Via ey.com
A new enterprise survey finds that building infrastructure to catch AI errors often makes the failures more visible, not less frequent.
Here is a counterintuitive finding for the enterprise AI playbook: companies that invest in the technology specifically designed to stop AI agents from hallucinating are reporting more hallucinations, not fewer.
A VB Pulse survey published in July 2026 found that 68% of enterprises traced at least one confident-but-wrong AI agent answer to missing or inconsistent business context over the prior six months. That figure was 57% in VentureBeat’s June 2026 survey. Recurring failures, meaning the same type of context breakdown happening more than once, climbed from 31% to 37% in the same window.
The survey polled 101 qualified enterprises with more than 100 employees, a tight sample but one that skews toward organizations actively deploying AI agents rather than just experimenting with them.
The detection paradox
Enterprises that have built, or are actively building, a governed context layer reported recurring failures at a 50% rate. Companies without such a layer saw recurring failures at just 21%.
As of July 2026, only 32% of survey respondents have a governed context layer fully deployed in production. Another 31% are piloting or actively building one. Add in organizations still evaluating, and engagement with this infrastructure category exceeds 80% of the sample.
RAG leads, but nothing dominates
Retrieval-augmented generation leads as the primary context approach at 31% of respondents. Governed semantic layers came in at 19%, with mixed or use-case-specific approaches at 17%.
What enterprises actually care about
When respondents ranked what matters most in evaluating context and AI agent solutions, access control and permissions tied with data ingestion capability, each cited by 24% of respondents as a top buying criterion. Response correctness ranked as the leading success metric at 38%.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our