ANALYSISINSURANCE
AI in Insurance: Record Adoption, Lagging Industrialization
Insurance is the economy's most AI-equipped sector — yet only 10% of insurers have industrialized it, while catastrophe losses top $100bn for the sixth year. Six use-case families, four breaking points, five 90-day decisions.
The best-equipped sector in the economy
Let us start with what sets this report apart from the previous ones. After retail, then transport and logistics — two sectors lagging on adoption —, insurance presents the opposite picture: it is one of the most AI-equipped sectors in the entire economy. In the United States, 33.9% of finance and insurance businesses report using AI to produce their services, against 19.8% on average across all sectors (US Census Bureau). In the United Kingdom, the regulator measures 95% adoption among the insurers responding to its survey — the highest rate in financial services (Bank of England / FCA).
Continental Europe confirms the picture, with its nuances by line of business. According to the European supervisor's survey of 209 undertakings, 50% of non-life insurers and 24% of life insurers already use AI, and three-year intentions push those figures toward 80% and 63% (EIOPA). On generative AI the shift has been remarkably fast: by early 2026, nearly two thirds of the 347 undertakings surveyed across 25 countries use it actively — 64% of reported use cases on internal processes, led by data extraction from invoices and medical reports, and already 36% facing the customer (EIOPA). In France, the national statistical survey excludes the financial sector from its scope — but the French supervisor documents the same dynamic, and the national average of 18% of AI-using businesses in 2025 sets the order of magnitude that banking and insurance comfortably exceed (INSEE).
Insurance adopts AI more than the rest of the economy — everywhere it is measured
This level of equipment is no accident. Insurance has been a data business from its origins: mortality tables, loss statistics, risk segmentation — actuarial science is a science of prediction, and machine learning is its natural extension. The American line-of-business surveys confirm it: 88% of auto insurers and 84% of health insurers already use, test or explore AI and machine learning (NAIC). The sector's question is therefore not "should we go": nearly everyone is already there.
The real question lies elsewhere, and it structures this entire report: adoption has not yet produced the value. Using AI somewhere in the organization has become commonplace; industrializing it across a complete chain — underwriting, claims, fraud, customer relations — with a measured result remains the exception. It is this gap between record equipment and minority industrialization that the maturity figures, further on, measure precisely. And it is widening at the worst possible moment: claims costs are not slowing down.
A hundred billion a year, six years running
If insurance is accelerating on AI, it is not out of taste for technology: its economic equation is tightening. Natural catastrophes cost $107 billion in insured losses in 2025 — the sixth consecutive year above the hundred-billion threshold — after $141 billion in 2024; the Los Angeles wildfires alone cost $40 billion insured, the costliest fire event in history (Swiss Re Institute). And 2025 is the most expensive year ever recorded for so-called secondary perils — floods, storms, wildfires — precisely those that strike with frequency (Munich Re).
France lives the same pressure at its own scale. Natural events cost €3.87 billion to French insurers in 2024, and the natural-catastrophe regime had to raise its surcharge from 12% to 20% on home and business policies on January 1, 2025 (France Assureurs). Add fraud, the system's other leak: €946 million detected in France in 2025, up 5% (ALFA); a record £1.16 billion in the United Kingdom (ABI); and an estimated $308.6 billion annual cost to American consumers (Coalition Against Insurance Fraud) — bearing in mind that detected fraud is, by construction, only the visible part.
A third strain is less visible but structural: the workforce. The American insurance sector was set to lose about 400,000 workers to retirement by 2026 — underwriters and claims adjusters first (US Bureau of Labor Statistics). Actuarial expertise and claims handling rest on skills that take years to build; when they leave, they are not replaced at the pace claims are rising. The US combined ratio did recover to 92.2 in 2025, its best level in a decade — but that recovery owes much to premium increases, whose acceptability has limits (AM Best). The conclusion is arithmetic: when losses rise structurally and expertise grows scarce, every hour of claims handling, every missed fraud file, every pricing error costs more than before. That is exactly AI's home turf.
Adoption is not industrialization
Back to the gap announced in the opening, because it can now be measured with precision. Of 344 property and casualty insurance executives surveyed worldwide, only 10% lead organizations that have genuinely industrialized AI — the "trailblazers" — while 60% remain at the exploration or proof-of-concept stage (Capgemini). The performance gap between the two groups is spectacular: up to 21% higher revenue growth and 51% higher share-price performance over three years for the former.
