We ask a lot of AI agents: read a ticket, find the right information, propose an action. When the result disappoints, the reflex is to blame the model. On the ground, the model is often the least guilty party. What is missing is a stable representation of the business, objects, relations, units, rights, that the agent can query without inventing.
A contract, not another warehouse
A semantic layer is not another data warehouse. It is a contract. It states what “customer”, “order”, “site”, or “invoice” means in your context, how those notions relate, and under which rules they can be read or written. Without that contract, the agent scavenges exports, wikis, and heterogeneous APIs. It can produce a fluent sentence and a wrong decision.
The problem becomes visible as soon as several systems talk about the same object with different definitions. An active customer for CRM is not necessarily a billable customer for finance. A site for operations is not always a site for energy billing. The agent has no way to decide if nobody formalized those distinctions. It then picks the most accessible source, not the most accurate one. That is the same challenge as a data API layer: make the landscape addressable without rebuilding it.
Readability before autonomy
Teams that hold in production invest first in that operable vocabulary. They expose business queries rather than raw tables. They log what the agent believed it understood. They refuse broad access “just in case”. Autonomy comes after readability, never before. Until meaning is shared, widening the agent’s scope mostly accelerates errors. Agent governance (logs, scope, escalation) completes that frame.
That is the spirit behind bricks like VERB: make data interpretable before handing it to an agent. If your agents hallucinate on business meaning more than on grammar, the issue is probably not the LLM. It is the absence of a layer that says what the words mean in your house. For more on the product, see also verbsolution.com.
