Companies are building governance-backed semantic layers, but AI agents are obtaining business context elsewhere, VentureBeat reported on Monday.

In VentureBeat's VB Pulse survey, the share of respondents saying they operate, run pilot applications of, or are building a semantic layer rose to 58 percent in June, 63 percent in July and 67 percent in August. A semantic layer sets definitions for company terms such as what counts as revenue and which customer record is the latest. But only 21 percent, 19 percent and 13 percent, respectively, cited it as a main source of context for agents.

In August, 64 percent of respondents pointed to a lack of context or mismatches as the reason agents had confidently given wrong answers in the past six months.

Document search was the most-cited source of agent context at 32 percent. Directly querying real systems using SQL, an API or an MCP server was 21 percent, up sharply from 11 percent in July. Feeding large volumes of documents into a model was 19 percent. Direct querying and bulk input do not, by default, include agreed term definitions.

Companies that adopted semantic layers experienced wrong answers more often, according to VentureBeat.

In the August survey, 78 percent of companies that had adopted a semantic layer said an agent had confidently given a wrong answer. Among companies reviewing adoption or with no plans, the figure was 37 percent.

The report offered two interpretations. One was that semantic layers serve as a standard for correct answers and lead companies to find more wrong answers. The other was that companies that first experienced errors built layers as a solution. VentureBeat said the survey alone could not determine which was the case.