When past stops predicting future: why insurance risk modeling needs a new data and AI foundation

Europe’s record-breaking summer heat reminded me the world is changing fast. A West Coast friend mentioned they now had four seasons: ‘Winter, Spring, Summer and Fire.’

For over a century, actuarial science has run on a simple assumption: yesterday's losses are the best guide to tomorrow's premiums. Trend tables, loss triangles, and smoothed historical curves have served insurers well in a world where floods, fires, and mortality rates changed slowly and predictably. Today’s world is less predictable, but the methods of risk management haven’t changed enough. Data from reinsurers, regulators, and consultancies now makes the case that weather volatility, AI-driven forecasting, and new governance requirements aren't industry trends… they are the critical mechanics of precise risk pricing.

The cracks in the trend line

According to Munich Re data cited in a 2026 industry review, global insured catastrophe losses hit $107 billion in 2025, with weather disasters responsible for the overwhelming majority of both total and insured losses. LexisNexis Risk Solutions' latest home trends report found claim severity at its highest point in seven years — up nearly 26% year-over-year and more than 93% since 2019. Industry surveys back this up: the 2026 RiskScan report found that insurers now expect natural catastrophes to become their single biggest risk category within five years, driven by perils — flooding, severe convective storms, excessive rainfall — that don't move in the gentle, linear patterns legacy trend models assume.

This is the core problem: a model built to extrapolate slow-moving history breaks down exactly when the underlying process becomes erratic. Insurers aren't just facing bigger losses — they're facing a fundamentally different loss-generating process, and the tools built for the old one aren't equipped for the new one.

Where AI is already closing the gap

The encouraging news is that better tools already exist, and they're producing measurable results. On the forecasting side, ECMWF's AI Forecasting System, which went fully operational in 2025, has shown performance matching or beating traditional physics-based models, including roughly 20% better accuracy in tropical cyclone track forecasts — data that feeds directly into the catastrophe models insurers rely on for pricing and reserving.

On the underwriting side, the Geneva Association, an insurance-industry research body, has found that AI implementation has driven a 43% improvement in risk assessment accuracy and a 31% reduction in processing time for complex policies. McKinsey's insurance research points in the same direction, finding that domain-focused AI deployment has driven measurable gains across underwriting, claims, and onboarding, including improved claims accuracy and faster onboarding. Individual carriers show what this looks like in practice: Zurich's sensor-driven AI system for commercial buildings has cut water damage claims by 45%.

The untapped value sitting in unstructured data

Much of the opportunity is about using more of the massive amount of disparate data, rather than scaling a process. Claims notes, adjuster narratives, medical records, inspection photos, and call transcripts have always carried signal; the industry simply lacked the tools to systematically make use of them. That's changing. Deloitte notes that Natural Language Processing (NLP) tools can now extract targeted insights from lengthy medical records, materially speeding up and sharpening risk assessments in life and annuity underwriting. Snowflake's industry research points to the same shift more broadly, observing that LLMs have made it possible for insurers to ingest and transform large volumes of unstructured data — claims records, legal documents, images, and customer interactions — data that was previously too costly or slow to use at scale. Academic research backs this up with hard outcomes: a 2026 study on fraud prevention found that integrating structured transactional data with unstructured sources like claim narratives and medical reports significantly improved predictive accuracy and anomaly detection compared to structured-data-only models. The same Snowflake research notes that third-party data — credit signals, weather feeds, purchasing behavior, even social data — is increasingly layered on top, giving insurers external context that historical loss triangles never captured. None of this is free of risk: unstructured and third-party data is messier, harder to audit, and more exposed to bias and privacy rules than a clean loss table ever was. But the evidence is consistent — insurers that build pipelines to responsibly ingest these additional data and files are extracting real, measurable accuracy gains that pure trend-based models simply cannot reach.

Why governance is the real bottleneck

None of this works, though, without data governance discipline — and the industry's own data confirms this as the binding constraint, not a compliance afterthought. A Grant Thornton survey found that 44% of insurance executives cite governance or compliance challenges as a barrier to AI implementation, even as 91% of insurers plan to keep investing in AI and 87% of CEOs now own AI decisions directly at the board level.

Regulators are moving in step. As of 2025, 24 U.S. states have adopted the NAIC's model bulletin requiring insurers to run a formal AI governance program covering data quality, bias testing, and board oversight across the AI lifecycle. New York's DFS guidance goes further, barring the use of AI in underwriting or pricing until insurers complete a full non-discrimination risk assessment. With NAIC data showing 88% of auto insurers and 70% of home insurers already using or piloting AI/ML models, these aren't theoretical rules — they apply to the models most insurers are already running today.

The opportunity

The evidence points to a clear shift: risk modeling is moving from reading the past to understanding the present in real time, using richer structured and much more unstructured data (parcel-level climate variables, sensor feeds, unstructured claims text) processed through AI systems that are only as trustworthy and audit-compliant as the governance wrapped around them. The insurers positioned to benefit won't be the ones with the most AI pilots — they'll be the ones who pair better data and models with the discipline to document and monitor them. In an unpredictable world, governed AI turns new data from risk into an actuarial edge.

Sources

1. Global RiskScan 2026: Cyber, Climate & AI Reshape Insurance Risks — https://beinsure.com/global-riskscan-cyber-climate-ai-insurance-risks/
2. Generative Artificial Intelligence and Its Implications for Weather and Climate Risk Management in Insurance (Gen Re) — https://www.genre.com/us/knowledge/publications/2025/september/gen-ai-and-its-implications-for-weather-and-climate-risk-management-en
3. AI is changing how P&C insurers model climate risk (Digital Insurance / LexisNexis Risk Solutions) — https://www.dig-in.com/news/ai-is-changing-how-p-c-insurers-model-climate-risk
4. Insurance industry navigates AI adoption and climate risks in 2026 (Munich Re, Grant Thornton, NAIC data) — https://eciks.org/10496-56740-insurance-ai-adoption-climate-risks-2026
5. AI Governance in Insurance: Risks, Compliance & 2026 Trends — https://cygeniq.ai/blog/ai-governance-in-insurance/
6. AI tools transforming the insurance profession (Geneva Association, Deloitte, Conning data) — https://www.idexconsulting.com/blog/2025/03/ai-tools-transforming-the-insurance-profession
7. The future of AI for the insurance industry (McKinsey) — https://www.mckinsey.com/industries/financial-services/our-insights/the-future-of-ai-in-the-insurance-industry
8. Generative AI in Insurance Industries: Transforming Underwriting and Claims (IRJET) — https://www.irjet.net/archives/V11/i7/IRJET-V11I767.pdf
9. Underwriter's edge: Harnessing Generative AI for optimal outcomes (Deloitte US) — https://www.deloitte.com/us/en/industries/financial-services/articles/generative-ai-insurance-underwriting.html

• 

  Phil Hailee

  Phil Hailee

  Senior Account Executive

  Collibra