A survey reported in September 2026  by S&P Global Ratings covers 121 insurers and reinsurers representing roughly 38% of the assets of rated companies. It suggests that most respondents have moved beyond concepts into some form of operational integration, while only about one third describe AI as fully integrated. Reported benefits are appearing first in customer experience, underwriting, risk management and claims.

Operational adoption is ahead of financial proof

S&P also indicates that the financial impact is not yet clear and that the evidence has not led to rating actions. Those caveats matter. A reduction in handling time or an improvement in document processing can be real without yet producing a measurable change in profit, capital or credit quality. The survey is a snapshot of respondents, not proof that every insurer will achieve the same result.

Why operational gains arrive first

AI is often introduced into a specific task: summarising a file, extracting data, routing a request, assisting an underwriter or drafting a response. The immediate measures are therefore operational—minutes saved, cases processed, errors flagged or response times improved. Financial outcomes sit further downstream and can be affected by volumes, pricing, staffing decisions, loss experience and customer behaviour.

A second reason is that organisations retain human review while systems mature. This can improve quality and manage risk, but it means savings are not equivalent to eliminating the original workload. Teams may also reinvest time in better analysis or service. That can be strategically valuable even if it does not appear as a direct cost reduction.

The measurement risk runs in both directions. Firms may overstate value by converting every saved minute into theoretical cash, even when headcount and costs do not change. They may also understate value by ignoring faster decisions, improved consistency or earlier detection of problems. A credible business case should distinguish productivity capacity, avoided loss, revenue effects and realised financial benefit.

Build evidence in stages

This editor recommends a three-level scorecard.

The first level measures use and process: adoption, cycle time, completion rates and exceptions.

The second measures quality and risk: accuracy, human corrections, complaints, leakage and compliance incidents.

The third measures economics: realised cost, conversion, retention, loss ratio or capital effect, depending on the use case. The causal link should be stated rather than assumed.

Governance should follow the same progression. A low-impact drafting assistant does not require the same validation as a system that affects underwriting or claims. Data provenance, model limitations, access controls, human escalation and monitoring should be proportionate to the consequence of error. Where AI contributes to a decision affecting a customer, the insurer should preserve an intelligible record of the input, output and professional judgement.

The S&P findings support a practical conclusion: operational evidence is useful, but it is not the end of the evaluation. Insurers should continue successful pilots while demanding stronger proof as systems move closer to material financial or customer decisions. The question is not whether a use case can generate an impressive demonstration. It is whether the organisation can operate it reliably, govern it proportionately and show where the value ultimately appears.

Avoid the pilot-to-portfolio gap

A use case can perform well in a controlled pilot and still struggle when connected to live systems, broader user groups and changing data. Before scaling, insurers should define ownership for model updates, workflow changes, incident response and retirement. They should also budget for evaluation and supervision as operating costs rather than treating them as temporary implementation work.

Portfolio management is needed once several AI tools affect the same customer or process. Separate assistants may create inconsistent messages, duplicate controls or hidden dependencies on one model provider. A central inventory, common risk classification and consolidated view of incidents can prevent local productivity gains from creating enterprise-level fragility.

The next review should test the recommendation against fresh operating evidence, identify any unintended consequences and record who owns the resulting action. That discipline keeps the proposal proportionate and allows governance to evolve as market practice, technology and regulation change.