The State of AI in Insurance: Who Built It, Who Bought It, and Who Is Winning
The insurance industry has spent roughly a decade talking about artificial intelligence (that’s right, more than 3 years). It has spent roughly three years actually deploying it. The gap between those two timelines explains almost everything about the current state of the market: why a handful of carriers are now publishing audited nine-figure savings while the majority remain stuck in pilot programs, why the vendor landscape has consolidated so quickly, and why the strategic question facing every carrier, MGA, and brokerage has quietly shifted from whether to adopt AI to how to adopt it without becoming one of the cautionary tales.
This piece is an attempt to map the full territory. Not the conference-panel version, where every carrier is transforming and every vendor is a category leader, but the version supported by earnings calls, annual reports, named deployments, and survey data from firms that have no incentive to flatter anyone. Who built what. Who bought what. What the chief executives are actually saying to analysts when the marketing teams are not writing the script. And, because the data now permits it, who is winning and who is falling behind.
A note on method before we begin. Every hard number in this piece is attributed to a named source. Where carriers have disclosed results themselves, through investor presentations or earnings calls, I have said so. Where the figure comes from a research house, a consultancy, or a trade publication, I have named it. The insurance industry has been burned repeatedly by AI claims that evaporate under scrutiny, and the only antidote is sourcing discipline.
How We Got Here: A Decade of Machine Learning Before the Headlines
It is worth correcting the record on timing, because the popular narrative that insurance discovered AI when ChatGPT launched is wrong, and the error leads to bad strategy. The industry's leaders did not start in 2023. They started around 2013, and the compounding difference shows.
Progressive began collecting driving behavior data through telematics devices in the late 1990s and has been refining usage-based pricing models for a quarter century. Aviva, by its CEO's account, has priced over 98 percent of its UK personal lines retail business with machine learning and spent years training more than 150 machine learning models in claims on proprietary data before generative AI arrived. State Farm has filed 326 AI-related patents since 2014. Travelers built its geospatial claims analytics and drone-based catastrophe response before large language models existed. Lemonade launched in 2016 with AI claims handling as its founding architecture, not a retrofit. The pattern matters because it explains the single most consistent finding across every serious study of the current wave: generative AI returns accrue disproportionately to organizations that had already industrialized their data, their measurement, and their model governance in the machine learning era. The carriers now publishing nine-figure AI results are not early generative AI adopters. They are late-stage machine learning organizations that added a new class of model to an existing operational discipline.
What changed in 2023 was not that AI became useful to insurance. It is that AI became useful for the specific thing insurance is made of: documents. Classical machine learning handled structured data brilliantly, which is why it colonized pricing and fraud scoring first. But the bulk of insurance work product, submissions, loss runs, policy forms, medical records, broker correspondence, inspection reports, is unstructured text, and until large language models, extracting meaning from it at production accuracy was somewhere between expensive and impossible. Datos Insights' finding that the top production use cases in 2026 are all variations on AI reading and summarizing documents is not evidence of a limited technology. It is evidence that the technology finally reached the substance of the work. Insurance is, at an operational level, a document processing industry with a balance sheet attached, and for the first time the documents themselves are computable.
That framing also explains why the current wave is diffusing so much faster than any prior insurance technology cycle. Core system replacements took decades because they required ripping out the operational spine. Document intelligence requires no such surgery. It layers onto existing workflows at the point where a human would otherwise read, extract, compare, or rekey, which is why adoption jumped from roughly a third of the market to nearly all of it in under three years, and why the constraint on value has shifted entirely from technology availability to organizational approach.
The Adoption Numbers Tell Two Different Stories
Start with the topline figures, because they are genuinely striking. Conning's 2025 industry survey found that 90 percent of insurers are now somewhere on the generative AI journey, with 55 percent in early or full deployment and machine learning adoption at 74 percent. Deloitte's research shows adoption correlates strongly with scale: among carriers with annual revenues above $10 billion, 91 percent are adopting or actively deploying AI, against 58 percent in the $100 million to $500 million segment. Evident, which tracks AI activity across 30 leading life, P&C, composite, and reinsurance groups in North America and Europe, measured an 87 percent year-over-year increase in AI deployments across the sector.
Datos Insights, in a Q1 2026 survey of North American insurance executives, declared the pilot phase over for property and casualty carriers. Most carriers now have AI in production. That is the first story, and it is true.
The second story is also true, and it is far less comfortable. The same Datos research found that almost no carrier has achieved widespread adoption, and that the top production use cases are overwhelmingly variations on a single theme: AI reading and summarizing documents. BCG's 2025 study of AI adoption in insurance found that only 7 percent of insurance companies surveyed have successfully brought their AI systems to scale, with roughly two thirds still piloting. Evident's data shows that while around 40 percent of insurers report tangible business benefits from AI, 77 percent of those benefits come from productivity gains and only 5 percent link to revenue growth. And a Simplifai report cited by CIO Dive in April 2026, drawing on McKinsey, EY, and Deloitte data, summarized the claims function in eight words that could describe most of the industry: lots of pilots, limited production, minimal P&L impact.
So the honest picture in mid-2026 is this. Adoption is nearly universal in name. Production deployment is common but narrow. Scaled, enterprise-wide deployment with measured P&L impact is rare, concentrated in perhaps a dozen carriers globally, and those carriers are pulling away from the pack at a pace this industry has not seen since the direct-to-consumer disruption of the early 2000s. The rest of this piece is about what separates them.
Claims: Where the Money Has Actually Moved
If you want to find real, audited AI returns in insurance, start in claims. It is the function with the highest transaction volume, the most unstructured documentation, the clearest cost baselines, and the shortest feedback loops. It is also where the single most instructive case study in the global industry currently sits.
Aviva's claims transformation is the deployment every other carrier now gets measured against. Working with QuantumBlack, McKinsey's AI arm, Aviva assembled a team of more than 50 data scientists, engineers, business leaders, and translators and deployed more than 80 AI models across its claims lifecycle. The published results, documented in McKinsey's case study and Aviva's own investor communications, are specific in a way that AI results in this industry almost never are: complex liability assessment times reduced by 23 days, routing accuracy improved by 30 percent, customer complaints down 65 percent, and more than 60 million pounds in reported savings from the motor claims transformation in 2024 alone. Speaking to diginomica in early 2026, Group CEO Amanda Blanc put the cumulative figure higher, saying “We have already saved nearly £100 million through our claims transformation” and pointing to agentic AI as the next unlock. Aviva has since told investors that a virtual agent capable of handling a simple claims call from beginning to end would be live by the summer of 2026, per its 2025 full-year results coverage in Insurance Times.
Two details of the Aviva program deserve more attention than they get. First, the foundation is proprietary data at scale: Blanc has noted that over 98 percent of Aviva's UK Personal Lines retail business is priced with machine learning and that the company has been training more than 150 machine learning models in claims on its own data for years. Second, Aviva measured outcomes at the workflow level, not the adoption level. It reported how fast files moved, how accurately they were routed, and what they cost, rather than counting logins or pilot launches. Most carriers still cannot do this, and the measurement gap is itself a competitive gap.
