Enterprise AI agents are giving confidently wrong answers because of missing or inconsistent company data, according to a survey of 130 people. The finding puts a practical question ahead of another round of agent demos: are organisations giving these systems dependable context to work with?
Watch Desk analysis
What happened
VentureBeat Intelligence surveyed people at organisations with at least 100 employees in August 2026. It found that 64% said an AI agent had produced at least one confident but wrong answer in the previous six months, traced to problems such as stale data or a missing definition of a business metric. The survey asked respondents to exclude errors caused by the model itself; 32% reported more than one context failure.
The survey also found that 67% of respondents said their organisation runs, pilots or is building a governed semantic layer: a shared set of definitions for business data, such as what counts as revenue. Yet among organisations running one in production, only 23% said it was agents’ primary source of business context.
Key findings
- Context failures were common among respondents
Nearly two-thirds reported at least one confidently wrong agent answer linked to company data in the past six months. - Semantic-layer status tracked with more reported failures
The figure was 78% among respondents whose organisations ran, piloted or built a semantic layer, versus 37% among those only evaluating one or with no plans. - Agents often get context elsewhere
Respondents most often named document retrieval as their agents’ primary context source (32%); 21% named direct queries to live systems and 13% a semantic layer. - Finding data may outrank finding the right data
Among respondents running production retrieval systems, 36% named ease of data ingestion as their main selection factor, compared with 11% who named retrieval accuracy.
Why it matters
A model can produce a polished answer from the wrong metric, stale record or conflicting source. A shared definition gives teams something to check against, but merely having a semantic layer does not mean an agent actually consults it. The survey suggests that data access and data governance are part of agent reliability, not backstage tidying to leave until after the launch party.
The higher reported failure rate among organisations building or using semantic layers is not evidence that those layers cause errors. The survey says it did not ask why the groups differed. Teams with better ways to check answers may simply be finding more failures; organisations may also have invested in a semantic layer after encountering them.
Our read
The useful takeaway is not that a semantic layer is a magic fix. It is that teams should check which data and definitions their agents actually use, and whether they can trace an answer back to them. A confident tone is not a data-quality strategy.
What to watch
- Whether organisations make governed definitions a primary source for agents, rather than just for dashboards.
- Whether future surveys measure independently checked accuracy, not only reported failures.
- How retrieval systems balance easy data ingestion with finding the right, current information.
Discussion spark: Should organisations make a governed semantic layer the default source for AI agents, or is that too rigid when different tasks need different data?
Sources and evidence
- 64% of enterprises find AI errors in data – VentureBeat (5 October 2026, 17:40 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.