Nick Burling, Chief Product Officer at Nasuni.getty​In conversations with advisory boards and customer user groups, AI governance consistently ranks among the top concerns. What remains far less clear is how organizations should govern AI agents as they gain broader access to enterprise data.Organizations can readily articulate how an AI model should behave. Ask how an agent arrived at a particular output, however, and the conversation often becomes much less certain. That disconnect is becoming one of the biggest blind spots in enterprise AI.Governance frameworks are becoming increasingly sophisticated. Still, most are designed to evaluate model behavior rather than control how AI interacts with enterprise data.An organization should be able to determine whether an agent accessed information that it was authorized to use and trace the output back to the source material that informed it. When those questions can't be answered, governance becomes difficult to enforce and even harder to prove. Without that audit trail, a governance framework is little more than a PowerPoint deck. How AI Agents Change The Governance Equation​Early AI governance focused on model behavior. That was the right place to start when the primary concern was what AI might generate. Organizations invested in responsible AI frameworks designed to reduce the risks associated with AI-generated outputs. AI agents introduce a different governance requirement because they operate on enterprise data rather than just prompts. Every output is influenced by the data an agent can reach and the permissions that determine what it is allowed to see. As agents become more deeply embedded in day-to-day operations, those interactions become just as important to govern as the models themselves.Microsoft has recently reached a similar conclusion, arguing that AI agents require the same governance rigor that organizations already apply to human identities, including clearly defined access and life cycle management. That raises the standard for governance.Organizations increasingly need to demonstrate that AI systems operate within established policy, because once an agent accesses information that it wasn't authorized to use, the governance failure has already occurred. In the age of AI agents, governance is increasingly measured by what an organization can prove. Why The Data Layer Remains Largely Ungoverned The first generation of AI governance focused on the outputs AI produces and the risks those outputs can create. Those efforts were both necessary and overdue.However, they assumed that the primary source of risk lies within the model itself. AI agents expose the limits of that thinking by extending decision-making into the data layer. In those environments, the greater risk often lies less in what the model generated than in the information it was allowed to access in the first place. The gap between governance ambition and governance reality is already visible. According to research conducted by my company this year, 97% of organizations have deployed or piloted AI agents. At the same time, 94% struggle to manage unstructured data, and only 31% say agent access is centrally managed and automatically enforced.Taking it a step further, Cisco's 2026 Data and Privacy Benchmark Study reported that roughly two-thirds of organizations struggle to efficiently access high-quality data for AI initiatives. The numbers show that many organizations are still wrestling with the fundamentals of data visibility and control, even as AI systems gain broader access to enterprise information.None of this suggests a lack of enthusiasm for AI. If anything, it suggests the opposite. Organizations are moving aggressively to adopt AI agents, but the governance mechanisms required to control how those systems interact with enterprise information are developing much more slowly. Reducing Fragmentation, Increasing Control Gaining control over AI governance starts with understanding how enterprise data is accessed and used. Visibility into the operational data layer lays the foundation for accountability, enabling the connection of AI-generated output back to the information that informed it. For many organizations, that begins with reducing fragmentation. Governance becomes significantly harder when data and permissions are scattered across disconnected systems. A fragmented environment makes governance difficult to enforce consistently because visibility is scattered across disconnected systems. Organizations already have governance practices for people accessing enterprise data. AI agents should operate within those same boundaries. As AI agents become more deeply embedded in day-to-day operations, those capabilities will determine how confidently organizations can scale AI. Moving Governance Down The Stack Early governance efforts focused on model behavior, and that made sense when the primary concern was what AI might generate. Now that agents have broader access to enterprise data, governance is shifting from intention to enforcement.In the age of AI agents, governance is increasingly measured by what an organization can prove. Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?