T3 Research · AI Index Series · 2025 Edition
AI IN
INSURANCE
INDEX 2025
A comprehensive analysis of AI adoption, investment trends, use cases, performance impacts, and regulatory considerations across the global insurance industry. Industry data through Q4 2025.
GenAI Adoption Has Hit an Inflection Point: 90% of Insurers Are Now Evaluating or Deploying
Source: Conning Survey on AI and Insurance Technology: The C-Suite Verdict, June 2025
Source: Conning Survey on AI and Insurance Technology, June 2025
Overall AI adoption jumped from 8% to 34% year-on-year. LLMs went from 18% to 63% in a single year. The driver is measurable operational returns from claims correspondence, policy summarisation, and underwriting support.
The NAIC's 2025 survey confirmed the breadth of adoption: 92% of health insurers, 88% of auto insurers, and 70% of home insurers report current or planned AI use. Even life insurance now reports 58%.
The speed of change: Full AI adoption jumped from 8% to 34% in a single year — a 26 percentage point increase. The industry is compressing what typically takes a decade into roughly 18 months.
How Insurers Split AI Spend: Traditional ML Leads, But Agentic AI Is Already in Budget
Source: IBM Institute for Business Value, Insurance in the AI Era, September 2025
Traditional ML holds the majority of AI spend, reflecting years of investment in pricing engines, fraud classifiers, and risk models already in production. GenAI at 21.5% is the fastest-growing layer.
Agentic AI at 11.8% of budget is already significant. IBM IBV data shows 77% of agentic use cases over the next year will concentrate in claims, making claims the primary proving ground for autonomous AI decision-making.
The governance implication is immediate. Insurers need frameworks that cover legacy ML models, GenAI document workflows, and agentic claims chains simultaneously — different risk profiles requiring different oversight protocols.
The Core Insurance Use Cases Are in Production, Not Pilot
Source: BCG / Datagrid, Insurance AI Agent Statistics, December 2025
Fraud detection leads because ROI is direct and measurable. AI fraud detection shows a 65% improvement in detection capability and a 60% reduction in overpayment rates versus rules-based systems. Deloitte projects AI-driven fraud detection could save P&C insurers $80–160 billion by 2032.
Claims processing has moved firmly to production at leading carriers. AI now handles 70–90% of simple claims in straight-through fashion, with decisions in minutes rather than weeks.
Underwriting support is the next high-growth area: currently at 58% adoption, projected to reach 70% by 2028, with 81% of insurance CEOs backing it as a top GenAI investment priority.
What separates firms that scale from those that stall: The insurers moving fastest from pilot to production share one characteristic that is rarely the headline; they invested early in Responsible AI infrastructure: bias testing, model validation frameworks, explainability documentation, and human oversight protocols. Without these, every deployment hits a compliance or governance wall before it reaches the business line at scale.
Back-Office AI Delivers Faster Returns Than Front-Office at Most Carriers
Source: Roots.ai, State of AI Adoption in Insurance 2025
The fastest returns in insurance AI are often invisible to policyholders: document extraction, contract processing, and knowledge management. Insurance is document-intensive by nature, and LLMs applied to this layer deliver efficiency gains with relatively low deployment risk.
53% of insurers are deploying AI to increase speed-to-quote. AI underwriting support is the highest-priority deployment for 2025, driven by the direct link between quoting throughput and premium growth.
The pattern is consistent across financial services: AI momentum starts in engineering and operations, then scales to claims and underwriting once the technical foundation is established.
The AI Testing Gap: Nearly a Third of Insurers Still Do Not Test Their Models for Bias
Source: Roots.ai, State of AI Adoption in Insurance 2025
Source: NAIC AI/ML Survey Report, January 2025
The EU AI Act classifies AI for risk assessment and pricing in life and health insurance as high-risk under Annex III. Obligations apply to both providers and deployers. Requirements before deployment: conformity assessments, human oversight protocols, incident logging. Non-compliance fines: up to 3% of global turnover or 15 million euros, whichever is higher.
Only 22% of insurers have AI solutions in full production. Of those, the NAIC's January 2025 survey found nearly one third still do not regularly test for bias or discrimination, despite the NAIC's AI Model Bulletin recommending this since December 2023.
In insurance, AI errors carry direct consumer harm potential. A model that miscalculates a risk premium or incorrectly denies a claim is a conduct failure with legal consequences under FCA Consumer Duty, Solvency II, and the EU AI Act.
EIOPA's 2025 consultation set a clear supervisory expectation: boards and CROs are accountable for AI decision quality in the same way they are for model risk under Solvency II. Firms that close this testing gap now will be ahead of enforcement in 2026 and 2027.
Claims Automation Is Delivering Real P&L Impact, Concentrated in Eligible Cases
The 75% time reduction applies to AI-eligible claims: high-frequency, low-complexity motor, home contents, and simple health cases. Complex liability, large commercial, and multi-party claims still require human expertise.
The Deloitte $80–160 billion fraud savings projection is contingent on multimodal AI integration across the full claims lifecycle. Most carriers today deploy point solutions. Closing the gap between single-function deployments and integrated architecture is the primary implementation challenge for 2025–2027.
The strategic question for 2026 is not whether to automate claims, but how to govern the boundary between what AI decides autonomously and what triggers human escalation. The EU AI Act's human oversight requirements make this a regulatory question as much as an operational one.
Talent and Data Quality Are the Binding Constraints, Not Technology
Source: Roots.ai, State of AI Adoption in Insurance 2025 (n=240 insurance executives)
Source: BCG / Datagrid, Insurance AI Agent Statistics, December 2025
Cultural resistance in insurance is not generic technophobia. The actuarial profession has built its authority on statistical rigour and explainability. AI models challenge that authority directly — the concern is principled: ceding professional judgment to opaque systems with unexplainable outputs.