Lemonade represents the other proven claims architecture: not AI layered onto an existing operation, but an operation built around AI from inception. The company's AI intake and claims agents, Maya and Jim, have been in production for years, and its Q2 2026 investor materials, covered by Investing.com, show the model at maturity: more than 50 percent of claims fully automated, producing a loss adjustment expense ratio of roughly 5 percent, about half the 9 percent typical of at-scale incumbents. President Shai Wininger told investors the company is “model-agnostic” on AI providers and can move from evaluating a new frontier model to implementing it in hours. Whatever one thinks of Lemonade's path to profitability, and its 15 percent share decline after those same Q2 results shows the market still has questions, the claims cost structure it has demonstrated is now the reference point for what full automation of simple claims actually yields.
The incumbent giants have moved too, mostly through purchased and partnered capability. Travelers, which handled 1.5 million claims in 2025, described by CEO Alan Schnitzer on the Q4 2025 earnings call as “about one every 20 seconds”, has built photo-first claims intake through its MyTravelers app, computer vision severity triage, a claims geospatial analytics platform, and a drone fleet of more than 700 FAA-certified pilots for catastrophe response, while paying out more than $23 billion in claims and closing 90 percent of catastrophe claims within 30 days, per Carrier Management's reporting. In January 2026, Travelers announced it was giving nearly 10,000 employees personalized Claude AI assistants through its partnership with Anthropic, with more than 30,000 employees having access to frontier models, according to Emerj's research on the carrier.
Claims Journal reported in June 2026 that State Farm, Allstate, Progressive, Liberty Mutual, Nationwide, and USAA have all acknowledged day-to-day AI use across underwriting, pricing, claim intake, and claims processing. Allstate disclosed in 2025 earnings calls that AI-assisted claims handling reduced cycle times on a portion of auto claims. The direction of travel among the US personal lines giants is uniform even where the disclosure is thin.
Then there is the infrastructure layer beneath all of it. CCC Intelligent Solutions, the cloud platform that effectively runs US auto claims, has more than 100 insurers using its computer vision and natural language processing across tens of millions of claims and repairs annually. Its December 2024 acquisition of EvolutionIQ, whose claims guidance AI had been adopted by 7 of the top 15 US disability carriers, extended AI-driven claims decisioning from auto physical damage into injury, disability, and workers' compensation. Shift Technology has built a billion-dollar fraud detection business, catching over $5 billion in fraud annually according to industry analyses, with deployments like Diot-Siaci reporting a 75 percent success rate in identifying claims fraud, per Actuidata. Tractable's computer vision reviews vehicle damage in seconds at accuracy levels its carrier customers publicly cite around 95 percent.
One caution belongs in any honest accounting of claims AI, and Risk & Insurance, drawing on Gallagher Re's Q4 2025 InsurTech report, named it precisely: the ROI paradox. Cycle times and handle times are falling across the industry, but the numbers boards and CFOs care about, loss ratios, net settlement costs, subrogation recovery, are not yet moving at the same pace for most adopters. Efficiency is real. Indemnity impact remains concentrated among the carriers that deployed at the decision level rather than the summary level. That distinction will come up again.
Underwriting: The Most Conservative Function Is Moving the Fastest
Underwriting was supposed to be the last holdout. It is the function where errors compound for years, where regulatory scrutiny is heaviest, and where institutional knowledge is most jealously guarded. Instead, over the past eighteen months, it has become the most active deployment surface in commercial insurance, and the reason is structural: underwriting is drowning in unstructured documents, and reading unstructured documents is the single thing large language models do best. Datos Insights found that the top production AI use cases across North American P&C are all variations on AI reading and summarizing documents. Underwriting is where those documents pile highest.
The named deployments now span the market. Zurich North America selected Sixfold to generate underwriter-ready risk narratives for its Middle Market team, saving underwriters an average of two hours per submission, according to CB Insights' customer documentation. Sixfold crossed one million underwriting submissions processed across more than 40 lines of business in late 2025, and its customer roster, Zurich North America, AXIS, Generali Global Corporate and Commercial, Guardian, Skyward Specialty, Mosaic, and New York Life among them, collectively represents roughly $265 billion in gross written premium, per APIs.io's provider documentation. When Skyward Specialty announced its Sixfold partnership in December 2025, CEO Andrew Robinson called it part of an “AI underwriting arms race”, which is about as candid as a public company chief executive gets about competitive dynamics. Sixfold's January 2026 Series B included strategic investment from Guidewire, and by June 2026 the company had launched an AI Underwriter product that actuary.info reports can be configured to produce bind-ready output, with performance data across 1.5 million submissions showing processing time reductions of 50 to 97 percent and hit ratio gains of 15 percent or more.
Cytora took the intake route and landed one of the most striking productivity numbers in the sector: Markel, an early customer of Cytora's Autopilot platform launched in March 2026, reported a 113 percent increase in underwriting productivity measured by written premium per full-time employee. Federato has built its RiskOps platform around portfolio-aware underwriting prioritization for complex commercial and specialty risks. Hyperexponential supplies pricing and decision infrastructure to commercial and reinsurance teams, particularly in the London Market. Gradient AI serves underwriting and claims risk scoring across group health and workers' compensation. N2G Worldwide reported a 40 percent increase in underwriter quote capacity and a 60 percent reduction in cycle times after deploying AI agents.
The property data layer has consolidated into genuine market infrastructure. Cape Analytics, acquired by Moody's in January 2025, supplies AI-derived property condition data, and its Roof Condition Rating is used by nearly half of the top 50 US property insurers and approved in 40 states. Zurich partnered with Nearmap to bring AI-driven roof scoring and high-resolution aerial imagery into US underwriting, now live nationwide, flagging damage and risk exposure in seconds to support risk selection and pricing in extreme-weather regions, per Zurich's own disclosures. ZestyAI plays in the same peril-scoring territory for wildfire and severe convective storm.
The personal lines giants have pushed furthest toward autonomy. GEICO now uses AI underwriting agents that approve auto policies by analyzing driving records, vehicle data, and credit information, cutting approval from days to minutes. Travelers' Business Insurance president Greg Toczydlowski described on the Q4 2025 earnings call how recently deployed generative AI agents mine internal and external data sources to assign business classifications and synthesize risk characteristics. And across the sector, the trajectory numbers are steep: industry analyses tracked by BuildMVPFast put AI underwriting adoption at 14 percent today with projections of 70 percent by 2028, alongside quote-to-bind time reductions of 60 to 99 percent among commercial adopters and loss ratio improvements of 3 to 5 points, a range worth treating carefully since a single point of loss ratio at a large carrier is worth tens of millions annually.
Two things should be said plainly about underwriting AI, because vendors will not say them. First, almost none of this is risk selection judgment being automated. It is document intake, data extraction, appetite checking, narrative generation, and classification, the eighty percent of underwriting that was never underwriting. The judgment layer remains human nearly everywhere, and the carriers deploying most successfully are explicit about that boundary. Second, the validation duty does not move. As actuary.info noted in its analysis of bind-capable AI underwriting under ASOP No. 56, when a model produces a decision without a human touchpoint, responsibility for validating that decision stays with the carrier, not the vendor. Any carrier treating an AI purchase as an accountability transfer is misreading both the actuarial standards and the regulators.
