Where AI could hurt insurers more than it helps

As the insurance industry races to make the best use of AI, some areas of caution are emerging.

Testing must always come before deployment, or else the AI could do irreversible damage, according to warnings from leaders at Chubb, Westfield Insurance and WTW. Regulators could also place limits on certain applications, either by directly limiting AI for certain tasks, or by limiting access to the power grid to preserve natural resources. AI is also changing the nature of coverage, as AI exclusions become more common. 

Read our ongoing coverage below.

Test before launch: execs warn against AI overreliance

Distribution leaders at Chubb, Westfield Insurance and WTW cautioned against rushing AI deployment without proper testing, governance and change management. Chubb's Nicholas Davis warned that premature launches carry irreversible consequences, stressing the need for oversight structures to validate AI output. WTW's Nabeel Tanveer identified carrier negotiations and client-facing broker functions as areas AI cannot reliably handle, while simpler lines such as pet insurance are more suitable for automation. Westfield is adopting a deliberate "trust and verify" approach. Across carriers and brokerages, the operational priority is clear: audit AI-generated materials before client delivery and build internal review processes before scaling.
Read more: Overreliance on AI: insurers' biggest mistake, execs say

Audit AI use cases before resource limits force the issue

AI governance in insurance extends well beyond data privacy — carriers face emerging regulatory and resource constraints tied to energy consumption, water usage and infrastructure costs. States like California, with established environmental regulatory frameworks, could move quickly to cap AI capacity, require environmental footprint reporting or restrict access during constrained periods. Carriers embedding AI across every claim summary, fraud flag and payment recommendation are building operational dependencies that become difficult and costly to unwind. The practical priority: inventory AI deployments now, distinguish essential use cases from convenient ones, build fallback workflows and track where AI delivers measurable value — before regulators define those boundaries instead.
Read more: Why carriers must weigh AI's value before it gets unsustainable

Kansas candidate backs AI in claims, not for denials

Kansas insurance commissioner candidate Dinah Sykes (D) supports AI adoption for claims processing, fraud detection and risk assessment, but draws a firm line at using algorithms to deny or delay legitimate claims. Sykes, a state senator running unopposed in her primary and likely to face Republican Dan Hawkins, says fraud-detection savings should translate to lower premiums for policyholders — not higher insurer profits. If elected, she would develop regulatory guidelines aligned with NAIC AI standards, requiring bias testing and human accountability for every coverage decision. Her platform signals the regulatory direction carriers may face: demonstrate that AI tools are accurate, transparent and consumer-protective or face commissioner scrutiny.
Read more: How to use AI in insurance, and what to avoid: Kan. candidate

AI exclusions are quietly eroding corporate coverage

With 88% of organizations integrating AI into daily workflows, according to McKinsey's State of AI 2026 report, underwriters are responding by embedding sweeping AI exclusions into standard policy renewals across general liability, tech E&O, cyber and D&O lines — often without fanfare. The Allianz Risk Barometer now ranks AI risk second globally, marking the largest single-year climb of any risk category. Enterprise buyers face five distinct liability vectors — financial loss, IP infringement, unauthorized data disclosure, physical harm and property damage — that existing policy language was never designed to address. During renewal cycles, risk managers must audit policy towers for hidden exclusionary language and establish vendor risk assessment protocols to build an insurable AI risk profile.
Read more: Why AI changes the rules of corporate liability

How embeddings are reshaping insurance risk modeling

Insurance carriers are moving beyond standardized data attributes toward neural network-generated embeddings — mathematical representations that capture the full complexity of raw data, including timing, sequences and semantic meaning. Unlike attributes, which summarize specific data points, embeddings surface patterns that data scientists may never have thought to code manually, such as a sequence of missed loan payments followed by new credit applications. Carriers already using embedding-based models report improved predictions for loss propensity, severity and customer churn. To stay competitive, underwriting and data science teams must evaluate whether current attribute-based models are leaving meaningful risk signals — and premium accuracy — on the table.
Read more: How insurance data is evolving from raw reports to embeddings