Conning's 2025 survey cited Bureau of Labor Statistics data showing a notable decline in lower-skilled clerical roles alongside significant growth in higher-paid, higher-skilled insurance occupations from 2002 to 2024. AI is accelerating this structural shift rather than simply replacing headcount.
Legacy system fragmentation is the deepest structural barrier. McKinsey's research suggests that upgrading legacy infrastructure delivers a 41% reduction in per-policy IT costs and a 40% increase in operational productivity.
Global AI Investment: The US Leads by an Order of Magnitude, Shaping Every Insurer's Model Stack
Source: Stanford HAI AI Index 2025
Source: ScienceSoft, AI for Insurance Claims, 2025
US private AI investment at $109.1B is nearly 12 times China's and 24 times the UK's. The foundation models, infrastructure, and tooling that global insurers deploy are overwhelmingly US-origin. European carriers face a structural dependency on US-anchored model stacks, which intersects directly with data sovereignty obligations and policyholder confidentiality requirements.
For Lloyd's of London, European composites such as AXA, Allianz, and Zurich, and Asian carriers, the core question is how to leverage US-built AI infrastructure while maintaining compliance with local data residency rules and EU AI Act obligations. Architecture decisions made now will be difficult to reverse.
The insurance AI market growing at a 32.3% CAGR is not a technology experiment. It is a structural transformation with a decade of sustained investment ahead of it.
Regulatory and Sovereignty Deadlines: What Has Passed and What Is Coming in 2026
Insurance AI now operates inside an accelerating regulatory perimeter. The EU AI Act, EIOPA governance expectations, FCA Consumer Duty obligations, and US state-level NAIC bulletins all moved from consultation to enforcement between February and December 2025. Carriers that have not yet mapped their AI deployments against these requirements are already in arrears. The 2026 deadlines below are binding, not advisory.
Sources: EU AI Act Official Journal 2024 • EIOPA May 2025 • FCA October 2025 • NAIC December 2025 • Fenwick December 2025
Sources: EU AI Act Official Journal 2024 • NAIC December 2025 • FCA AI Strategy signals 2025 • EIOPA consultation timeline • Fenwick December 2025
The window to shape proportionate requirements through EIOPA consultation responses, FCA bilateral engagement, and EU AI Act implementation input is closing. The August 2026 high-risk system deadline is not a future concern — the conformity assessment, technical documentation, and human oversight infrastructure it requires takes 12 to 18 months to build. Carriers starting that work now are on the right timeline. Carriers that are not have a material compliance gap.
On data sovereignty: carriers building AI pipelines on non-sovereign infrastructure today face costly re-platforming as GDPR enforcement tightens and EU AI Act data governance requirements are applied. Architecture decisions made in 2026 will be difficult and expensive to reverse.
Inference Costs Collapsed 280x in Two Years: Insurance AI Economics Have Fundamentally Changed
Source: Stanford HAI AI Index 2025
Sources: IntuitionLabs, February 2026 • TLDL, February 2026 • Official provider documentation, verified February 2026
* Input prices are typically 5–12x lower
The 280x collapse in inference costs has changed the economics of insurance AI. The marginal cost of AI analysis on a motor claim is now measured in fractions of a cent. Three tiers map naturally onto claims complexity.
Premium tier ($5 to $168/M): Complex liability claims, large commercial underwriting, legal document analysis, reinsurance treaty review. Model quality materially affects decision quality.
Mid-tier ($1 to $15/M): Standard underwriting support, compliance drafting, policy summarisation, broker communications. The workhorse layer for most carrier operations.
Budget tier ($0.05 to $2/M): High-volume FNOL intake, simple claims triage, document classification, customer-facing assistants. Cost per interaction must stay below the economics of human handling.
The question is no longer "can we afford AI?" It is "are we routing tasks to the right model tier, and is that routing decision auditable and defensible to regulators?"
The Data Is Clear.
The Leaders Are Pulling Away.
T3 helps insurers move from signal to strategy with regulation-first AI governance, implementation, and training designed for the insurance operating environment.
This page is published by T3 Consultants Ltd for informational purposes only. It does not constitute legal, regulatory, compliance, investment, insurance, or financial advice. Readers should seek independent professional counsel before making decisions based on any information contained herein. Regulatory information: All references to the EU AI Act, EIOPA guidelines, FCA obligations, NAIC bulletins, and other regulatory frameworks reflect T3's understanding of publicly available materials as of the date of publication. Regulatory requirements, enforcement timelines, and supervisory expectations are subject to change. Nothing on this page should be relied upon as a definitive statement of legal obligations. Compliance determinations require qualified legal advice specific to your jurisdiction, business model, and AI deployments. Market data and statistics: Industry figures, adoption rates, and market size projections are sourced from third-party research organisations as cited. T3 has not independently verified all underlying data. Where primary data does not exist, the closest defensible proxy is used and noted explicitly. Statistics should be read in the context of their original methodology, sample sizes, and publication dates. Forward-looking projections: Market size forecasts, CAGR projections, cost savings estimates, and adoption trajectories are based on third-party research and involve assumptions that may not materialise. Actual outcomes may differ materially. T3 makes no representation as to the accuracy or completeness of forward-looking statements. AI model pricing: Pricing data for AI models was verified against provider documentation and third-party pricing sources in February 2026. Model pricing changes frequently and without notice. Output token prices are displayed; input pricing differs by model. Readers should verify current rates directly with providers before making procurement or architecture decisions. No endorsement: Reference to specific AI models, vendors, or third-party research does not constitute endorsement by T3 Consultants Ltd. Last reviewed: February 2026.
