A VB Pulse July 2026 survey found that enterprises are more often tracing confidently wrong AI agent answers to missing or inconsistent business context. Among 101 qualified companies with more than 100 employees, 68% said they had linked such an error to context issues in the past six months — up from 57% in a June survey. Reports of recurring failures rose as well, from 31% to 37%.
How context is provided matters
Every AI agent needs to understand what business terms and metrics actually mean, whether definitions are consistent, and whether documents are current. The survey highlights that companies provide that context in very different ways, with varying reliability.
Retrieval over documents remains the most common approach: it is the primary source for 31% of respondents. However, 13% primarily run agents by loading long-context directly into the model’s context window rather than using structured retrieval, and 5% give agents no structured context at all, relying only on the model’s general knowledge. In total, nearly one in five enterprises are either forcing context into the model or not feeding structured business context.
Even the dominant approach can produce confidently wrong outputs. Retrieval matches a question to text that looks similar in meaning; similar wording does not guarantee identical meaning. Srijith Rajamohan, AI research lead at Redis, illustrated this problem earlier: two sentences containing the same words can assert opposite facts, and an embedding-based retrieval followed by text search may not distinguish them.
Buying preferences shifted toward governance, but accuracy trails
When choosing a retrieval system, ease of data ingestion and access control now lead selection criteria, each cited by 24% of respondents. This is the first time in this survey series that a governance property has driven purchasing. Retrieval accuracy was chosen by 15%.
Once systems are in use, correctness remains the primary way enterprises judge them: 38% said response correctness is their primary success metric, roughly double the next-most-cited metric, security and access control (19%). In short, buying is moving toward governance, while success is still measured by whether answers are right.
Companies building governed layers report more failures because they can trace them
A governed context layer aims to fix these issues by providing a single, shared model of what business data means. Adoption is growing but incomplete:
- 32% run a governed layer in production.
- 31% are piloting or building one.
- 20% are evaluating one.
- 14% have no plans.
- 4% don’t know.
Comparing adoption to failure reports in the 91 companies able to say whether they had experienced the failure shows an inversion: those running or building a governed layer report recurring failures at 50%, while those without such a layer report recurring failures at 21%.
That does not mean the layer causes the failure. Rather, a governed layer makes the failure visible and traceable: linking a bad answer to a broken definition or stale table requires a shared, governed reference point. Without that reference, the same wrong answer may occur but be attributed to the model or never traced.
Kyle Nesbit, founder of the semantic layer startup Credible Data, noted that this is a long-standing issue: the lack of governed data analysis has been a pain point for 30 years, and AI has amplified the chaos and pain.
Company size and instrumentation affect detection
Larger companies report more recurring failures. Enterprises with more than 1,000 employees reported recurring failures at 55%, versus 30% for firms with 101–1,000 employees. Interestingly, the larger companies were less likely to have a governed layer in production (24% vs. 37%). More instrumentation and more people asking why a number was wrong turns up more failures, not fewer. A clean record may indicate lack of checks rather than a healthy context layer.
Practical implications for enterprises
The survey suggests several takeaways for organizations building on context layers:
- Retrieval alone will not close the context gap. RAG remains the default, but more documents or a bigger index won’t fix definitions that differ across systems.
- Budget commitments are outpacing production deployments. Sixty-three percent are building or running a governed context layer, but only 32% have one in production. Spend is going toward build and pilot activity, not universal resolution.
- A spotless failure record is a warning sign. Twenty-two percent report no context-related failure at all; this group is likely the least active in checking.
- Most organizations plan to avoid a single-provider lock-in. Seventy-nine percent intend to keep at least part of the context layer outside any one vendor’s stack (best-of-breed or mixed approach), while just 12% plan to consolidate onto a single provider’s native context stack.
As Constellation Research analyst Michael Ni put it earlier this year: “Whoever controls runtime context, controls the AI decision layer for enterprise data.” The VB Pulse data show that companies are actively wrestling with that control: building governed context layers increases visibility into errors and enables root-cause tracing, even as it temporarily raises reported failure rates while organizations surface and remediate problems.
