More builders, not just central projects
In a customer story published July 31, OpenAI says Dutch cooperative insurer Univé treated ChatGPT Enterprise as an organizational capability rather than another IT rollout. Leadership sessions, privacy and security review, inherited enterprise permissions and human accountability preceded employee experimentation.
Vendor-published metrics
OpenAI and Univé report 97% license activation, 85% weekly active use, 40 prompts per active user per week, and about 1,500 employee-built custom GPTs. Pet-insurance claim preparation reportedly fell from hours to minutes, while trained claims handlers retain every final decision.
For underwriting, agents prepare missing-document and risk signals from approved sources before a person decides. Permissions are said to follow the source systems so AI cannot see more than the user is authorized to access.
What can transfer—and what cannot
Deployments should measure sustained use, approval boundaries, correction time and business outcomes instead of seat counts alone. Evidence for recommendations and a named final decision-maker are especially important in claims and underwriting.
These are self-reported customer-story figures, not an independent audit or causal experiment. No comparative cost, error-rate or customer-outcome data was published, so the same result cannot be assumed elsewhere.