The Next Enterprise AI Advantage May Depend on How Businesses Define Their Data

An AI agent can access a company's databases, retrieve documents, and generate a convincing answer in seconds. Yet it can still get a fundamental business question wrong. What counts as revenue? Which customer record is authoritative? Does a particular metric refer to gross sales, net sales, or recognized revenue?

These are not necessarily failures of artificial intelligence in the conventional sense. They can be failures of business context.

New survey findings from VentureBeat Intelligence reveal a striking disconnect. By August 2026, 67% of surveyed enterprises said they were running, piloting, or building a governed semantic layer to give business data consistent meaning. Yet only 13% identified that layer as their AI agents' primary source of business context.

Meanwhile, 64% of August respondents said that, during the preceding six months, they had traced a confidently wrong AI agent answer to missing or inconsistent business context.

The numbers point to a problem that is becoming central to enterprise AI adoption. Companies are investing in better ways to organize and govern their data, but many agents still obtain the information they need through other routes. The next stage of enterprise AI may therefore depend less on making models more intelligent and more on ensuring that they understand what a business actually means.

The News Behind the Numbers

The findings emerged from three waves of VentureBeat's VB Pulse survey conducted in June, July, and August 2026.

Across those waves, the proportion of respondents whose organizations were running, piloting, or building a governed semantic layer increased steadily.

• June 2026. 58%.

• July 2026. 63%.

• August 2026. 67%.

However, the proportion naming a governed semantic layer as their agents' primary source of business context moved in the opposite direction.

• June 2026. 21%.

• July 2026. 19%.

• August 2026. 13%.

At the same time, context-related failures remained common. The share of respondents reporting a confidently wrong agent answer caused by missing or inconsistent business context was 57% in June, 68% in July, and 64% in August.

The figures do not establish that semantic layers are ineffective. They reveal a gap between organizations investing in governance and the way their AI agents actually obtain business information.

That distinction matters. A company can build a well-governed data architecture without ensuring that every AI application uses it.

What Is New About the Enterprise AI Context Problem?

Traditional business software generally operates within predefined rules, data structures, and workflows. A reporting system may use an established formula for calculating revenue, while a customer database follows agreed procedures for identifying the current account record.

AI agents introduce a different operating model. They can retrieve information from documents, query databases, call APIs, and coordinate multiple steps to answer questions or complete tasks.

That flexibility creates opportunities for automation. It also introduces ambiguity.

Suppose a finance executive asks an agent to calculate revenue for a particular quarter. The underlying warehouse contains several fields that might appear relevant. One represents gross sales, another reflects net sales, and another contains recognized revenue.

A model may retrieve the correct records and perform its calculations accurately while still selecting the wrong definition.

The resulting answer can sound authoritative because the model has produced a coherent explanation based on the information it received. But the answer may not correspond to the company's official financial definition.

A governed semantic layer attempts to address this problem by establishing consistent meanings for business concepts, metrics, and relationships.

The challenge is ensuring that those definitions become part of the information available to agents when they make decisions.

Why Business Context Matters More Than Ever

Enterprise AI is moving beyond isolated chat interfaces toward agents that perform multistep tasks using company information.

An agent might summarize sales performance, investigate a customer complaint, prepare a financial report, or retrieve information from several internal systems. Each task depends on more than access to data. It depends on interpreting that data correctly.

Without consistent business definitions, different agents can produce incompatible answers from the same underlying information.

One system might count a customer as active based on a recent transaction. Another might use an account-status field. A third might infer activity from a document that has not been updated.

Each answer could appear reasonable in isolation.

Across a business, however, these inconsistencies can undermine reporting, customer operations, and executive decision-making.

The problem becomes more consequential when agents are allowed to take actions rather than simply generate recommendations. An incorrect interpretation of a customer record or business metric can propagate through a workflow, affecting subsequent decisions.

This is why enterprise AI governance cannot be reduced to permissions, model selection, or security controls alone. Businesses also need a reliable way to communicate the meaning of their information.

How Enterprise AI Agents Obtain Business Context

The VB Pulse findings identify several approaches that enterprises use as their primary source of business context.

These percentages describe the primary source respondents identified, not every component of their AI architecture. The approaches are not necessarily mutually exclusive at the technical level.

A direct query can be supported by a semantic layer. A document retrieval system can incorporate governed definitions. An agent can combine several sources during the same workflow.

The important question is not simply where an agent retrieves information. It is whether the information arrives with enough business meaning to support a correct answer.

Where the Data Reveals a Surprising Contradiction

The survey found that enterprises running, piloting, or building a semantic layer reported confidently wrong AI agent answers more frequently than organizations still evaluating one or reporting no plans to adopt one.

