In August, 67% of respondents to an enterprise survey said their organizations were operating, testing or developing a governed semantic layer. Just 13% identified it as the main source of business context for their agents. That disconnect exposes a deployment problem: building agreed business definitions doesn’t ensure an AI system uses them when answering a question.
Adoption Is Rising, but Agents Read Elsewhere
Across three survey waves, semantic-layer activity increased while its role as the main context source declined. Enterprises appear to be investing in governed meaning while sending their agents through other routes to reach information.
| Survey wave | Operating, testing or developing a semantic layer | Layer identified as main agent context source |
|---|---|---|
| June | 58% | 21% |
| July | 63% | 19% |
| August | 67% | 13% |
A governed semantic layer gives business terms an agreed interpretation. Revenue needs a definition; customer records need a basis for deciding which one is current. Without that meaning, an agent can retrieve valid data and still produce an answer that misrepresents the business.
August respondents most often identified document retrieval as their agents’ main context route, at 32%. Queries against live systems accounted for 21%, compared with 11% in July. That increase was large enough for the survey to distinguish, although the difference between live queries and document retrieval was not.
Another 19% relied primarily on long-context loading, which places substantial amounts of documents or records into a model’s context window. Each route supplies material for an answer. None automatically establishes what the organization means by the material.
An agent executing SQL, invoking an API or accessing an MCP server can receive a column value without receiving its business definition. Loading more records into a prompt leaves a similar gap unless agreed definitions accompany them. Both methods can work alongside semantic governance, but the survey did not establish whether respondents had connected the two.
More Reported Failures Do Not Prove Worse Performance
Context errors remain common. In August, 64% of respondents reported identifying absent or conflicting business context behind an agent’s confident but incorrect answer during the preceding six months. The corresponding figures were 68% in July and 57% in June.
Organizations already operating, testing or developing a semantic layer reported these failures more frequently than those merely considering one or having no plans: 78% versus 37% in August. A separate July sample of 101 respondents showed the same direction, at 89% versus 35%.
Removing organizations that lacked agents working on internal data, or did not investigate root causes, left the difference intact. August’s figures became 79% versus 41%; July’s became 90% versus 46%. The pattern deserves scrutiny, but it does not establish that semantic governance causes more errors.
The August survey report offered two possible explanations. Governed definitions could make incorrect answers easier to detect because teams have a reference against which to check them. Alternatively, earlier context failures could have prompted organizations to build a layer in the first place.
Respondents were not asked what caused the difference. The evidence therefore cannot separate better detection from adoption driven by previous failures, or determine whether the layer changes the actual error rate. Treating reported incidents as a simple performance ranking would overread the data.
Vendors Are Placing Context at Different Points
The product response is becoming more specific than “give the model more data.” OpenAI, Keewano and Snowflake place context at different points in the agent workflow. Those choices affect what persists, what accompanies a query and how much exploration a model must perform.
OpenAI’s Data agent, introduced within ChatGPT Work in September, connects to warehouses including Snowflake, Databricks and BigQuery. Alongside warehouse access, it gathers information from Slack, BI dashboards and file storage during the same operation.
Arpan Shah, general manager of the enterprise technology vertical at OpenAI, explained that the agent brings together existing context rather than creating a persistent context layer spanning customer tools. Its approach is assembly at use time. Enterprises evaluating it should distinguish that capability from maintaining a durable set of shared definitions.
KeewanoDB, also introduced in September, gives agents MCP access to raw event sequences. That sounds like straightforward live querying, but Mark Kardashov, co-founder and CEO of Keewano, described an ingestion process that uses large language models to construct a semantic layer and enrich arriving events with context.
The distinction changes how a query should be assessed. Keewano says meaning is attached before an agent reads the events, so direct access need not mean access to context-free records. The relevant question becomes what enrichment occurred upstream.
Snowflake addresses the issue through model routing and preparation. Its Cortex AI Gateway initially assigns work to a smaller model, moving to a more capable one when necessary. That approach depends on receiving a task with sufficient context already supplied.
Baris Gultekin, Snowflake’s vice president of AI, explained in August that poorly prepared context forces models to explore: generate SQL, test it, search data and retry unsuccessful steps. He described that work as costly and generally requiring a more capable model. Preparing the context beforehand removes that exploratory burden.
Snowflake says Horizon Context and Cortex Sense support this preparation, while existing permissions determine which data and models an agent may access. Sanjeev Mohan, principal at SanjMo, argued in August that the differentiation lies in keeping routing within governed data boundaries, with permissions, tagging and cost attribution attached.
Deployment Decisions Need to Follow the Answer Path
Even among August respondents already operating, testing or developing a semantic layer, only 20% identified it as their agents’ main context source. Governance adoption alone is therefore a weak guide to what actually shapes an answer.
The vendor approaches suggest concrete questions for deployment reviews:
- Context assembly: Does the agent combine existing information during a task, or retain a shared layer between tasks?
- Ingestion enrichment: What meaning is attached before a record becomes available for querying?
- Task preparation: Which definitions reach the model before it begins exploring, and which access controls remain in force?
These questions separate retrieval capability from business interpretation without assuming that one architecture is always superior. OpenAI assembles context during use; Keewano describes enrichment during ingestion; Snowflake couples preparation with governed access and routing. Each places responsibility for meaning somewhere different.
As live querying gains ground, enterprises need to examine that placement rather than count semantic-layer projects. The practical priority now is to trace an agent’s answer back through its data route and verify that the organization’s definitions reach the point where the model uses them.