Distribution and the Agent Layer: The Quiet Majority
The distribution side of the industry gets less coverage than carriers, which is a mistake, because it is where adoption has broadened fastest. Industry data compiled by Perspective AI puts AI usage among US insurance agencies at 64 percent in 2026, using AI in at least one core workflow, up from 38 percent two years earlier. The drivers were not agents discovering technology on their own. Carriers like GEICO, Lemonade, Progressive, and State Farm dragged the customer experience bar forward until agencies either adopted or lost quotes.
The workflow pattern on the distribution side is pragmatic: 49 percent of agencies now use AI somewhere in the claims process, most commonly first notice of loss intake, document classification, and status communications, while 44 percent have adopted it in service, largely for after-hours coverage, certificate of insurance generation, and policy-change questions. Voice AI has found its first durable niche in after-hours coverage at independent agencies, where more than 80 percent of inbound volume is telephone and the alternative is voicemail. The same research finds broker adoption lags carrier adoption by roughly 18 to 24 months on customer-facing surfaces while leading on internal productivity, which matches what anyone who has spent time in both worlds would expect: agencies adopt what saves their own hours first.
There is a strategic subtext here that surfaced, of all places, on Chubb's Q1 2026 earnings call. Asked whether brokers using AI to lower their own expenses creates an opportunity for carriers to reduce acquisition costs, Chairman and CEO Evan Greenberg answered that it does, at the right moment, and went considerably further: he called intermediation costs in numerous parts of the business excessive and said that in an age of digitalization and AI, technology will ultimately bring those costs down. When the CEO of one of the world's most profitable commercial insurers says the distribution cost structure is a target, every brokerage and MGA should understand that their AI adoption is not just an efficiency play. It is a defense of their economics. The intermediaries that use AI to demonstrably add value per dollar of commission will keep their economics. The ones that do not will find carriers increasingly willing to route around them.
State Farm illustrates the opposite instinct, and it is equally rational for its model. The largest US P&C insurer, with more than 96 million policies and accounts, has built its AI roadmap around an explicit augment-not-replace thesis for its 19,200 agent offices, per Perspective AI's analysis of the carrier's 2026 program. State Farm joined OpenAI's Frontier platform as a launch partner in late 2025, deployed a knowledge assistant in its contact centers, and has filed 326 AI-related patents since 2014, among the highest totals in US P&C, covering claims triage, autonomous vehicle fault analysis, and underwriting. The strategic constraint is genuine: the agent network is the moat, and any deployment that erodes the personal relationship damages the business it is meant to grow. USAA has followed a similar shape, conservative on customer-facing generative AI, aggressive on internal agent-assist, per the same research.
The Broker Giants Move: February 9 and What Followed
If you want a date for when AI stopped being an operational topic and became an existential one for insurance distribution, the market supplied it: February 9, 2026. That day, two AI-powered insurance apps went live inside ChatGPT, a car insurance comparison tool from Insurify and a home insurance quoting app from Spanish digital insurer Tuio, and public broker valuations convulsed. By MarshBerry's accounting, WTW fell 12 percent, Aon 9.9 percent, and Arthur J. Gallagher 9.3 percent in a single session, with the MarshBerry Broker Composite Index down 8.9 percent before the ripple reached European insurance stocks the following day. Bank of America then put a number on the anxiety, estimating $15 billion in low-complexity insurance commissions at risk from AI disintermediation. The sell side largely counseled calm, with Goldman Sachs calling the move “overdone” and McKinsey concluding AI would reshape distribution models rather than disintermediate them, and the consensus is probably right about the timeline. But the mechanics that triggered the repricing were real and new: for the first time, an insurance product could be quoted with a real price for a real buyer inside an AI platform where hundreds of millions of people already do their research. The apps were trivial. The channel was not.
What has followed is the most consequential development of 2026 in distribution: the global brokers have stopped treating AI as a productivity initiative and started treating it as strategy, publicly, with capital and named partners attached.
Marsh McLennan used its Q1 2026 earnings call to make the most explicit competitive claim, with CEO John Doyle telling analysts directly “why we believe Marsh will be an AI winner” and grounding the argument in scale, capacity to invest, proprietary data assets, and the trusted adviser position, per Coverager's call analysis. Notice that this is nearly word for word the same advantage checklist Amanda Blanc articulated for Aviva, translated into brokerage. The delivery is already visible in product: the LenAI platform and Sentrisk supply chain analytics documented in Marsh's brokerage market review, AI-enabled tools like ADA, Centrus, and GC Quotebox, and the Risk Companion suite unveiled at RIMS RISKWORLD 2026, where Renewal Companion lets clients model retentions, limits, and program design in real time and Captive Companion automates captive reporting and benchmarking, per Insurance Business. The organizational signals are just as loud: Martin South moved into a Chief Client Officer role focused on AI-enabled client experience, and Oliver Wyman's AI Quotient practice has become the consultancy's fastest-growing unit, advising on more than $50 billion in AI-related capital deployment. Marsh is positioning to profit from the AI cycle twice, as an adopter and as an adviser to everyone else's adoption.
Brown & Brown made the boldest single commitment of any distribution firm. On July 23, 2026, the brokerage announced it was becoming an AI-first enterprise and enlisted Anthropic, McKinsey, and Accenture to rewire the business, per its release and Business Insurance's coverage. The disclosed pilot results explain the conviction: productivity gains of up to eightfold in participating teams, software troubleshooting times cut 80 to 90 percent, engineering tasks that took days completed in hours using Claude Code, and roughly 80 percent of participating employees rating the tools at the highest value level. The plan extends Claude across approximately 23,000 employees and embeds AI into customer service, operations, technology, and corporate functions, governed by a dedicated value management office, with the company stating the initiative “is not being undertaken as a workforce reduction effort”. Two details make Brown & Brown the case study worth watching. First, the sequencing: the firm's own Q1 2026 investor materials describe a technology journey running since 2015 through platform rationalization and data standardization before the AI layer, exactly the data-readiness-first pattern Datos Insights identifies as the strongest predictor of AI outcomes. Second, the honesty about timing: management has told investors the meaningful margin impact arrives in years three to five, which is what a credible enterprise AI program sounds like and what pilot-purgatory programs never say. And it is worth stating plainly what Brown & Brown did not do: a brokerage with the scale to build almost anything chose to partner with a frontier lab and two consultancies rather than construct its own stack. The largest distribution firms are reaching the same buy-and-partner conclusion the largest carriers already reached.
The through line from February 9 to these announcements is the one Evan Greenberg identified from the carrier side: distribution economics are being repriced, and the brokers know it. The public market showed the giants what the penalty for perceived AI passivity looks like, and within months the giants responded with named partners, named products, and named numbers. The firms further down the distribution stack, the regional brokerages, wholesalers, and MGAs without a frontier lab partnership or an investor day, should read that sequence as their own preview. The repricing arrives for everyone. The only variable is whether a firm meets it with pilots or with production.
