Most property and casualty insurers are still struggling to move artificial intelligence projects beyond the testing stage, as weaknesses in data, governance and internal processes slow wider adoption.
Around 60% of insurers remain at the exploration or proof-of-concept stage, according to Capgemini's 2026 World Property and Casualty Insurance Report. The global study, which included interviews with 344 senior insurance executives across 18 markets, also found that 42% of insurers did not use key performance indicators to measure whether their AI investments were working.
The findings point to a gap between insurers' spending on AI technology and their readiness to use it across the business. Capgemini found that insurers directed 72% of their AI-related spending toward technology, compared with 28% for change management.
New research from insurance operations and technology services provider ReSource Pro found similar problems across the US P&C insurance market. Its 2026 AI Lessons Learned Report is based on interviews with more than 40 executives from across the insurance value chain. It asked organizations what they had learned from using AI, what obstacles they faced and what they planned to do next.
"What the industry is telling us loud and clear is that the organizations winning with AI are the ones treating it as a business transformation, not an IT initiative," said Mark Breading, Senior Partner at ReSource Pro. "The five lessons in this report are hard-won and remarkably consistent. No matter the segment, the same barriers and the same mindset shifts keep coming up."
The report identified poor data quality as the most common barrier to wider AI use. Many insurance organizations continue to store information across separate legacy systems, often in different formats and at varying levels of quality. This can reduce the reliability of AI-generated results and make it harder to use the same system across several teams or business functions.
However, fixing the data alone is not enough. ReSource Pro found that insurers also need clear workflows, documented procedures and defined responsibility for AI projects. Employees must continue to use human judgment when checking AI outputs, while management must decide who approves new uses, monitors performance and intervenes when a system produces inaccurate or unreliable results.
The report also found that some AI projects failed to move beyond proof of concept because responsibility for the technology was unclear or senior executives were not convinced that the expected benefits outweighed the security and implementation risks. Capgemini reached a similar conclusion. Its research found that insurers with more mature AI programs were more likely to have central governance, clear responsibilities for employees and cross-functional teams working toward shared performance targets.
This suggests that access to AI technology is no longer the main dividing line between insurers. The larger challenge is whether an organization can make the operational changes needed to use the technology at scale and show that it is producing measurable results. That readiness gap is starting to matter beyond individual insurers' own bottom lines, as US regulators move to make AI governance an explicit compliance requirement rather than a matter of internal best practice.
For carriers, these governance weaknesses are also becoming a regulatory issue. The National Association of Insurance Commissioners' model bulletin on insurers' use of AI calls for companies to establish written governance programs that reflect the risks posed by their systems. Regulators may ask insurers to provide information and documents during investigations or examinations, particularly when AI is used to support decisions that affect consumers.
The bulletin covers uses of AI in product development, marketing, underwriting, pricing, policy servicing, claims management and fraud detection. It also states that decisions involving consumers must comply with existing laws on unfair trade practices and unfair discrimination.
As of July 7, 2026, 25 US jurisdictions had adopted the NAIC model bulletin. California, Colorado, New York and Texas had introduced separate insurance-specific AI rules or guidance.
The NAIC is also testing an AI Systems Evaluation Tool in 12 states. The pilot began in March 2026 and is being used through formal examinations, surveys, data calls or a combination of these methods. The tool is designed to help regulators assess how insurers govern their AI systems and whether their controls are strong enough to manage the risks. Areas under review include board and senior management oversight, data accuracy, inventories of AI systems, bias monitoring, vendor management and the role of human reviewers. A revised version of the tool is expected to be considered at the NAIC's 2026 Fall National Meeting.
These controls may become more important as insurers begin testing agentic AI, which can carry out multi-step tasks with less direct human involvement. NAIC working-group materials have identified several potential risks, including AI systems acting beyond their authority, errors spreading across connected processes, problems with real-time data management and new cyber threats.
Against this backdrop, executives interviewed by ReSource Pro said their immediate priorities included improving data infrastructure, creating formal governance frameworks and giving employees more training. Organizations were also looking for narrower business problems where AI could deliver clear and measurable results before expanding it across underwriting, claims, distribution or customer servicing.