In August, the figures were 78% versus 37%. A separate July wave involving 101 respondents found a similar pattern, at 89% versus 35%.

The gap remained when the analysis excluded enterprises that did not run agents on their own data or did not track the root cause of failures. In August, the comparison was 79% versus 41%; in July, it was 90% versus 46%.

At first glance, this seems counterintuitive. Why would organizations investing in more rigorous data governance report more context failures?

The survey did not establish the cause, and two explanations deserve consideration.

First, enterprises with semantic layers may be better equipped to detect mistakes. Once an organization has an agreed definition of revenue or customer status, it has a clearer standard against which to evaluate an agent's answer. Businesses without such standards may fail to recognize that an answer is wrong.

Second, the relationship may run in the opposite direction. Organizations that experienced serious context failures may have been more likely to start building semantic layers in response.

In that case, the semantic layer would be associated with a higher reported failure rate because it was adopted by companies already confronting the problem.

The findings do not prove that semantic layers increase errors, reduce errors, or have no effect on underlying failure rates. They demonstrate why measuring reported failures requires an understanding of how organizations detect and investigate them.

For executives, that is an important lesson in itself. Better governance can make problems more visible before it makes them disappear.

How Vendors Are Approaching the Context Gap

Several vendors are addressing the relationship between live data access and business context, but their approaches differ.

OpenAI focuses on assembling context across tools

OpenAI's Data agent, launched inside ChatGPT Work in September, connects directly to data warehouses including Snowflake, Databricks, and BigQuery. It can also retrieve context from Slack, business intelligence dashboards, and file storage during the same process.

According to Arpan Shah, OpenAI's general manager of the enterprise technology vertical, the Data agent does not create a persistent context layer across a customer's tools. Instead, it combines the available contexts for the task at hand.

This approach emphasizes assembling relevant information when a user asks a question. It can be useful when information is distributed across multiple systems and the goal is to bring it together quickly.

However, assembling information and establishing permanent, governed business definitions are different functions. A dynamically assembled answer still depends on the quality and consistency of the underlying information.

Keewano attaches context during ingestion

Keewano takes a different approach with KeewanoDB, which launched in September.

The system allows agents to read raw event sequences directly through MCP, placing it within the category of direct-query approaches measured by the survey. However, Keewano co-founder and CEO Mark Kardashov described a semantic layer that uses large language models to attach additional context to events as they enter the system.

That distinction matters because a direct query does not necessarily mean an agent is working with raw, unexplained information.

If context has already been attached during ingestion, the agent may receive a more meaningful representation of an event even when it queries the underlying data directly.

The broader architectural principle is that context can be prepared before a request arrives, rather than reconstructed entirely during inference.

Snowflake connects context preparation with model routing

Snowflake approaches the problem through data governance, context preparation, and model selection.

Its Cortex AI Gateway automatically routes tasks to a smaller model first and escalates when necessary. But routing is only useful if the model receives sufficient information to complete the task.

Snowflake's vice president of AI, Baris Gultekin, has explained that models without good context may need to explore data themselves, write and test SQL, search through information, and retry failed approaches. This can increase costs and create a need for more capable models.

Snowflake says its Horizon Context and Cortex Sense tools help prepare context before inference, while existing access controls govern the data and models available to agents.

The commercial proposition is therefore broader than selecting the cheapest model. It is to combine appropriate context, controlled access, and model routing so that tasks can be completed with less unnecessary exploration.

As industry analyst Sanjeev Mohan observed in the original report, this approach connects routing with governance and cost attribution within an established data boundary.

The three approaches illustrate an emerging distinction in enterprise AI architecture. Context can be assembled dynamically, attached to data during ingestion, or prepared as part of a governed data platform.

No single approach is automatically right for every workload.

Platform or Product. Where Does the Competitive Advantage Lie?

The semantic-layer debate highlights a familiar tension in enterprise technology. Businesses can invest in a foundational platform, but the value of that investment ultimately depends on the applications using it.

A semantic layer can establish shared definitions, but it does not automatically determine what every agent reads. The survey illustrates this gap: although two-thirds of August respondents said their organizations were running, piloting, or building a semantic layer, only 20% of that group identified it as their agents' primary source of business context.

That figure differs from the 13% reported across all respondents because it refers specifically to organizations adopting a semantic layer.

For platform providers, the opportunity is to make governed context accessible throughout the AI stack. For application developers, the challenge is to use that context correctly in the workflows that matter.

The most valuable product will not necessarily be the one with the largest collection of enterprise integrations. It may be the one that consistently interprets company data correctly, respects access restrictions, and produces results that users can verify.

In other words, connecting more systems is not the same as understanding them.