The Specialty and E&S Frontier: Where the Operational Math Is Most Brutal
No segment of the market makes the case for AI adoption more starkly than excess and surplus lines, and the argument is hiding in plain sight inside WSIA's own stamping office data. Surplus lines premium across the 15 stamping office states reached $90.3 billion in 2025, up 7.8 percent from $83.8 billion in 2024, capping a run in which the segment posted its seventh consecutive year of double-digit or near-double-digit growth, per WSIA's annual reports and AM Best's segment tracking. But the number that should keep every wholesale and specialty operations leader awake is the other one in the same report: item counts rose 14.1 percent in 2025, nearly twice the rate of premium growth.
Read those two figures together and the operational story writes itself. Transaction volume is growing much faster than premium, which means average premium per transaction is falling, which means the revenue available to fund the manual handling of each submission, quote, binder, endorsement, and filing is shrinking while the volume of that handling explodes. At midyear 2025 the stamping offices processed 3.7 million items; full-year 2024 saw nearly 7 million. Every one of those items represents documents to be read, data to be extracted, forms to be checked, and compliance requirements to be verified across a patchwork of state-specific surplus lines rules. AM Best's November 2025 revision of its E&S outlook from positive to stable, citing moderating premium growth and early rate softening, sharpens the point further: the segment spent seven years being paid hard-market rates to do soft-market volumes of paperwork, and that subsidy is ending.
This is precisely the environment where document AI stops being an efficiency nicety and becomes a margin survival requirement. E&S risks are, by definition, the risks that do not fit standard forms and standard appetites, which means their submissions are messier, their comparisons harder, and their manual processing costs higher than the admitted market's. The MGAs and wholesalers that have moved earliest on AI-driven intake, policy checking, and compliance workflows are not chasing innovation credit. They are responding to a transaction economics problem their own filing statistics have been announcing for years. Conning's MGA research and AM Best's delegated authority segment reports have both documented the extraordinary growth of the MGA channel, and that growth has been built substantially on operational leverage. AI is now the primary source of the next increment of that leverage, and the specialty distribution firms that fail to capture it will find their expense ratios diverging from competitors within two renewal cycles, not ten.
Agentic AI: The 2026 Inflection
If 2024 was the year of copilots and 2025 the year of production document intelligence, 2026 is unmistakably the year agentic AI crossed from concept into early insurance production, and it is worth being precise about what that means, because the term is being abused daily. An agent, in the operative sense, is a system that completes a multi-step workflow, deciding, acting, checking, and escalating, rather than producing a draft for a human to carry forward. The distinction is the difference between software that helps someone do a task and software that owns a task.
The market data shows the transition mid-leap. Gartner expects 40 percent of enterprise applications to embed task-specific AI agents by the end of 2026, up from under 5 percent in 2025, yet only around 17 percent of organizations have actually deployed agents so far, per analysis compiled by consultant Tommaso Maria Ricci. In insurance specifically, Datos Insights confirms agentic AI has crossed from concept to early production, and analyst projections tracked by Roots suggest more than 35 percent of insurers will deploy AI agents across at least three core functions by late 2026.
The live deployments make the abstraction concrete. Aviva told the market in March 2026 that a virtual agent would handle simple claims calls from beginning to end by summer, moving the technology directly into the most sensitive customer moment insurance has. Sixfold's AI Underwriter, live across six carriers, can be configured to produce bind-ready output without a human touchpoint. Cytora's Autopilot automates the full submission pipeline from broker email to structured, guideline-checked data. GEICO's underwriting agents approve auto policies in minutes. Lemonade has been running the pattern for years and now automates more than half of claims end to end.
The sane deployment doctrine for 2026, and the one visible in every successful program, is scoped autonomy: agents own the low-risk, high-volume steps, first notice of loss intake, document handling, data entry, appetite screening, status communication, while humans hold the consequential decisions, with clean escalation paths between the two. Aviva's own architecture, which its team describes as seamlessly switching between digital and human tracks, is the template. The carriers getting agentic AI wrong are the ones treating autonomy as a binary rather than a dial, and the regulatory framework discussed later in this piece will punish that error before the market does.
The Economics of the Wave: Follow the Capital
The financial architecture underneath all this deployment activity deserves its own accounting, because capital flows are the most honest forward indicator any market produces, and right now every category of capital, venture, strategic, and internal, is pointing the same direction.
On the venture side, global insurance startups raised roughly $3.9 billion in seed-through-growth financing in 2025, and insurtech funding topped $1 billion in February 2026 alone, with AI-focused deals dominating the mix, per industry funding trackers compiled by BuildMVPFast. The composition of that funding is more revealing than its size: the largest rounds are concentrating in domain-specific workflow AI rather than horizontal tools, and Series A valuations in insurance AI have reached levels that only make sense if investors believe the category winners will own durable workflow positions inside carriers and distribution firms for a decade. Strategic capital is confirming the thesis through acquisition: Moody's buying Cape Analytics, CCC buying EvolutionIQ, Applied taking Cytora, Guidewire investing in Sixfold. Incumbent technology platforms are paying premiums for deployed, validated AI rather than funding internal equivalents, which is the clearest possible market pricing of the build-versus-buy question.
On the demand side, McKinsey projects generative AI could unlock $50 billion to $70 billion in new insurance industry value, concentrated in customer operations, marketing, and software engineering, and market sizing aggregated by Openkoda projects the global AI-in-insurance market nearly tripling to $43.85 billion by 2030. Those projections deserve the standard skepticism owed to all market sizing, but the unit economics visible in individual deployments do not require projection. An illustrative fraud detection case documented in Openkoda's industry analysis describes a system costing roughly $1.8 million upfront generating annual savings around $1.25 million, payback in under two years, and fraud is consistently the fastest-ROI category across the market. Aviva's claims program returned its investment multiple times over in a single year. Markel measured a 113 percent productivity gain. N2G measured 40 percent more quote capacity. When Chubb's CEO dismisses AI token costs as a minor fraction of the efficiency gains, he is describing an ROI profile that most technology categories never achieve at any point in their lifecycle.
The corollary for buyers is a pricing environment worth exploiting deliberately. Vendors flush with venture capital are competing ferociously for referenceable deployments, which means proof-of-concept terms, implementation support, and contractual flexibility are more buyer-friendly right now than they will be once the category consolidates. The carriers and distribution firms negotiating multi-workflow relationships in 2026 are locking in economics that late adopters will not see.
The Talent Equation: The Adoption Driver Nobody Chose
There is one more force behind adoption that has nothing to do with technology enthusiasm, and in candid conversations across the industry it comes up before any other: the people are leaving, and they are not being replaced. A US Bureau of Labor Statistics projection cited throughout the industry put expected insurance workforce losses at 400,000 by 2026, driven by a retirement wave the sector has seen coming for a decade and a recruiting pipeline that has never recovered its appeal to younger professionals. Millions of insurance professionals currently spend the majority of their working hours on manual document handling, data entry, and rekeying between disconnected systems, exactly the work the departing generation performed and the arriving generation will not accept as a career.
