The Biggest Risk for AI Agents May Be the Company’s Own Data

An agent can answer a question in seconds and still be wrong for reasons that have little to do with its language model. It may draw from an outdated customer record, combine figures from systems that disagree or apply a definition of revenue that differs from the one used by the finance team. To the employee receiving the answer, however, the distinction is usually invisible. The response sounds authoritative, and the failure appears to be an AI hallucination.

That confusion is becoming a serious enterprise problem. In a VentureBeat Intelligence survey of 130 qualified respondents, 64% said they had traced at least one confidently wrong AI agent answer during the previous six months to missing or inconsistent business context.

The survey covered organizations with at least 100 employees. Its findings point to a difficult reality for companies rushing to put agents into customer support, analytics, operations and internal decision-making: improving the model may not fix an answer that is built on contradictory corporate information.

When the data disagrees

The failures described in the survey were not necessarily caused directly by the model. Instead, agents encountered gaps in the information needed to interpret a request correctly.

One system may contain a recent customer status while another still shows an old one. A sales database may define an active account differently from a finance system. Two departments may use the word “revenue” while counting different transactions. Even a well-performing retrieval system can return an answer that is internally coherent but wrong for the business if the underlying sources conflict.

These are familiar organizational problems, but AI makes them harder to see. A human analyst who notices conflicting figures may ask a colleague which source to trust. An agent can select one result, combine several results or present a qualified guess in polished language. The result feels more certain than the underlying evidence warrants.

Nearly half of the respondents who reported a context failure said it had happened more than once, according to the survey. That repetition suggests the issue is not limited to an isolated bad document or an occasional stale record. In many companies, it reflects a persistent weakness in how information is organized and governed.

For employees, the consequences can range from wasted time to bad decisions. An agent used for reporting might misstate a business trend. One supporting sales staff could recommend the wrong account action. A system answering questions about policy might give a technically fluent response based on an obsolete rule. The more an organization treats the agent as an authority, the more costly those errors become.

Retrieval is not the same as understanding

Many companies have responded to unreliable AI answers by adding retrieval. The basic idea is straightforward: instead of asking a model to rely only on what it learned during training, retrieve relevant documents or records and place them in the context of the question.

Retrieval remains the most common primary context source among respondents, at 32%. Direct queries to live systems followed at 21%. Those approaches can improve access to information, but access alone does not guarantee that the information is consistent, current or properly defined.

The survey found that only 13% of respondents named a governed semantic layer as their primary context source. A semantic layer is intended to provide a shared interpretation of enterprise data. Rather than simply pointing an agent toward files or databases, it can help establish which fields matter, how business terms are defined and which sources should take priority.

That distinction resembles the difference between giving someone a library and giving them a catalog maintained by an authority. The library may contain the answer, but the catalog helps explain which edition is current, how subjects are organized and whether two apparently similar books should be treated as the same work.

For AI agents, this shared structure could become a control point between raw corporate systems and the people or software asking questions of them. It may also give companies a way to measure whether an answer was based on an approved definition rather than merely on a plausible passage retrieved from somewhere inside the organization.

A correlation that needs caution

The survey found that 67% of respondents either operated, piloted or were building a governed semantic layer for agents and business intelligence tools. Yet organizations in those categories were more likely to report a context failure. The rate was 78%, compared with 37% among companies that were only evaluating a semantic layer or had no current plans.

At first glance, that result could suggest that semantic layers are failing to solve the problem. The survey does not establish that conclusion. Companies with stronger data governance may simply be better at detecting and documenting wrong answers. Organizations may also have started building semantic layers precisely because they had already experienced failures.

This is an important distinction. A company that cannot identify a wrong answer may appear more reliable than one with sophisticated monitoring. Better measurement can make a problem look larger before it makes the system better.

The finding also exposes a challenge for executives evaluating AI programs. A low incident count does not necessarily mean an agent is trustworthy. It may mean that no one has traced the answer back through its sources, definitions and freshness. Reliability requires not only producing an answer, but also understanding why the system produced it and whether the business agrees with the assumptions behind it.

The economics of choosing a context layer

The survey offers another warning about how companies are selecting their AI infrastructure. Among respondents running production retrieval systems, ease of data ingestion was the leading selection factor at 36%. Only 11% named retrieval accuracy.

Ease of ingestion is understandable. Enterprises often have years of accumulated systems, documents and databases, and connecting them is expensive. A tool that can absorb information quickly may appear to offer the fastest route to deployment.

But convenient ingestion can create a larger problem if it imports unresolved conflicts. A system that gathers every available source without ranking authority, checking freshness or defining key terms may increase the amount of information available to an agent while reducing confidence in the answer.

The strategic market is also unsettled. Thirty-seven percent of respondents expected to keep best-of-breed standalone tools. Twenty-eight percent anticipated a mixture of provider-native and independent systems, while 22% expected to consolidate onto one model provider’s stack.

That contest is not only about which model is most capable. It is also about who controls the layer that tells the model what the company means. Providers may offer increasingly integrated tools, while independent vendors may compete by supporting multiple models and existing enterprise systems.

For businesses, the choice will shape more than architecture. It will determine where definitions are maintained, who can change them and how easily an organization can audit an answer after something goes wrong.

The survey’s central lesson is therefore less dramatic than the usual debate over model intelligence, but more consequential for daily work. An agent can be fluent, fast and technically advanced, yet still fail because the company has not agreed on its own facts. Before asking which model should run the business, organizations may need to answer a more basic question: which version of the business does the model have permission to believe?