Competitive Risks and Strategic Trade-Offs

The emerging competition involves more than model accuracy.

Vendors that assemble context dynamically can offer flexibility, particularly when information is spread across many tools. But they must handle inconsistent definitions and changing source data.

Systems that attach context during ingestion can reduce the need to reconstruct meaning for every request. However, the quality of their results depends on the accuracy, completeness, and maintenance of that prepared context.

Governed data platforms can provide consistency and access controls, but businesses must still integrate those capabilities into the agent's execution path. A governance system that exists but is rarely consulted may have limited influence on actual outputs.

These approaches also raise questions about portability. If business definitions, metadata, and agent workflows become tightly coupled to one platform, changing vendors may become more complicated.

For enterprises, the objective should be to preserve clear business definitions and make them accessible across the systems that need them, rather than assuming that any single vendor will solve every context problem.

How the Global Enterprise AI Market Could Change

The shift toward context-aware AI has implications for organizations across Europe, the United States, and other major markets.

As AI agents become more deeply integrated into finance, operations, customer service, and analytics, businesses will need stronger ways to ensure that different applications interpret the same information consistently.

This creates an opportunity for data-platform providers, enterprise software vendors, and companies building tools for AI governance.

It also changes how organizations should evaluate AI investments. Model benchmarks and reasoning capabilities remain important, but they are only part of the equation. The quality of data access, business definitions, and context preparation can determine whether an otherwise capable model succeeds in a real enterprise environment.

For multinational businesses, the problem can become more complex when different departments, regions, or business units use different definitions for apparently identical metrics. A shared semantic layer may help establish common definitions while preserving legitimate local differences, provided those distinctions are explicitly represented.

The survey does not measure how these factors differ across countries or industries. Nevertheless, the underlying issue is relevant wherever AI agents must operate across fragmented enterprise systems.

What Businesses Should Do Next

Enterprises do not need to wait for a perfect architecture before improving AI reliability. They can begin by making the path from business question to final answer more transparent.

First, define critical business terms. Establish authoritative definitions for concepts such as revenue, active customers, retention, and completed orders. Identify who owns each definition and how exceptions should be handled.

Second, map the agent's context sources. Determine whether each agent uses document retrieval, direct database queries, long-context loading, a semantic layer, or a combination of these approaches.

Third, test answers against business definitions. Evaluate not only whether the model retrieved accurate source data, but also whether it applied the correct business rules and selected the appropriate records.

Fourth, distinguish error detection from error prevention. Track how frequently agents produce incorrect answers, how often those mistakes are detected, and whether specific interventions reduce the underlying error rate. Better monitoring may initially increase reported incidents by revealing problems that previously went unnoticed.

Fifth, align context architecture with the workload. Dynamic context assembly may suit exploratory tasks spanning several applications. Prepared semantic context may be valuable for recurring business metrics and structured workflows. Other tasks may need a combination of both.

Finally, connect governance to the execution path. A semantic layer provides limited protection if agents routinely bypass its definitions. Organizations should verify that the relevant business context reaches the model when it matters and that access controls remain effective throughout the workflow.

These steps turn semantic governance from an abstract data initiative into an operational discipline tied to AI reliability.

Enterprise AI's Next Breakthrough May Be About Meaning

The enterprise AI market has spent considerable effort improving models, expanding infrastructure, and connecting agents to more data. The next challenge is more fundamental: ensuring that AI systems understand the meaning of the information they receive.

The VB Pulse findings show that companies increasingly recognize the need for governed semantic layers, yet relatively few identify those layers as their agents' primary source of context. They also show that organizations building such layers report more context failures, although the survey cannot establish whether those companies experience more errors or simply detect them more effectively.

The distinction is crucial. Technology leaders should not judge a semantic layer solely by whether reported errors fall immediately after implementation. They should examine whether business definitions become more consistent, whether agents use those definitions, and whether incorrect answers become easier to identify and prevent.

OpenAI, Keewano, and Snowflake illustrate different ways of addressing the problem, from assembling context dynamically to enriching data during ingestion and preparing governed context before model inference.

The wider lesson is that access to information is only the beginning of enterprise intelligence. Reliable AI requires a connection between data, business meaning, governance, and the task being performed.

As companies move from experimental chatbots to agents that influence real operations, that connection could become one of the most important sources of competitive advantage.

The Next Enterprise AI Advantage May Depend on How Businesses Define Their Data

enterprise AI context, semantic data layers, AI agent governance, business data meaning, enterprise AI agents, data context management, AI retrieval systems, governed business data, AI context failures, Snowflake Cortex AI, OpenAI Data agent, KeewanoDB, enterprise data governance, AI agent reliability