This reframes the entire AI conversation in a way vendor marketing rarely captures. For most carriers, MGAs, and agencies, AI adoption is not primarily a headcount reduction program. It is a headcount replacement program for headcount that cannot be hired. The after-hours voice AI at independent agencies exists because the alternative was voicemail, not layoffs. The submission intake automation at wholesalers exists because requisitions for processing roles sat unfilled for two quarters. Travelers giving 30,000 employees frontier model access and State Farm building augmentation for 19,200 agent offices are bets that the scarce human expertise remaining, senior underwriting judgment, complex claims handling, client relationships, gets multiplied rather than spent on paperwork. The demographic math was going to force an operational transformation of this industry regardless of what happened in AI research. The technology arrived at the moment the industry ran out of alternatives, which is the least discussed and most important reason adoption curves in insurance are outrunning most other conservative industries.
The Vendor Landscape: Who Built What and Who Adopted It
The insurance AI vendor market has sorted itself into distinct layers over the past two years, and mapping them clarifies who is actually competing with whom. The categories below are drawn from named deployments, disclosed customer rosters, and funding activity through mid-2026.
The claims infrastructure layer is the most mature. CCC Intelligent Solutions sits at the center of US auto claims with more than 100 insurer customers and an ecosystem connecting carriers, repair shops, parts suppliers, and total loss buyers, with AI modules for photo estimation, total loss, and injury prediction layered on top. Its EvolutionIQ acquisition brought claims guidance AI already adopted by 7 of the top 15 US disability carriers into the fold, extending CCC from auto physical damage into disability and workers' compensation. Tractable supplies computer vision damage assessment to auto and property carriers globally. Shift Technology anchors fraud detection, with agentic capabilities expanding from flagging suspicious claims toward orchestrating full investigations.
The underwriting intelligence layer is where the current land grab is happening. Sixfold, with Zurich North America, AXIS, Generali GC&C, Guardian, Skyward Specialty, Mosaic, and New York Life on its roster and Guidewire money on its cap table, has staked out risk analysis and submission narratives for sophisticated commercial carriers. Cytora, now under Applied Systems ownership according to industry platform analyses, owns submission intake and risk digitization, with Markel's 113 percent productivity gain as its flagship result. Federato's RiskOps platform serves carriers and MGAs managing complex commercial and specialty portfolios. Kalepa competes in the same underwriting workbench territory. Planck supplies generative AI commercial risk data. Hyperexponential holds the pricing and decision infrastructure position in the London Market and beyond.
The property analytics layer has effectively become market plumbing: Cape Analytics under Moody's ownership, Nearmap powering Zurich's US roof scoring, ZestyAI scoring wildfire and severe weather peril, and the geospatial capabilities the large carriers have built internally on top of them.
The workflow automation layer is the newest and, for the day-to-day economics of MGAs, wholesalers, and retail brokerages, arguably the most consequential, because it targets the unstructured document work that consumes the majority of operational headcount: submission intake, policy checking, quote comparison, renewal processing, and compliance workflows. FurtherAI has emerged as the reference point in this category, raising one of the largest Series A rounds in insurance AI history, $25 million led by Andreessen Horowitz just six months after its seed round, and counting Accelerant, Millennial Specialty Insurance, Leavitt Group, and Upland Capital among its disclosed customers and partners, with supported customers writing over $15 billion in premium across all 50 states and a UK and European expansion announced in May 2026. Andreessen Horowitz partner Joe Schmidt captured why the category is being funded this aggressively when he described the founding team as people whose customers see them as “true AI partners, not just AI tools”. That phrase, whatever you think of venture capital rhetoric, names the actual purchasing criterion that has emerged across every layer of this market. Nobody is buying software anymore. They are buying an operating capability that somebody else maintains.
It is a mistake, though a common one, to treat all of this as a P&C story. The life, health, disability, and workers' compensation markets are running the same playbook with their own casts. EvolutionIQ built its claims guidance franchise in group disability, individual disability, and workers' compensation, reaching 7 of the top 15 US disability carriers before CCC acquired it, and its generative medical summarization now extends across life, accident, and casualty lines. Gradient AI supplies underwriting risk scoring and claims decision intelligence to group health and workers' compensation writers. Sixfold's roster notably includes Guardian and New York Life, two of the most conservative franchises in American insurance, and its platform supports life and health lines alongside its 40-plus P&C lines. Aviva has cited medical underwriting as one of its earliest successful generative AI capabilities. The workers' compensation market in particular is a natural habitat for this technology: long-tail claims generating enormous unstructured medical documentation, return-to-work outcomes that improve measurably with earlier and better-guided intervention, and expense structures under permanent pressure. When the most risk-averse buyers in insurance, life carriers and disability writers, are putting AI into production on claims decisioning, the argument that the technology is too immature for regulated insurance work has formally expired.
Above all of these sits the core systems layer, Guidewire and Duck Creek, which is responding to the AI wave the way platform incumbents always do: through investment and integration rather than native invention. Guidewire's strategic stake in Sixfold and Applied's ownership of Cytora are the pattern in miniature. The core system vendors know that whoever owns the AI workflow layer will eventually influence which core systems get renewed, and they are buying position accordingly.
Consolidation is the signal worth watching in all of this. Moody's bought Cape Analytics. CCC bought EvolutionIQ. Applied took ownership of Cytora. Guidewire invested strategically in Sixfold. In each case, established infrastructure players purchased proven, deployed, carrier-validated AI rather than building equivalents internally, despite having engineering organizations orders of magnitude larger than their targets. The sophisticated acquirers in this market, the companies whose entire business is insurance technology, are themselves buyers rather than builders. Carriers evaluating their own build-versus-buy decisions should sit with that fact for a moment.
What the Chief Executives Are Actually Saying
Earnings calls and investor presentations are where AI strategy gets stated under legal exposure rather than marketing incentive, which makes them the most reliable window into how seriously the industry's leadership actually takes this. Read across the 2025 and 2026 transcripts, and a clear hierarchy of conviction emerges.
At the top sits Aviva's Amanda Blanc, who has done something no other major insurance CEO has: tied AI directly to reported financial results, repeatedly and specifically. In Aviva's 2025 full-year results, alongside a 25 percent rise in operating profit to 2.2 billion pounds, Blanc credited AI models for significant claims indemnity benefits and pricing sophistication, per actuary.info's analysis of the presentation, and identified AI as one of three long-term investment priorities alongside growth and the Direct Line integration. Her argument for why Aviva specifically wins is a checklist other carriers should be nervous reading: millions of customers, the ability to deploy and reuse at scale, capacity to invest, and proprietary customer and claims data. Notice what is on that list and what is not. Nowhere does she claim Aviva's advantage is superior AI technology. The advantage is everything Aviva wraps around the technology.