Why do so few cross the line? The same surveys give the answer, and it is not technological: 42% of insurers track no performance indicator for their AI initiatives, and 55% report no clear return on investment (Capgemini). You cannot industrialize what you do not measure. The warning counts double for the autonomous-agent wave: more than 40% of agentic AI projects, across all sectors, will be abandoned by the end of 2027 for lack of demonstrated business value and risk controls (Gartner). Across UK financial services — insurers first among them —, 55% of use cases already involve some automated decision-making, but only 2% amount to fully autonomous decisions (Bank of England / FCA): full autonomy remains, rightly, the exception.
Record equipment, minority industrialization — the gap that will set tomorrow's ranking
While the majority experiments, capital has already chosen: global insurtech funding turned up again in 2025 — $5.08 billion, the first increase since 2021 — and 77.9% of last-quarter volumes went to AI-centered companies (Gallagher Re). Customers, too, are ahead of their insurers: nearly 70% say they have used consumer generative-AI tools on their own initiative when buying insurance, while remaining vigilant about privacy and attached to human contact in difficult moments (Geneva Association). The environment — capital, customers, competitors — is moving; the internal maturity gap is becoming the sector's main strategic risk.
Six use-case families, one common logic
Where does AI produce measured results in insurance? The documented deployments — official releases, annual reports, regulatory filings — outline six use-case families, each backed by published, verifiable figures. None requires inventing a technology: all require prepared data and a chosen scope.
Claims handling. This is the family with the most spectacular before-and-after gap, because a claim concentrates everything: documents, photos, expertise, compliance, customer relations. Allianz has put its first agentic claims AI into production in Australia — seven specialized agents that check coverage, validate the weather event, screen for fraud and calculate the payout. On its first scope, simple capped claims — spoiled food after a power outage —, claim processing and settlement time falls by 80%, from several days to one day, sometimes hours — and the payout decision is never automated (Allianz). At Lemonade, about 55% of claims are settled end-to-end without human intervention, and the cost of handling a pet claim has fallen from $44 to $14 in four years (Lemonade). Established carriers follow the same slope: at Travelers, more than half of claims are now eligible for straight-through processing, and claims call-center staffing has been cut by about a third (Travelers).
Underwriting. This is the family of the long document: medical files, risk statements, questionnaires — the raw material exists, waiting to be read. Zurich cut the turnaround time of its UK commercial underwriting by 78% in two years, combining document AI with aerial imagery for property assessment (Zurich). Aviva launched in late 2025, after eighteen months of testing on a thousand files, a generative-AI tool that summarizes medical reports of more than ninety pages for individual life underwriting — with the underwriter keeping the final decision (Aviva). And the reachable order of magnitude is known: at China's Ping An, 94% of life policies are underwritten with a decision rendered in seconds (Ping An).
Fraud detection. This is the family with the most directly quantifiable return: every fraudulent file set aside is a net saving. The global reference is French: Shift Technology, which has analyzed more than 4 billion policies, claims and documents for more than 120 insurers across some thirty countries, puts the fraud uncovered in the United States alone at $2 billion in 2025 (Shift Technology). Generali France, a client since 2016, has doubled its detection in five years on construction claims (Generali). At the scale above, Ping An saves more than 10 billion yuan of claims per year — about $1.45 billion — through intelligent detection, for the third consecutive year (Ping An).
Customer relations. This is the fastest family, because the cost per contact is already known to every customer-service executive. The ceiling is documented: at Ping An, AI agents handled about 1.7 billion interactions in 2025 — 80% of the group's entire customer service (Ping An). In France, health insurer Alan instantly resolves 40% of member requests — 70% of simple coverage questions — with satisfaction held at 4.5 out of 5 and a stable support team despite a third more members (Alan). In the United States, nearly all of the some 50,000 daily messages sent to Allstate's claims customers are drafted by generative AI, with advisers now stepping in only to verify (Allstate).
Pricing and risk modeling. This is the family closest to the actuarial core — and the most regulated, as we shall see. The tooling has industrialized: French insurtech Akur8, specialist in transparent AI pricing, equips more than 250 insurers across some forty countries, and more than 3,000 actuaries use it daily (Akur8). Behavioral data is changing the very base of the calculation: Progressive feeds its models with tens of billions of telematics driving miles per year, and added 3.7 million policies in 2025 (Progressive). Facing climate-driven losses, fine-grained modeling of secondary perils is becoming this family's other decisive use — it is what will decide the accuracy of tomorrow's premiums.
