Even when companies introduce governance-based context layers to prevent AI agents from producing wrong answers, the share reporting such errors rises instead, a survey found.

A recent VentureBeat report said 68 percent of companies over the past six months found that answers AI agents delivered confidently were wrong, and that company-related information was missing or incorrect.

That is a sharp increase from 57 percent in a June survey. The share saying wrong answers were repeated at least twice also rose to 37 percent from 31 percent in June. The survey was conducted in July by VB Pulse, run by VentureBeat, of 101 companies with more than 100 employees.

Over the same period, the share of companies actually operating governance layers increased to 32 percent from 25 percent. Error reporting rose alongside adoption of the layer.

The most common way companies provided context to agents was RAG, at 31 percent.

Long-context approaches that put entire documents into a model’s context window accounted for 13 percent. Another 5 percent relied only on a model’s general knowledge without separate structuring. Srijes Rajamohan (스리지스 라자모한), an AI research lead at Redis, said, "The two sentences 'Rome is closer than Paris' and 'Paris is closer than Rome' cannot be distinguished with embedding search alone."

Among companies already operating or building governance layers, the rate of repeated wrong answers was 50 percent, higher than 21 percent at companies without such layers.

Among large companies with more than 1,000 employees, the repeated-error rate was 55 percent, higher than 30 percent at companies with 101 to 1,000 employees. Michael Ni (마이클 니), an analyst at Constellation Research, said, "Whoever controls the runtime context (that AI refers to) will come to control the enterprise data AI decision-making layer."