Zurich's Mario Greco has made the most aggressive structural commitment. In launching the Zurich AI Lab in October 2025, a research collaboration with the University of St. Gallen and ETH Zurich's Agentic Systems Lab, Greco said “AI has proven its significant value” for customer service, response times, and risk information, and described the lab, per Reinsurance News, as the company's “moonshot factory” aimed at revolutionizing the business model itself. Zurich's AI360 strategy explicitly targets becoming an AI-native insurer, and its Q4 2025 results, record business operating profit of $8.9 billion, up 14 percent, give Greco something most transformation-minded CEOs never have: the financial credibility to fund conviction. Greco has also positioned Zurich as an insurer of the AI boom itself, noting the company insured more than 200 data center and technology projects in the US last year with a total insured value of $150 billion, per The Insurer's coverage of the 2025 results.
Chubb's Evan Greenberg is the industry's most valuable skeptic, and his commentary rewards close reading precisely because he refuses the hype register. On the Q2 2026 call, asked about rising AI token costs eating into efficiency gains, Greenberg said those costs represent a minor fraction of the efficiencies and improvements Chubb gains, per Reinsurance News, and he has consistently framed AI as supporting combined ratio through insights and efficiencies rather than as a revolution. His most pointed observation, the one about excessive intermediation costs coming down in an age of AI, is not a prediction about technology at all. It is a prediction about which parts of the value chain will be repriced. Greenberg is telling the market that AI changes bargaining power, and almost nobody in distribution is listening carefully enough.
Travelers' Alan Schnitzer has taken the practitioner's register: specific volumes, specific deployments, specific use cases. Claims call center reductions, generative AI classification agents in Business Insurance, frontier model access for more than 30,000 employees. Travelers is notable for pairing one of the industry's largest internal claim organizations with aggressive external AI partnership, a combination that quietly refutes the idea that operational scale requires building your own models.
And at the insurtech end, Lemonade's Daniel Schreiber and Shai Wininger continue to run the industry's most instructive live experiment. Their model-agnostic stance, swapping frontier models in hours as better ones ship, is the purest expression of a thesis the rest of the industry is slowly converging on: the foundation models are commodities that improve on someone else's capital expenditure, and durable advantage lives in the workflow layer, the data, and the operational integration wrapped around them. Skyward Specialty's Andrew Robinson said the quiet part out loud with his arms race framing. When specialty carrier CEOs start using competitive language about underwriting AI on the record, the adoption question is settled. Only the execution question remains.
Step back from the individual voices and notice what changed in the questions, because the analyst side of these calls is its own data set. In 2023 and 2024, analysts asked insurance CEOs whether AI was real and what they were experimenting with. By the 2025 and 2026 cycles, the questions had become actuarial: whether broker AI adoption lets carriers compress acquisition expenses, whether token costs erode the efficiency capture, how AI-driven vulnerabilities change cyber underwriting, how savings flow through to combined ratio guidance. The sell side has stopped asking if and started modeling how much, and chief executives are now answering with numbers because vagueness on these calls has begun to read as weakness. That shift in the register of scrutiny, more than any vendor announcement, marks the moment a technology stops being a story and becomes a line item. Insurance AI crossed it sometime in the past eighteen months, and there is no crossing back.
The Winners: What the Leaders Have in Common
Naming winners in a market this early carries obvious risk, but the evidence base is now strong enough to support some conclusions, provided we define winning honestly: measured financial impact, disclosed publicly, sustained across more than one reporting period.
Aviva is the clearest carrier winner globally. Nearly 100 million pounds in claims transformation savings by its CEO's own account, more than 80 models in production, 98 percent of UK personal lines retail priced by machine learning, and AI cited as a driver in results that hit the company's 2026 targets a year early. Zurich has paired record profits with the most ambitious institutional commitment to AI research of any carrier and has named deployments, Nearmap in underwriting, Sixfold in Middle Market, that show the strategy operating at the workflow level rather than the press release level. Travelers has converted AI into underwriting income growth and claims efficiency at enormous scale while being unusually transparent about specific use cases. Chubb is winning in the way Chubb always wins, capturing efficiency without ever overpaying for fashion, and Greenberg's token cost comments suggest the economics are working comfortably.
Among the US personal lines giants, GEICO's underwriting automation and Progressive's long-standing analytical culture keep them at the front of the pack, and Allstate has disclosed measurable cycle time gains. Lemonade wins on a narrower but important definition: it has proven what an AI-native cost structure looks like, a loss adjustment expense ratio around half the incumbent norm, even as its overall economics remain a work in progress. In specialty, Markel's 113 percent underwriting productivity gain and Skyward Specialty's early and public commitment to underwriting AI mark them as the commercial market's pace setters. And the broker tier now has its own frontrunners: Marsh McLennan, positioned to monetize the AI cycle through both adoption and advisory, and Brown & Brown, whose disclosed eightfold pilot productivity gains and enterprise-wide commitment alongside Anthropic, McKinsey, and Accenture constitute the most aggressive distribution-side program on public record.
On the vendor side, the winners are the companies with named, referenceable, premium-weighted customer rosters: CCC across claims infrastructure, Sixfold and Cytora in underwriting intelligence, Cape Analytics as property data plumbing, Shift in fraud, and the emerging workflow automation leaders whose customers publicly credit them with measurable operational gains. The consistent pattern among winning vendors is depth over breadth: insurance-specific, workflow-specific, deployed inside the systems where decisions actually happen.
Equally instructive is what the winners have not done. None of them led with customer-facing chatbots, the most visible and least valuable deployment surface. None of them announced a moratorium while waiting for the technology to settle, a posture that sounds prudent and functions as forfeiture. None of them attempted a big-bang enterprise transformation; even Zurich's moonshot rhetoric sits on top of workflow-by-workflow deployments, and State Farm explicitly sequences minimum viable products in 12-to-18-month windows rather than attempting everything at once. And none of them, despite collectively employing tens of thousands of technologists, chose to build their document AI stacks from scratch. The restraint is as much a part of the playbook as the ambition.
Across every winner, carrier or vendor, four traits repeat. Proprietary data treated as the asset it is. Deployment at the point of decision rather than the point of summary. Workflow-level measurement rather than adoption-level measurement. And, tellingly, a willingness to buy or partner for the technology layer while concentrating internal effort on data, distribution, and judgment. Not one of the winning carriers built its stack alone. Aviva partnered with QuantumBlack. Zurich buys from Nearmap and Sixfold. Travelers partners with Anthropic and Cambridge Mobile Telematics. State Farm launched with OpenAI's Frontier platform. The winners are sophisticated buyers, not heroic builders.
The Patterns Behind the Laggards
I am not going to name the laggards, partly because public failure attribution in this industry is usually unfair to the people involved, and partly because the more useful truth is that lagging is a pattern, not a company. The patterns are well documented, and if you work inside a carrier or brokerage, you will recognize whether they describe your shop within the first sentence of each.
