In September 2026, KPMG published research showing that insurance leaders’ confidence in AI transformation exceeds their reported readiness in important areas. The release appeared on 29 September, drawing on research conducted from 20 to 29 May and supplementary insurance research.

For insurers and benefits providers, the useful response is to test what an AI workflow delivers and what supports it. A management assessment of relative leadership cannot by itself establish financial value, dependable service or adequate controls.

Treat survey confidence as a starting point

KPMG reports that 44 percent of respondents place their organisations in the top quartile for AI transformation, while 11 percent report strong data foundations and governance. Separately, 11 percent describe their view of AI return on investment as very clear. These measures address different questions and should not be combined.

The findings are self reported. The release describes participating leaders across 20 countries and six regions, and cites 53 insurance respondents in supplementary research. Those 53 should not automatically be treated as the sample behind every percentage. The results provide a diagnostic prompt, rather than an audited ranking of insurers.

Measure an entire workflow

An insurer should define the task that is expected to improve and record a baseline before implementation. For a benefits enquiry, that might include the time required to resolve the question, the accuracy of the answer and the need for subsequent correction. Faster drafting is useful only if the completed response remains suitable.

Measurement should include the work moved to other teams. A system can appear productive while creating more review, remediation or customer follow up elsewhere. Counting generated outputs or users does not capture those costs. Compare a complete case with the previous process and explain material differences in case complexity.

Financial evaluation should include integration, data preparation, oversight and recurring operating costs. A pilot with unusually clean records or intensive expert support may not reproduce its apparent economics at scale. The business case should show which assumptions have been tested and which remain estimates.

Make benefits data fit for the task

Global benefits workflows often require a clear relationship between country, employer, policy period and benefit wording. A response drawn from an outdated document can be fluent and still inappropriate. The assistant needs a way to select the applicable source and show that source to the reviewer.

Data readiness should therefore be assessed through actual questions. Test a changed eligibility rule, an expired policy, conflicting provider information and an enquiry spanning two countries. The expected behaviour may be to escalate or request clarification. A system that can recognise insufficient evidence may provide more dependable service than one that always completes an answer.

Keep editorial knowledge and client specific records distinguishable. A professional article may explain a market practice without establishing the contractual entitlement of a particular employee. Permissions and source labels should preserve that distinction throughout retrieval and response generation.

Give the business an accountable owner

A business sponsor should own the intended outcome, with technology and control functions supporting delivery. Document which decisions the system may assist, which actions require approval and who can suspend the workflow when it fails.

The board should ask for examples of errors and their resolution, not only successful demonstrations. A useful report links an incident to the source, process or permission that caused it and shows whether the corrective action was verified. Management should distinguish an isolated correction from a change that prevents recurrence.

Training should reflect the task and the employee’s responsibility. Reviewers need to know how to check the evidence, identify uncertainty and handle exceptions. Attendance at a general AI session does not demonstrate that someone can supervise a benefits enquiry or claims workflow.

Expand on demonstrated results

This editor recommends a small number of defined workflows with measured outcomes, clear ownership and documented limits. Expand deployment when the evidence supports the next use, rather than treating organisational confidence as an approval criterion.

KPMG’s research helps frame the questions. The answer for each insurer must come from its own data, operating experience and controls, with anticipated benefits kept separate from results already observed.