Internal support and cross-functional work. This is the quiet family — and often the first profitable deployment, because it touches neither pricing nor the payout decision. AXA opened its secure generative AI to the whole group: 65,000 employees had accessed it by late 2024, including 16,000 regular weekly users, on uses ranging from file summaries to underwriting preparation (AXA). Covéa has six generative-AI tools in daily use — including automatic claims-file summaries that save up to fifteen minutes per complex file (Covéa). Travelers counts more than 20,000 regular users of its internal tools (Travelers); CNP Assurances is working through some thirty use cases in test-and-learn mode (CNP Assurances).
Two speeds of return — and what conditions them
Six families, one common trait: the consolidated file. The insurer that structures its policy, claims and customer-exchange data serves its policyholders — and makes each of these families accessible in the order of its choosing.
The proof, in French
The classic objection — "these are the results of American or Asian giants" — holds even less here than elsewhere: two of the world references in insurance AI are French companies, and the major domestic groups document their deployments under the same rules as everywhere: a narrow scope, prepared data, a measured result. Four cases suffice to show it.
A fifth French signal deserves attention, because it concerns method rather than technology: in May 2026, MAIF signed with six representative unions a company agreement dedicated to AI — a commitment to no AI-related economic redundancies, training for all 9,000 employees, and a monitoring commission attached to the works council (MAIF). In a business whose value rests on experienced claims handlers and underwriters, bringing the teams along is not a nicety: it is a condition of the result.
The regulatory rail: life and health pricing is already classified
Insurance has a particularity that neither retail nor logistics knows to this degree: part of its AI use cases is already explicitly classified as high-risk by law. The European AI regulation lists risk-assessment and pricing systems in life and health insurance in its Annex III; the May 2026 European agreement postponed the application of those obligations to December 2, 2027 — a delay, not a dispensation (AI Act). In France, the ACPR — expected to become the supervisory authority for these uses — opened a public consultation on algorithmic fairness in the summer of 2026: the supervisor is preparing its doctrine ahead of the deadline (ACPR).
The rest of the framework is already in force, and it covers nearly all six families. The European supervisor published in August 2025 its opinion on AI governance — data, traceability, fairness, explainability, human oversight — applicable to all systems, including those the AI Act does not classify (EIOPA). Digital resilience has applied since January 2025 to all European insurers, AI and cloud providers included (DORA). And the GDPR strictly frames the health data that feeds underwriting and claims handling (CNIL). The United States is moving along the same slope through its own channels: the model bulletin on AI has been adopted by 24 states plus the District of Columbia, Colorado has enforced its predictive-model governance since October 2025, and New York requires proof of no proxy discrimination in underwriting (NAIC).
A framework already in place — and a single explicitly designated "high-risk" regime
None of these rules forbids anything: they separate the light-compliance families — internal support, customer relations, fraud — from the one that will demand a complete file. Choosing the first use case is therefore also choosing a regulatory regime.
A project's journey — and its four breaking points
AI projects that stall in insurance do not hit the walls of retail or logistics: they hit their own. Superimposing the field reports — the 10% who industrialize as much as the 60% who experiment without a framework (Capgemini) — four breaking points specific to the sector emerge. Each is prevented by a decision that costs nothing, provided it is taken before launch.
- Breaking point 1 — the pilot before the file. The organization launches an underwriting or claims assistant on top of generations of core systems that do not talk to each other: policies in a legacy administration system, claims in another, documents in the archive, exchanges in mailboxes. The tool computes on sand and the team concludes "AI doesn't work". The successful cases do the opposite — Aviva tested its medical summarization for eighteen months on a thousand prepared files before opening the service (Aviva). The countermeasure: start by consolidating the file of the one chosen case, even though it is the least visible workstream.
- Breaking point 2 — the opaque model facing actuarial validation. Insurance is the only sector where every model that touches risk goes through formal validation — actuarial, risk management, audit. The brilliant pilot built on an unexplainable model stops precisely there, after months of development: impossible to document what the supervisor already requires — explainability, traceability, human oversight (EIOPA). The countermeasure: require explainability when choosing the tool, not at the end of the journey — it is the positioning of Akur8's transparent AI pricing, and Aviva's rule, where the underwriter keeps the final decision.
- Breaking point 3 — health data and fairness discovered after development. The case touches pricing or health data, and the regulatory review comes afterwards: proxy discrimination untested, information of individuals missing, fragile legal basis — exactly what the AI Act classifies as high-risk in life and health, what GDPR frames, and what supervisors are turning into doctrine, from algorithmic fairness to proof of no proxy discrimination (ACPR; NYDFS). The countermeasure: build the compliance review into the case's costing — and keep life and health pricing for when the governance is ready.