The first pattern is pilot purgatory. BCG's finding that only 7 percent of insurers have scaled their AI systems means the median large carrier is now several years and several million dollars into a portfolio of pilots that have never touched the P&L. MIT's Project NANDA research put the general enterprise version of this number at 95 percent of generative AI pilots failing to deliver measurable P&L impact, and identified the cause as a learning gap rather than a technology gap: tools that do not learn the organization's workflows, deployed by organizations that do not adapt their workflows to the tools. Datos Insights added the sharpest sentence in any 2026 industry report on this subject, observing of carriers pursuing identical document-summarization roadmaps that “The ones who think they're getting an edge are probably wrong”.
The second pattern is the hero project. MIT's research found more than half of enterprise AI budgets in 2025 flowed to sales and marketing pilots, high visibility and low return, while the durable returns came from unglamorous back-office automation. The insurance version of this is the customer-facing chatbot launched with a press release while the submission intake team still rekeys ACORD forms by hand. Budget allocated by internal lobbying rather than workflow economics is the most reliable leading indicator of a stalled program.
The third pattern is the internal build that quietly dies. Every large carrier has at least one: the in-house document extraction engine, the homegrown summarization tool, the model built by a talented data science team that worked beautifully in the demo and then decayed in production as document formats drifted, models were deprecated, the lead engineer left, and no one owned the maintenance. MIT's lead researcher noted that almost everywhere the research team went, enterprises were trying to build their own tools, and the data showed purchased solutions succeeding at roughly twice the rate. These failures rarely make the trade press because nobody issues a press release for a write-off, but every operations leader in this industry can privately name three.
The fourth pattern is measurement theater: programs that report adoption statistics, tool logins, and pilot counts to the board because they cannot report cycle times, loss ratio effects, or expense impacts. The ROI paradox that Gallagher Re's research surfaced lives here. If a program cannot state its baseline, it cannot state its return, and programs that cannot state their return get defunded in the first hard market for capital.
The fifth pattern is governance last. These are the carriers where deployment ran ahead of documentation, where the model inventory the state insurance department will eventually request does not exist, and where the answer to how the tool reached its conclusion is a vendor shrug or, worse, an internal shrug. The Earnix survey finding that fewer than one in three executives strongly believe their governance reviews keep pace with regulatory demands suggests this pattern is far more widespread than disclosure indicates. Governance-last programs do not fail visibly. They accumulate liability quietly and then fail expensively, in a market conduct exam or a litigation discovery process, at the moment of maximum inconvenience.
Running underneath all five patterns is a sixth phenomenon that MIT's research surfaced and that anyone managing insurance operations in 2026 will confirm off the record: the shadow AI economy. While official pilots stall in committee, employees are using consumer AI tools on their own initiative, pasting policy language and claims narratives into public chatbots because the sanctioned tools are worse or nonexistent. The productivity is real, which is why it persists. The data governance exposure is also real, which is why it should terrify compliance officers. Shadow AI is the tax an organization pays for the gap between its employees' needs and its official capability, and its prevalence inside a company is the most reliable diagnostic that the official AI program has failed the people it was meant to serve.
What unites all these patterns is that none of them is a technology failure. The models work. The failures are failures of ownership, allocation, integration, and maintenance, which is to say they are failures of approach, and approaches can be changed.
Build Versus Buy: The Question the Data Has Already Answered
Every one of the threads above converges on a single strategic question, so it deserves direct treatment: should an insurer build its AI capability internally or buy it from specialists? Two years ago this was a genuine debate. The evidence has since settled it for the overwhelming majority of the market, and it is worth walking through exactly why, because the reasoning matters more than the conclusion.
Start with the base rates. MIT Project NANDA's research found that purchased AI solutions succeed roughly 67 percent of the time against roughly 33 percent for internal builds, a two-to-one gap, and that pilots pairing internal specialists with external expertise reached a 67 percent success rate against 22 percent for IT-only internal efforts, a three-to-one gap. Only 5 percent of custom enterprise AI tools reach production at all. Mid-market organizations moving with vendors get from pilot to implementation in around 90 days; large enterprises building internally average nine months or longer. These are not insurance-specific numbers, but BCG's insurance-specific finding, 7 percent scaled, two thirds stuck piloting, tells you the industry is not beating the base rate. It is underperforming it.
Now consider why, because the why is what most carrier boards get wrong. The instinct behind building is that AI capability is strategic and strategic capabilities belong in-house. The premise is correct and the conclusion does not follow, because it misidentifies what the capability actually is. A production AI system in insurance is not a model. It is a living operational system: document pipelines that break when a wholesaler changes a submission format, guardrails that need retuning when a frontier model version is deprecated, evaluation harnesses that catch accuracy drift before an underwriter does, integrations into policy administration systems that were not designed to be integrated with, and audit trails that satisfy a regulator applying the NAIC model bulletin. Maintaining that system is a full-time, specialized, continuously evolving discipline. It is LLM operations, and it is a categorically different skillset from the actuarial and data science excellence that large carriers genuinely possess. Being world-class at pricing models does not make an organization competent at production document AI maintenance any more than being a world-class driver makes someone a mechanic. The carriers that concede this are not admitting weakness. They are reading their own org charts accurately.
The economics compound the point. A specialist vendor amortizes the cost of that maintenance discipline, the engineers, the evaluation infrastructure, the model migration work, across dozens of customers. A carrier building internally bears it alone, forever, against a technology substrate that changes quarterly. When Lemonade, the most technically capable insurance operator on earth, tells investors it swaps underlying frontier models in hours precisely because it refuses to be married to any of them, it is describing the maintenance treadmill every internal build must run without Lemonade's engineering density. And when Moody's, CCC, Applied, and Guidewire, companies whose core business is insurance technology, choose to acquire deployed AI capability rather than build it, they are pricing that treadmill honestly.
The talent market delivers the same verdict from another direction. An internal build competes for machine learning and LLM operations engineers against technology companies that pay in equity denominated in AI-boom valuations, offer work on frontier problems, and concentrate similar engineers by the hundreds. Insurance carriers can win some of that talent, the leaders demonstrably have, but they cannot win enough of it to staff and permanently retain a production AI engineering organization at every carrier that would need one, and the departure of two key engineers can strand an internal system in a way no vendor relationship replicates. The total cost accounting boards should demand, and rarely receive, includes not just the build but the forever after: the model migrations every time a frontier lab deprecates a version, the evaluation infrastructure, the security reviews, the prompt and pipeline maintenance as document formats drift, and the key-person risk premium. Priced honestly across a five-year horizon, the internal build is almost never the cheaper option even before its two-to-one disadvantage in success probability is applied, and pricing things honestly across time is supposed to be the one discipline this industry owns.
There is also a distinction worth preserving that the build-versus-buy framing can blur. Large insurers do have world-class quantitative organizations, and nothing here argues otherwise. Actuarial science, pricing sophistication, catastrophe modeling, and portfolio analytics are genuine internal strengths and should remain internal, because they encode the risk judgment that is the company. But LLM operations and production document AI maintenance is a categorically different discipline from any of those, different tooling, different failure modes, different talent, different cadence, and organizational excellence in one is routinely mistaken for readiness in the other. The most sophisticated carriers in the world have already demonstrated they understand the difference: they kept their actuaries and bought their document AI. The mid-market carrier convinced it should build because it has strong pricing talent is misreading the leaders it believes it is imitating.
