The third wave of VentureBeat's Pulse tracker on enterprise context layers landed this week, and the interesting number isn't adoption — it's the widening gap between what teams are building and what their agents actually read from.
Across three waves since June, the share of surveyed enterprises (100+ employees) running, piloting, or building a governed semantic layer went 58% → 63% → 67%. The share naming that layer as their agents' *primary* source of business context went the other way: 21% → 19% → 13%. You can build the thing and still not be in the path.
Where context actually comes from, per the August wave: retrieval over documents at 32%, direct queries to live systems at 21%, long-context loading at 19%, semantic layer at 13%. The direct-query number is the one to sit with — it was 11% in July. Doubling in a month is a real signal, and it has an obvious cause. Wiring an agent to an MCP server or a SQL endpoint is now a afternoon's work. Curating definitions is not.
The definition problem is not a retrieval problem
The architectural cost is specific. A direct query returns rows; it does not tell the agent what `revenue_net` excludes or which customer table is canonical. Long-context loading has the same hole at a different layer — you hand the model raw material and hope it infers the business meaning. Together that's 40% of primary context paths carrying no governed definition by default, and it's the path that's growing.
VentureBeat's framing of context failure holds up: 64% of August respondents traced a confidently wrong agent answer to missing or inconsistent business context in the prior six months, with the question explicitly excluding model errors. July was 68%, June 57%.
The counterintuitive result is detection. Among enterprises running, piloting, or building a semantic layer, 78% reported a context failure, against 37% among those only evaluating or with no plans. July showed the same split, 89% versus 35%. VentureBeat says it didn't ask why the groups differ, so read this as an instrumentation artifact, not causation — teams with a canonical definition can tell when an answer contradicts it. Teams without one just ship the answer.
Why the cheap path wins
A NeMo Retriever paper circulating this week puts numbers on the alternative: NVIDIA reports agentic retrieval lifting nDCG@10 by 8.7 points while taking roughly 160x longer. That ratio explains the drift. When the high-quality retrieval path costs two orders of magnitude in latency, a direct query against a live table looks rational right up until it answers with an ungoverned column.
If you own a RAG or agent platform, the actionable piece is logging which source answered each question, not just whether the answer was right. The 13% number only exists because someone asked where context came from. Most teams can't answer that about their own stack.