- Breaking point 4 — automating the moment of truth. The claim is the instant when insurance proves its worth — and the one contact customers fear losing: nearly 40% worry about the disappearance of human contact, at the very moment they adopt AI for everything else (Geneva Association). The organization that automates end-to-end all the way to claim denial gains minutes and loses trust — complaints, churn, regulatory attention; the market knows it: across UK financial services, only 2% of use cases amount to fully autonomous decisions (Bank of England / FCA). The countermeasure: mandatory human review of every unfavorable decision — Allianz built it into the very workings of its claims agents — with full automation reserved for simple requests and favorable outcomes.
Each breaking point is prevented by a decision that costs nothing — provided it is taken before starting, not after.
What the executive must decide — and when
Neither a three-year transformation plan nor a prudent wait while the reference texts move: the trajectory that works, judging by the deployments carried through, holds in five decisions over ninety days — each with its indicator, each answering one of the breaking points above.
- Choose a family, not a technology. Among the six documented families, take the one where your current cost is most measurable — customer relations, fraud and internal support are the most frequent entry points, and the least exposed to regulation. Indicator: a validated starting-cost figure (cost per contact, minutes per file, underwriting turnaround, fraud detected).
- Audit the file of the chosen case — not the whole company. Policy history, claims files, attachments, customer conversations: does the data exist, is it consolidated, since when? Indicator: a written verdict within fifteen days.
- Run the case's compliance review before signing. Health data, information of individuals, explainability, proxy discrimination: which regime — AI Act, GDPR, supervisory doctrine — applies to this precise case, and what must be documented? Indicator: a written review attached to the costing — before signature.
- Cost in total cost — teams included. License, integration with policy and claims systems, file consolidation, support for claims handlers and underwriters — with the teams in the loop from day one, in the manner of the agreement signed at MAIF. Indicator: a budget where data preparation and human support are visible on page one.
- Set the decision date and instrument measurement from day one. Industrialize or stop, at twelve or twenty-four weeks depending on the family; the ending figure is compared with the starting figure, on the same scope, with human review maintained until quality is demonstrated. Indicator: the date is on the executive committee's agenda, the before/after dashboard is live from week 1.
The third scenario — doing nothing — also has a price, even though it appears on no budget. The 10% of insurers who industrialized show up to 21% higher growth and a total shareholder return six times higher (Capgemini; McKinsey); catastrophe losses have exceeded $100 billion for the sixth consecutive year (Swiss Re Institute); and every euro of undetected fraud is paid for good. In a sector where nearly everyone is already equipped, the gap is no longer made on adoption: it is made on discipline — not more resources, more method.
Method
This report draws on three materials. First, public statistical surveys and studies by reference institutions — US Census Bureau, NAIC, Bank of England / FCA, EIOPA, INSEE, ACPR, CNIL, France Assureurs, ALFA, ABI, Swiss Re Institute, Munich Re, AM Best, Capgemini, Gallagher Re, Geneva Association, Deloitte, Gartner, McKinsey — whose full references appear below. Second, the YPrtnrs Institute knowledge base: more than 330 verified company cases and data points, including nearly sixty specific to insurance, each tied to an original publication. Third, YPrtnrs Institute's own analysis, whose estimates are systematically flagged as such and never presented as measurements. Company figures come from official communications, annual reports, regulatory filings or statements by their executives as reported by the trade press cited in the resources; projections (McKinsey, Deloitte, Gartner) are dated and explicitly designated as projections.
Resources
ACPR. (2025, September). Surveillance de l'IA dans le secteur financier — réunion de place; ACPR. (2026, July). L'ACPR lance une consultation publique sur l'équité algorithmique. Autorité de contrôle prudentiel et de résolution.
Allianz. (2025, November). When the storm clears, so should the claim queue — Allianz launched its first agentic AI to automate claims [corporate article].
Allstate. (2025, February). Chief information officer's statements to the Wall Street Journal, as reported by The Register: Insurance giant finds claims rep that gives a damn (it's AI).
AM Best. (2026, March). Best's Special Report: US Property/Casualty Industry Sees Underwriting Income Nearly Triple to $61 Billion in 2025.
Association of British Insurers. (2025, November). Fraudulent insurance claims continue to top £1 billion.
Aviva plc. (2025, November). Aviva to launch groundbreaking AI Underwriting Tool [press release].