AI turns build-vs-buy into a more nuanced question for insurers

AI is dismantling the traditional build-or-buy binary in insurance procurement, replacing multi-year sourcing decisions with continuous portfolio management. The calculus has shifted: agentic development tools now lower build costs enough that even non-differentiating systems may warrant in-house development, while core functions like claims, underwriting and onboarding increasingly justify insourcing to protect brand trust. Four factors drive sound decisions: total cost of ownership, client trust, brand impact and competitive differentiation. Hidden risks demand attention — shadow AI introduces data exfiltration and prompt-injection exposure, single foundation-model dependence mirrors vendor lock-in, and frontier model costs can quickly outpace budgets. Self-hosted, air-gapped models are gaining traction as geopolitical and vendor-failure hedges.
Read more: How the rise of AI is forcing a build-or-buy rethink

UnitedHealth targets 80% real-time prior authorizations by 2027

UnitedHealth Group's $3 billion AI investment is producing measurable operational results, with executives targeting 80% of prior authorizations processed in real time by end of 2027 — a benchmark other health insurers will likely face pressure to match. AI now touches nearly every provider and consumer interaction, automating complex claims, enabling real-time data sharing with health systems and flagging at-risk members for proactive outreach. 
Read more: Where UnitedHealth says its AI investments are paying off

California telematics bill creates AI ambiguity for auto insurers

AB 311, the Consumer Driving Data Protection Act, would let California auto insurers use scoring models on telematics data to set rates — but whether that permits AI remains disputed. The bill's author says AI is explicitly excluded; Consumer Watchdog warns predictive algorithmic modeling could qualify, and that telematics data collected today could train future AI models. The California Department of Insurance would retain oversight of predictive algorithms if the bill passes. For insurers navigating compliance planning, the ambiguity warrants direct engagement with the CDI and legal counsel before building AI-dependent telematics infrastructure around the legislation.
Read more: Why California's telematics bill leaves insurers guessing on AI

AI adoption linked to higher retention rates, Liberty Mutual finds

Independent agencies reporting retention gains are more likely to embrace AI: 52% view AI as an opportunity, and 51% say AI tools have saved them notable time, according to Liberty Mutual's 2026 Independent Agency Growth Study of roughly 1,200 U.S. agency principals and staff. The average agency retention rate sits at 84%, with most reporting flat or declining figures year-over-year and only 18% posting gains of 5% or more. Automating annual reviews and deploying AI-powered policy-checking tools — which flag coverage gaps and cross-sell opportunities — are among the highest-impact applications for agencies looking to scale retention efforts without sacrificing client relationships.
Read more: AI optimism aligns with agency retention: Liberty Mutual study

How neural network embeddings sharpen insurance risk models

Embeddings — compact numerical representations learned by neural networks — offer insurers a path beyond brittle, manually engineered feature sets that require constant rebuilding as data volumes grow. Unlike traditional models dependent on predefined attributes, embeddings learn risk relationships directly from geospatial data, behavioral patterns, claims history and text, then reuse those representations across underwriting, pricing, segmentation and fraud detection. Properties or customers that differ on the surface but behave similarly under loss conditions cluster together in embedding space, enabling more precise risk assessment. For carriers managing expanding data sources and shifting risk conditions tied to climate or economic cycles, the practical advantage is reduced model maintenance and greater consistency across analytical teams.
Read more: Why neural networks and embeddings improve insurance models

Process change, not new tech, closes the insurance performance gap

Claims inflation is outpacing earned premium growth, compressing combined ratios across P&C, while regulators and rating agencies increasingly evaluate operational resilience alongside financials. The core problem isn't outdated technology — it's unstructured institutional knowledge and rigid processes that prevent rapid response to pricing shifts, catastrophe volatility and regulatory requirements. Replacing legacy systems without first organizing that knowledge and streamlining workflows will not produce the speed or adaptability the market now demands. The path forward is bimodal: prioritize quick wins in high-value pain points, structure data for AI readiness and treat change as continuous iteration rather than discrete transformation projects.
Read more: Why managing process change is more important than new technology

How Travelers built a cheaper, faster in-house LLM

Travelers' proprietary large language model, TravelersLLM — trained on millions of internal insurance documents and evaluated against tens of thousands of domain-specific questions — is outperforming general-purpose frontier models on cost and speed for insurance-specific tasks. The insurer credits AI broadly with contributing to strong Q2 earnings. Travelers measures ROI not on model cost alone but on decision quality, consistency and enterprise scalability across underwriting, claims and operations. Tens of thousands of employees use its generative AI platform monthly. The company's approach — embedding governance into AI design rather than treating it separately — offers a replicable framework for carriers operating in regulated environments.
Read more: How Travelers measures ROI for its in-house LLM

This roundup was created with AI assistance. A Digital Insurance editor reviewed each item before publication.