Summary

  • Sapiens’ survey of 475 insurance professionals records high interest in agentic AI, but only 16% say it is already in production.
  • The report’s 24% figure measures the average share of six business processes with agentic AI live; it is not the share of insurers using it.
  • The index measures foundations such as cloud, master data, governance and workforce readiness. It does not count autonomous claim decisions, and human oversight remains extensive.

Analysis

A forecast meets its operating baseline

A new insurance-market index puts a striking horizon beside a much smaller present-day footprint. About two-thirds of respondents expect insurance AI to be fully autonomous within two to three years. The report’s final takeaway gives the figure as 66%; Sapiens’ launch release rounds it to 67%. Either way, it is a forecast from people working in insurance, not a deployment plan or a measured capability.

The same report says 87% regard agentic AI as important to their business, but only 16% have it running in production. Half describe themselves as fully or mostly ready to adopt it. These numbers describe three different states: strategic priority, readiness and live deployment. Treating any one of them as a proxy for the others would make the market look further along than the evidence shows. (Sapiens Insurance AI Maturity Index 2026; Sapiens’ 7 October launch release)

The denominator changes the story

A second statistic can be misread if it is detached from its label. The report says agentic AI is live in an average 24% of six process areas: distribution, underwriting, servicing, claims, fraud and reinsurance. That is process coverage, not the proportion of insurers that have adopted the technology. The same comparison puts ordinary AI at 42% and rule-based automation at 40% of process areas.

The distinction matters commercially. A small number of insurers could have agentic tools in several workflows while most have none; another pattern could involve many firms testing one narrow process. The aggregate 24% cannot tell a buyer which pattern exists. The 16% production figure answers a different question: how many surveyed organisations report any agentic AI running live. A sound market comparison needs both company-level adoption and workflow-level depth, with the sample and period attached.

The study also defines agentic AI as a system that acts across multiple steps with less direct human input. It defines multi-agent systems separately, as several agents coordinating through a process. A conversational assistant, a model that drafts a claim summary and a system that gathers records, assesses a claim and issues payment are not interchangeable uses of the word “AI”.

Readiness is not autonomy

Sapiens groups respondents into four maturity bands: 25% Autonomous Leaders, 28% Operational, 23% Enabled and 24% Reactive. The names can sound like a map of autonomous decisions, but the score is built from cloud migration, master-data strategy, AI governance and HR readiness. The top band has those foundations in place; it is not a count of insurers handing claims decisions to machines.

That construction still offers useful market information. Ninety per cent of respondents report at least some cloud infrastructure, yet 36% use a hybrid model and 18% use SaaS. Cloud presence is therefore not the same as a modern, uniform operating core. Among insurers not yet using agentic AI, respondents cite privacy and security, regulation and compliance, and integration with existing systems as leading barriers. More than a quarter say their core systems need a significant overhaul.

For software suppliers, the commercial opening is not simply to sell an agent. It is to help a carrier connect reliable data, permissions, legacy platforms and staff to a bounded workflow. The survey suggests that this work can be more difficult than the product demonstration, but it does not price the opportunity or compare vendors.

Human oversight is part of the baseline

Respondents who use AI or agentic AI report confidence in the decisions and in their organisations’ checks. Yet the report says a person reviews an average 53% of agentic AI output, and people override 15% of AI or agentic-AI decisions. That is not evidence that the systems have failed: a review may be a planned control, and an override can be appropriate. It is evidence that the current operating model still includes human judgement at scale.

Nor does the survey measure the time spent reviewing, the cost of an incorrect decision, the share of checks that change an outcome or the distribution of errors across products and customers. Confidence is not an error rate. Production use is not proof of safe autonomy. Those missing denominators matter when an insurer compares an AI tool’s promised speed with the cost of integration, oversight and recovery.

A useful signal with a bounded reach

Research in Finance conducted a 10-minute online survey for Sapiens from July to September 2026. The 475 respondents were evenly drawn from the UK and Ireland, the Nordics, Benelux, South Africa and North America, and split between transformation leaders and business users. They worked across Life and Property & Casualty insurance, at firms ranging from fewer than 250 to more than 20,000 employees. Sapiens says qualitative interviews with executives will form a second stage.

This is a structured view across five regions, not a census of every insurance market. Sapiens commissioned the work and sells insurance software, so its commercial interest belongs beside the findings. The index is best read as a useful baseline for the questions it defines, not as independent proof that the whole industry will reach autonomy on the stated timetable.

Sources