AXA. (2023, July). AXA propose une IA générative sécurisée à ses collaborateurs [press release]; Le Monde Informatique. (2024, December). Comment Axa avance dans la GenAI.
Bank of England, & Financial Conduct Authority. (2024, November). Artificial intelligence in UK financial services — 2024.
Capgemini Research Institute. (2026, May). World Property & Casualty Insurance Report 2026 [summary release].
CNIL. (2025, February). IA et RGPD : la CNIL publie ses nouvelles recommandations pour accompagner une innovation responsable. Commission nationale de l'informatique et des libertés.
CNP Assurances. (2025, January). Interview with La Revue du Digital: « Une véritable envie de faire évoluer tous les métiers avec l'IA générative ».
Coalition Against Insurance Fraud. (2022, August). The Impact of Insurance Fraud on the U.S. Economy.
Colorado Division of Insurance. (2025). SB21-169 — Protecting Consumers from Unfair Discrimination in Insurance Practices [amended Regulation 10-1-1]; New York State Department of Financial Services. (2024, July). Insurance Circular Letter No. 7.
Covéa. (2025, September). Les usages de l'IA générative chez Covéa [corporate report].
Deloitte Center for Financial Services. (2025, October). 2026 Global Insurance Outlook.
EIOPA. (2024, April). Report on the digitalisation of the European insurance sector; EIOPA. (2025, August). Opinion on AI governance and risk management; EIOPA. (2026, February). Survey on Generative AI shows swift but cautious adoption among Europe's insurers. European Insurance and Occupational Pensions Authority.
Eurostat. (2025, December). 20% of EU enterprises use AI technologies [ICT survey 2025]; Statistisches Bundesamt — Destatis. (2024, November). Jedes fünfte Unternehmen nutzt künstliche Intelligenz.
Finyear. (2024, September). Akur8 boucle sa Série C à 120 millions de dollars menée par One Peak [after Akur8's release].
France Assureurs. (2026, January). L'assurance des événements naturels en 2024; ALFA — Agence de lutte contre la fraude à l'assurance. (2026, July). 2025 detected-fraud figures [as reported by News Assurances Pro].
Gallagher Re. (2026, February). Global InsurTech Report Q4 2025.
Gartner. (2025, June). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 [press release].
Geneva Association. (2025, November). Gen AI in the Insurance Customer Journey [survey of 6,000 insurance customers, six countries].
INSEE. (2026, July). Les TIC et le commerce électronique dans les entreprises en 2025 (Insee Première No. 2120). Institut national de la statistique et des études économiques.
Lemonade, Inc. (2026, February). Annual Report 2025 (Form 10-K, SEC filing); Lemonade, Inc. (2026, February). Q4 2025 Shareholder Letter.
MAIF. (2026, May). Company agreement on artificial intelligence [as reported by L'Assurance en Mouvement].
McKinsey & Company. (2023, June). The economic potential of generative AI; McKinsey & Company. (2026, July). How AI will reshape the economics of insurance: A CEO's guide to strategy.
Munich Re. (2025, January). Natural disaster figures 2024; Munich Re. (2026, January). Natural disaster figures 2025 [NatCatSERVICE].
NAIC. (2023-2026). Line-of-business AI/ML surveys (auto, homeowners, life, health) and Implementation of NAIC Model Bulletin: Use of AI Systems by Insurers [map as of April 1, 2026]. National Association of Insurance Commissioners.
Ping An Insurance (Group) Company of China. (2026, March). Ping An Reports 2025 Performance [annual results release].
Progressive Corporation. (2026, March). 2026 investor meeting [as reported by Repairer Driven News].
Shift Technology. (2026). Fraud detection — key figures [official site]; Microsoft. (2025, April). Shift Technology helps hundreds of insurers fight claims fraud using Azure OpenAI Service [customer story]; Generali France. (2025, April). Generali renouvelle sa confiance en Shift Technology [press release].
Swiss Re Institute. (2025, December). 2025 marks sixth year insured natural catastrophe losses exceed USD 100 billion [sigma, press release].
Travelers. (2026, January). Fourth-quarter 2025 results and related statements [as reported by Carrier Management]; European Union. (2024-2026). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence [AI Act], Annex III and the "Digital Omnibus" agreement of May 6, 2026; EIOPA / ESAs. (2025). Digital Operational Resilience Act (DORA) — applicable since January 17, 2025; US Bureau of Labor Statistics. (2023). Insurance-sector attrition projections [as reported by Insurance Business America].