Enterprises are building managed context layers to help AI agents understand business data, definitions, and processes more accurately. However, companies with these systems in place are currently more than twice as likely to report recurring AI failures—not necessarily because the technology creates more errors, but because it makes existing problems easier to detect.
Over the past six months, 68% of enterprises said they had traced an AI agent’s confidently incorrect answer to missing or inconsistent business context. Of those, 37% said the issue occurred more than once, while 32% said it happened only once. The findings come from VB Pulse’s July 2026 survey of 101 eligible companies with at least 100 employees.
The overall figure is up from 57% in VB Pulse’s June survey. Recurring failures also increased from 31% in June to 37% in July. This is the second time VB Pulse has asked companies the same question, and the results show that failure rates are rising even as more organizations move managed context systems into production. The share reporting failures increased from 25% in June to 32% in July.
How AI agents receive context determines whether they get the answer wrong
Every AI agent needs a reliable way to understand what a business means, how metrics are defined, and whether internal documentation is current. That business context is essential for reducing AI hallucinations and confidently incorrect answers.
The challenge is that enterprises provide context to AI agents in very different ways, and those approaches are not equally dependable.
Document search remains the most common method, serving as the primary information source for 31% of companies. Other organizations rely on less structured approaches. About 13% primarily use long context windows, feeding documents directly into the model instead of retrieving relevant information. Another 5% provide no structured context at all and rely largely on the model’s general knowledge. In total, nearly one in five companies either provides business context through long context loads or does not provide it in a structured way.
Even advanced retrieval systems can produce confidently incorrect answers. Search typically identifies content that is semantically similar to a question, but similar wording does not always mean similar intent or meaning. Srijith Rajamohan, Redis’ AI research leader, described this gap in an interview with VentureBeat earlier this year.
“If you had a sentence like ‘Rome is closer than Paris’ and a sentence like ‘Paris is closer than Rome,’ and you did an embedded search and then a text search, you wouldn’t be able to tell the difference,” Rajamohan said. “The same word is present in both sentences.”
Governance is influencing AI search purchases, but accuracy still lags
The way companies select enterprise search and retrieval systems does not always address the root cause of incorrect AI answers. Access control and permissions, along with ease of data ingestion, are now the leading selection criteria, each cited by 24% of respondents.
This is the first time in the research series that governance-related capabilities have played such a significant role in purchasing decisions. Search accuracy, however, was cited by only 15% of companies. The capabilities most directly connected to preventing confidently incorrect answers are not yet the primary factors driving most purchases.
After an AI system is deployed, companies still need to determine whether it is working accurately. Correct answers are the primary success metric for 38% of organizations—twice the share that identified security and access control as the most important measure, at 19%.
In other words, companies are becoming more governance-focused when purchasing AI infrastructure, but they continue to measure success primarily by whether an AI agent delivers the right answer.
Companies with managed context layers are reporting more failures
A managed context layer is designed to provide a shared, controlled model of what a company’s data means. AI agents, business intelligence tools, and analytics systems can reference the same definitions instead of interpreting business terms independently.
Adoption remains fragmented. Approximately 32% of enterprises have one managed context system in production, while 31% are building or piloting one. Another 20% rated their implementation at level one, 14% have no plan, and 4% do not know their current status.
When adoption data is compared with reported failures, the pattern appears reversed. Among the 91 companies that disclosed whether they had experienced a failure, 50% of organizations that run or are building a managed context layer reported recurring failures. By comparison, 21% of companies without such a layer reported recurring failures.
This does not mean managed context layers cause more AI errors. Instead, they make errors visible. A shared reference point helps organizations identify broken definitions, outdated tables, inconsistent metrics, and incorrect answers. Without that visibility, the same errors may continue to occur without being properly attributed or investigated.
The underlying problem predates generative AI by decades. Kyle Nesbit, founder of semantic layer startup Credible Data, explained the issue to VentureBeat last month. “This is the same problem people have had for 30 years: lack of controlled data analysis,” Nesbit said. “With the advent of AI, the same problems arise, but with orders of magnitude more disruption and pain.”
Company size reinforces the same conclusion. Enterprises with more than 1,000 employees reported a 55% recurring failure rate, compared with 30% among companies with 101 to 1,000 employees. This was despite larger organizations being less likely to have a managed context layer in production—24% compared with 37% of smaller enterprises.
The more systems a company operates and the more employees rely on its data, the more opportunities there are for inconsistent definitions and incorrect answers to surface. A clean failure record is not necessarily evidence of a healthy AI context layer. It may simply indicate that no one is actively checking.
What managed context layers mean for enterprise AI
These findings have several implications for organizations building AI agents and data infrastructure:
Search alone cannot solve the business context gap. Retrieval-augmented generation remains the default way to provide context, while nearly one in five enterprises relies on long context windows or has no structured context layer. Adding more documents or expanding an index will not resolve a definition that means different things in different systems.
AI budgets are moving faster than production infrastructure. Approximately 63% of enterprises have either built or are operating a managed context layer, but only 32% have one in production. The biggest gap is not necessarily awareness or spending; it is the transition from experimentation to dependable infrastructure.
A clean AI failure record can be a warning sign. The 22% of companies reporting no context-related failures are not necessarily the best-managed organizations. They may simply be the least likely to monitor and investigate AI answers. Larger enterprises report recurring failures at nearly twice the rate of mid-market companies, even though they are less likely to have managed context systems in production.
Most companies do not plan to rely on a single provider. About 79% of organizations intend to keep at least part of their AI context layer outside one vendor’s technology stack. They favor best-of-breed tools or explicit combinations of systems, while only 12% plan to adopt a single provider’s native context stack.
The preference for maintaining control over enterprise AI context is consistent with VentureBeat’s reporting throughout the year. Constellation Research analyst Michael Ni summarized the strategic importance of this infrastructure when DataHub’s context-layer initiative emerged.
“Those who control the runtime context will control the AI decision-making layer of their enterprise data,” Ni said.
Source: venturebeat.com