Agentic AI Has a Foundation Problem

There's considerable excitement surrounding agentic AI and its potential to transform enterprise operations. Organizations are exploring autonomous agents capable of retrieving information, coordinating activities across applications, making recommendations, and executing increasingly complex workflows. While these capabilities are promising, I keep coming back to a fundamental concern: we're introducing greater autonomy into enterprise environments that have struggled with data quality, interoperability, and architectural complexity for decades.

These problems aren't new, and they certainly weren't created by AI. Enterprise systems have historically been developed to satisfy specific program, organizational, or operational requirements. Each system often establishes its own data structures, definitions, interfaces, and assumptions about how information should be interpreted. Over time, organizations introduce integration layers, middleware, and APIs to exchange information between these environments. While these approaches enable connectivity, they don't necessarily resolve the underlying differences in meaning, authority, or intended use of the data.

This distinction becomes particularly important with agentic AI. A conventional integration may transfer information between two systems according to predefined mappings and business rules. An AI agent, however, may retrieve information from multiple sources, interpret relationships between them, and use those interpretations to determine subsequent actions. If the underlying systems use inconsistent definitions or conflicting assumptions, the agent may produce a response that appears reasonable but is based on information that was never semantically aligned.

Consider something as simple as an enterprise reporting process. Two systems may report the same business event using different definitions of when that event is considered complete. Both systems may be operating correctly within their respective domains, yet their information is not necessarily interchangeable. An AI agent tasked with consolidating those reports must have some basis for determining which definitions apply, whether the information is comparable, and which source has authority for the decision being made. Providing the agent with access to both systems does not establish that understanding.

This is where I believe many discussions about enterprise AI readiness are missing an important architectural consideration. We've spent years emphasizing data accessibility and connectivity, but considerably less attention has been given to preserving the context necessary for information to be interpreted consistently across systems. Metadata, semantic relationships, provenance, and governing authority are not simply supporting documentation. They are essential to understanding what data represents, where it originated, and under what conditions it can be relied upon.

The challenge extends beyond semantics. Many enterprise environments carry substantial technical debt in the form of tightly coupled applications, duplicated information, undocumented dependencies, and integration mechanisms that have evolved through years of incremental modifications. Agentic AI may be able to navigate these environments more efficiently, but that doesn't mean the underlying architectural problems have been resolved. In fact, introducing autonomous workflows across previously isolated systems may expose inconsistencies that were manageable when humans remained responsible for interpreting the information and reconciling differences.

And Then There's the Issue of Governance…

Organizations frequently have extensive governance frameworks that identify data stewards, establish quality standards, and define responsibilities. Yet those frameworks are not always connected to the operational mechanisms through which information is accessed, interpreted, and used. As agents begin moving beyond information retrieval into decision support and execution, the distinction between having a documented policy and being able to enforce that policy becomes increasingly significant.

For example, an agent may have technical access to information without having the authority to use it for a particular purpose. A source may be authoritative within one operational context but inappropriate for another. Two sources may disagree, or the information necessary to establish an authoritative interpretation may be missing altogether. These conditions require explicit governance and architectural mechanisms. They should not be left to an AI model to infer simply because it can retrieve and process the available information.

None of this suggests that organizations should abandon agentic AI or wait until every legacy system has been modernized. There are legitimate opportunities to improve productivity and operational effectiveness within existing environments. However, we should distinguish between demonstrating that an agent can complete a workflow and establishing that the underlying information environment can support dependable, repeatable, and appropriately governed autonomous operations.

I also think we need to reconsider how we define AI readiness. Model performance, infrastructure, security, and workforce adoption are important, but they represent only part of the problem. Organizations must also understand the architectural foundations upon which their AI capabilities will depend. That includes the meaning and provenance of information, the authority associated with its use, the relationships between systems, and the mechanisms available to identify and resolve conflicts.

For years, we've attempted to address enterprise interoperability by connecting systems that were never designed to share a common understanding of their data. Agentic AI doesn't eliminate that problem. It introduces a new participant that can operate across those boundaries at a speed and scale that makes the consequences of unresolved inconsistencies potentially more significant.

Perhaps the greatest contribution of agentic AI will be that it finally forces us to address some of the architectural and data governance problems we've been carrying forward for decades. But that will only happen if we recognize that the foundation matters as much as the intelligence we place on top of it.