New Data: Most Enterprise GPUs Sit Half-Idle, and Over Half of Companies Have Caught Their AI Agent Being Confidently Wrong

What happens when an enterprise AI agent answers a finance question with total certainty, cites a number, and moves on, and the number is wrong? According to new VentureBeat Research data, that scenario isn't hypothetical: 57% of enterprises say they've traced a confidently wrong AI agent answer back to bad or missing business context in just the past six months, and 31% say it happened more than once.

That figure comes from a VB Pulse survey of 101 enterprises, part of a broader VentureBeat Research effort that fielded five parallel surveys of the "agentic stack" among 573 technical leaders at companies with 100 or more employees in June 2026.

The other half of the story: idle GPUs

The same research project surfaced a second, seemingly unrelated problem: 86% of enterprises report that their GPUs run at half capacity or less. Companies are buying AI infrastructure faster than they can put it to work, a mismatch fueling an active debate on Wall Street over whether the AI buildout is running ahead of actual usage.

Underused hardware and confidently wrong software might look like separate issues, but they point at the same underlying condition: enterprises are deploying agentic AI ahead of the operational discipline needed to run it well. Consistent with that, 54% of surveyed companies reported at least one AI agent security incident or near-miss in the past 12 months that was caught before it caused damage, meaning the remaining 46% either had no incident or didn't catch one in time.

The proposed fix, and who actually has it

Researchers behind the survey point to what they call an "agentic context layer": a shared, governed model of what a company's business data actually means, built once and referenced consistently by every agent that touches it, rather than re-derived ad hoc each time. Only 25% of surveyed enterprises run one in production today. Another 34% are building one. The remaining 41% haven't started.

Oddly, the data shows companies already running or building a context layer report confident-wrong failures at a much higher rate (78%) than companies with no plans to build one (20%). That's not necessarily evidence the fix doesn't work; more likely, the companies furthest along in deploying agents at scale are also the ones with enough real usage to notice when something goes wrong.

What this means for small businesses

Most small and mid-sized companies aren't running fleets of GPUs, so the capacity numbers may feel abstract. But the underlying pattern applies at any scale: adopting an AI agent for customer support, reporting, or internal search without first agreeing on what your own numbers and definitions actually mean is how you end up with a tool that sounds authoritative and is wrong. Before scaling up agent use, it's worth asking a blunt question: if this agent gave a confidently wrong answer today, would anyone on your team actually notice?

Sources: VentureBeat, VentureBeat