Agents Without Semantics are Flying Blind

It has been a while since my last blog post. The reason is straightforward: we have been focused on the hard work of bringing AI into production and making it generally available within Extreme Platform ONE. With that milestone now achieved, it’s a good time to pause, reflect, and continue sharing the lessons we’ve learned along the way.

Why agent governance matters

In earlier blogs, I have talked about the different ways users can interact with AI in our system. The journey always starts simple... talking to AI through conversational interactions. But the real shift happens when you move to the next stage: letting AI act on your behalf. That is where trust, adoption, and exponential value start to compound.

The moment you cross that threshold—when AI goes from answering questions to taking actions—governance becomes critical. Governance is not only about controlling access to the underlying data and helping to ensurecompliance with regulations. Just as important, it’s about giving users the ability to define how their agents operate: how broadly they can be deployed across the organization (the deployment scope), how deep their actions can go (the action scope), and how autonomous they should be—whether they need a human in the loop, or the freedom to act independently. These dimensions together define the safe and scalable use of AI agent.

Why semantics are the foundation

Among the many strategies that keep agents on task, one stands out as foundational: embedding rich metadata and precise semantics in a domain-aligned ontology. Put simply, AI is only as intelligent as the data it truly understands. This may sound obvious, but when organizations move from experimentation to deploying agents at scale, the absence of a semantic backbone becomes painfully clear.

We often celebrate breakthroughs in model design, multi-agent systems, or clever prompt engineering. These innovations matter. But none of them can deliver consistent value if the agents lack a machine-readable understanding of what enterprise data means. Without that shared understanding, agents operate in a fog of ambiguity. They misinterpret context, make brittle decisions,and produce outputs that vary from inconsistent to outright misleading.

Real-world examples of semantic failure

To illustrate, I asked AI itself to provide examples of how poor semantics can derail agents:

• Healthcare AI: Misdiagnosing conditions because terminology is inconsistent across hospital systems. For instance, “MI” could mean “myocardial infarction” or “mitral insufficiency.” The consequence? Potentially life-threatening errors.
• Retail recommendation engines: Recommending irrelevant products because product metadata is flat and unstructured. An agent might confuse “running shoes” with “dress shoes,” leading to a poor customer experience and lost revenue.
• Financial agents: Making incorrect forecasts because time-series data lacks the right context or because metrics are misaligned. “Revenue” and “net revenue” are not the same—but an agent without semantics might treat them as interchangeable.

These failures are not the result of bad models —they are the result of bad semantics. A model can be advanced, but without clear meaning in the data, it will misinterpret context and produce flawed results. In multi-agent systems, those mistakes compound quickly as one agent’s errors cascade to others. The reverse is also true: once the semantic layer is well-structured and aligned, every new data set strengthens the system. By tying data into a consistent ontology, agents can reason more effectively, collaborate more smoothly, and deliver outcomes that grow exponentially as the system scales.

In summary, if youwant AI to truly scale across your enterprise —powering everything from customer service to engineering, finance, and operations —the key isn’t just more models or more agents. The real foundation is the data infrastructure that gives those agents meaning: rich metadata, well-structured ontologies, and a semantic layer that allows intelligence to compound. Invest there, and you create the conditions for AI to deliver consistent, trusted, and exponential value across every corner of your business.