McKinsey’s recent article on bridging the gap between AI agents and ERP systems captures a reality many enterprises are only now confronting: meaningful AI value does not come from better answers, but from executed outcomes.
As long as AI remains assistive—summarizing, recommending, drafting—it improves individual productivity but rarely shifts enterprise performance.
“Value emerges only when AI can initiate and complete work across systems of record.”
That shift—from assistance to execution—is where ERP, CRM, HR, finance, and identity platforms become central. On this point, McKinsey is unequivocally right.
They are also right to emphasize that once AI agents act, questions of risk, security, and governance move from theoretical to operational. An agent that can transact is no longer just software; it becomes part of the enterprise’s execution fabric.
However, large organizations encounter a deeper boundary even before ERP integration becomes the dominant challenge. In practice, the harder divide is not technical. It is organizational.
When an AI agent approves an expense, provisions access, disables an identity, or commits a transaction, the enterprise does not primarily ask how the decision was made. It asks who was authorized to let it happen, who owned the outcome, and who could have intervened if it went wrong.
“At the moment of irreversible execution, enterprises care less about how a decision happened than about who was accountable for it.”
These are not questions of architecture or policy enforcement. They are questions of authority and accountability.
Most current approaches to agentic AI focus on strengthening controls. Policies are refined, identity boundaries tightened, logging improved, approvals introduced, and audit trails expanded. All of this is necessary. None of it is sufficient.
Controls explain how something happened and what occurred. They do not answer who was responsible at the moment an action became irreversible. In traditional organizations, this gap is not solved by ownership alone. It is solved structurally—by combining accountability with clearly bounded authority. Managers are comfortable owning outcomes not because they are named as owners, but because roles are explicitly defined, scopes are limited, and the tools available to an employee are deliberately tailored to their job. Responsibility becomes acceptable only when autonomy is constrained by design.
When AI agents begin executing work, these structures do not automatically carry over. Without an explicit organizational model, enterprises end up with autonomous actions that are technically compliant but organizationally ambiguous. Failures may be logged perfectly and still leave leadership asking the only question that really matters: who owned this?
“Without explicit ownership, autonomous execution becomes observable but not controllable.”
This is one reason so many agentic initiatives succeed in pilots but stall at scale. The technology works, integrations function, and controls are in place—yet progress slows. The bottleneck is not models or APIs. It is the absence of an operating model for autonomous execution.
At some point, enterprises must decide whether an AI agent is merely a tool or whether it is an actor embedded in the organization. As long as agents are treated purely as technology components, governance remains external and accountability diffuse. But once agents operate end to end, organizations intuitively want the same things they demand of human employees: defined scope, clear ownership, explicit authority, and the ability to intervene when needed.
This is why some organizations are beginning to talk about moving from AI agents to what are sometimes called digital employees—not as a metaphor, but as an organizational construct. The terminology may evolve, but the need is clear. Enterprises require a way to embed autonomous execution into their operating model without losing control.
McKinsey is right to push the conversation beyond copilots and chat interfaces. Bridging AI agents to ERP systems is a necessary step toward real value. The next step is to bridge autonomous execution to enterprise authority.
The future of enterprise AI will not be decided by smarter models alone, but by how well organizations learn to place autonomy inside structures of accountability they already understand.