Enterprise agents ship to production despite failing internal trust tests

The survey's most pointed finding isn't the failure rate itself but the direction of travel: organizations that already distrust their evals are simultaneously moving toward removing the last human checkpoint from the deployment loop, which means the failure surface is expanding while the safety net is shrinking.

This is largely disconnected from recent activity in our archive, as we have no prior coverage to anchor it to. It belongs to a broader conversation that has been building across the agent infrastructure space, where vendors including major cloud providers and specialized eval startups have been pitching observability and testing tooling as the solution to production reliability. The survey data complicates that pitch directly: the problem isn't that enterprises lack eval coverage, it's that the evals they have don't predict real behavior. That distinction matters commercially because it shifts the value proposition away from 'more tests' toward 'better simulation of production conditions,' which is a harder and less mature product category.

Watch whether any of the major agent platform vendors (Salesforce, ServiceNow, Microsoft) publish updated evaluation methodology documentation or announce dedicated red-teaming partnerships within the next two quarters. If they don't respond to this kind of survey data with concrete tooling changes, it signals the industry is betting on failure rates staying below the threshold that triggers regulatory attention.