None of this means carriers should outsource judgment, data, or accountability. The winning pattern, visible in every carrier named in this piece, is precise about the division: own your data, own your risk appetite, own the validation duty that ASOP No. 56 and the regulators place on you regardless of vendor, and buy the technology layer from specialists whose entire existence depends on keeping it accurate, current, and maintained. The MIT research phrase for the winning 5 percent was tightly scoped initiatives, domain-specific focus, and smart partnerships. That is not a compromise position between building and buying. It is simply what winning has turned out to look like.
One more implication follows, and it is the one the industry still resists saying plainly. Because the foundation models are available to everyone and improving on someone else's capital expenditure, using AI is not a strategy and confers no differentiation whatsoever. Announcing that your company uses AI in 2026 carries exactly the informational content of announcing in 1996 that your company uses computers. Differentiation lives in what the winners actually have: proprietary data, distribution, underwriting judgment, and the operational discipline to deploy purchased capability where decisions get made. Carriers should spend accordingly.
The Regulatory Layer Nobody Gets to Skip
Whatever architecture a carrier chooses, it now operates inside a hardening governance perimeter. The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, had been adopted by more than 26 states as of early 2026, giving state insurance departments a ready framework for governance expectations, and Colorado's algorithm and predictive model regulations have gone further still. The message regulators are sending is uniform: deployment without documented governance is not a defense, and neither is pointing at a vendor. Nor are regulators operating blind: NAIC surveys have already documented AI use in claims estimation, triage segmentation, and data collection across home and auto lines, per Claims Journal's reporting, which means state departments know precisely where the models sit before the market conduct questions arrive.
The industry appears to have heard the first half of that message and not the second. Earnix's 2026 trends report, based on a survey of 400 global insurance executives, found that 92 percent of insurers now conduct formal AI governance reviews on a regular cadence, yet fewer than one in three executives strongly agree those reviews are sufficient to keep pace with regulatory demands, and 38 percent cite regulatory and legal exposure as their primary ethical concern in deployment. Governance activity is nearly universal. Governance confidence is scarce. That gap is itself an argument for working with specialists who maintain compliance artifacts, audit trails, and model documentation as a core product feature rather than an afterthought bolted onto an internal build by whoever has spare capacity that quarter.
What Comes Next: The Shape of the Sorted Market
Forecasting is cheap, so let me confine it to extensions of trends already visible in the data rather than speculation, and attach the evidence to each.
The vendor market will consolidate hard, and soon. The acquisition pattern is already established, Moody's, CCC, Applied, Guidewire, and the funding data shows capital concentrating in category leaders while the long tail of undifferentiated wrappers starves. Datos Insights' observation that most carriers are running identical document-summarization roadmaps implies its own conclusion: identical capabilities become procurement line items, and procurement line items consolidate to two or three vendors per category. Carriers should be underwriting their vendors' survivability the way they underwrite any counterparty, favoring named customer rosters, premium-weighted deployments, and strategic backers over demo quality.
Underwriting autonomy will expand faster than the industry is publicly comfortable admitting. Bind-capable systems are in production today at some of the most conservative carriers in the market. The projections tracked across industry analyses, adoption moving from 14 percent toward 70 percent by 2028, will prove directionally right even if the numbers wobble, because the economics of quote-to-bind compression are too large for competitive markets to leave unclaimed. The constraint will be governance capacity, not technology.
The expense ratio gap will become visible in public financials. Through 2025, AI's impact hid inside combined ratios influenced by a dozen larger forces. But when one carrier's claims operation runs at Lemonade's loss adjustment expense structure, or captures Aviva's indemnity precision, and a peer's does not, the divergence compounds quarterly, and analysts have already started asking the question on earnings calls: the exchanges with Greenberg about token costs and broker expenses were not curiosity, they were early modeling. Within two years, AI operational maturity will be a standard component of how the sell side differentiates carriers, and the laggards will discover that pilot purgatory has a share price.
Distribution economics will be repriced, and February 9 was the market's opening bid. Greenberg said it plainly: intermediation costs are excessive and technology will bring them down. The brokerages and MGAs that survive that repricing with their margins intact will be the ones that used this window to make themselves demonstrably cheaper to trade with and demonstrably better at packaging risk, which is to say the ones that treated AI as a defense of their commission rather than a threat to their staff. The global brokers have already shown what the response looks like at scale; the middle of the distribution stack now has both the warning and the template.
And the regulatory perimeter will keep tightening, state by state, with governance documentation becoming as standard a filing artifact as rate support. The 26-plus state adoption of the NAIC bulletin in barely two years is among the fastest regulatory diffusions in modern insurance history, and nothing about the political environment suggests deceleration. The winners will treat compliance artifacts as a product feature they procure and maintain, not a project they improvise.
The Takeaway
The AI adoption question in insurance is over. Ninety percent of insurers are on the journey per Conning, the pilot phase is formally declared dead by Datos Insights, and the CEOs of Aviva, Zurich, Chubb, and Travelers are discussing AI on earnings calls in the language of combined ratios and reported savings rather than the language of exploration. What remains open is the distribution of outcomes, and the early evidence says that distribution will be brutal: a small group of carriers converting AI into audited nine-figure results while the majority remain stuck in patterns, pilot purgatory, hero projects, dying internal builds, measurement theater, that are organizational rather than technological and therefore entirely avoidable.
The winners' playbook is no longer secret. Treat proprietary data as the asset. Deploy where decisions are made, not where demos look good. Measure at the workflow level. Buy the technology layer from domain specialists who live and die by maintaining it, partner deeply rather than transactionally, and concentrate internal effort on the things no vendor can supply: your data, your judgment, your accountability. The two-to-one success gap between buying and building is not a talking point. It is the most expensive lesson of the first generation of enterprise AI, already paid for by the 95 percent, available free to everyone who comes after.
Insurance has absorbed every prior technology wave slowly and then completely. This one is being absorbed quickly and unevenly, and the unevenness is the opportunity. The carriers, MGAs, and brokerages that internalize the buy-side discipline now, while most of their competitors are still funding pilots and press releases, will spend the next decade with a structural expense and decision-quality advantage that no amount of catch-up spending fully closes. The technology is available to everyone. The approach, it turns out, is the differentiator.
Fabio Faschi is an Enterprise AI Solutions and Sales leader helping carriers, MGAs, and brokerages put artificial intelligence to work across underwriting, claims, and distribution. A National Producer, Board Member of the Young Risk Professionals New York City chapter, and Committee Chair at RISE, he brings over a decade of insurance industry experience and has built and scaled more than a dozen national brokerages and SaaS-driven insurance platforms. He is the founder of ScholarusAI.com and Hogglet.com for Enterprise AI transformation and risk management. Fabio's expertise has been featured in publications like Forbes, Consumer Affairs, Realtor.com, Apartment Therapy, SFGATE, Bankrate and Lifehacker.
