient scribes go from pilot project to clinical infrastructure
- LegalLegal AI vendors go from research assistant to agent that produces the deliverable
- Real estateLeasing and resident calls move to voice AI agents that never sleep
- Sales and marketingCustomer support agents now get paid per resolved ticket, not per seat
- Jobs, HR and recruitingAI runs the first-round interview, not just the resume filter
- CybersecurityAutonomous AI agents take over penetration testing and vulnerability hunting
- Consulting and accountingAI-written client reports become a recurring, public Big Four embarrassment
- Retail and e-commerceShopping agents get a checkout button, and Amazon spends 2026 deciding who's allowed to press it
- EducationNYC and LA ban AI for students, not for teachers
- Media and creativeAI video splits into model labs that train and aggregators that resell
- Manufacturing and roboticsHumanoid robots move from demo livestreams to paid warehouse shifts
- Transport and robotaxisRobotaxis go from demo to metered fare across a dozen-plus US cities
- Energy and data-centre powerGrid rules get rewritten mid-buildout, state by state
- GovernmentGSA turns AI adoption into a government-wide subscription, then flips 'free' to metered
- Agriculture and foodAutonomous tractor fleets learn to run the tools behind them, not just the wheel