As McKinsey reports, fewer than 10% of enterprises that have deployed AI agents have scaled them to deliver tangible business value. As a consequence, Gartner projects more than 40% of agentic AI projects will be cancelled before the end of 2027 — citing escalating costs, unclear business value, and inadequate risk controls. This is referred to as the gen AI paradox, and it doesn't necessarily reflect a technology problem.
The AI models work and the integrations hold. The agents execute the tasks they are configured to execute. But what fails is everything organized around them: the accountability structures that don't account for a non-human actor in the decision chain; the oversight protocols that assume when a human being is reviewing each step; the workflow architecture built for handoffs between people, not between a person and an autonomous system running a multi-step task sequence without pausing for approval. The organizations implementing agentic AI don't redesign the context, but deploy directly into it and thus assume all the contextual dysfunctions that come with.
We keep seeing this mistake get repeated. An organization identifies a workflow and deploys an agentic AI into it. It points to the AI as the unit of transformation, but what it doesn't touch is the organizational structure the AI is operating inside – who owns its outputs, who is responsible when it acts on incomplete information, who is accountable when it hands off to the wrong downstream process. The AI does exactly what it was built to do. The organization fails to redesign itself to govern what the AI is doing. When something goes wrong, the failure narrative is technological: the model hallucinated, the integration was incomplete, the training data was biased. The more uncomfortable diagnosis, however, is sitting in plain view. It's often the case that the organization failed to adapt to the technology properly. Put simply, organizations running agentic AI have thus far not updated their "operating system" to accommodate such an impactful technology.
The conversation about accountability in agentic AI is not ours alone. McKinsey has named it. Deloitte, PwC, Harvard Business Review, and researchers at Berkeley's California Management Review have each, in the past twelve months, published analyses pointing toward governance and organizational design as the defining failure modes. That volume of coverage is itself a signal, and a strong one at that. The diagnosis is largely accurate, especially considering AI in enterprise, AI in tech, and AI in knowledge work. But it also extends outward – the same failure pattern also appears simultaneously across sectors.
Organizations running agentic AI have thus far not updated their "operating system."
In financial services, agentic AI is executing transactions and flagging compliance issues without clear protocols for human review before consequential decisions proceed. In healthcare, the FDA loosened oversight of AI-enabled clinical decision support tools in January 2026 – framed as innovation-enabling. While the deregulation may spur some degree of innovation, it's also accountability-reducing, and at the exact moment when clinical workflow integration is accelerating. In media, autonomous editorial agents are operating in content pipelines where the boundary between AI-initiated and human-reviewed output has not been architecturally defined.
The organizational design question is identical across all three: who is responsible for what AI decided, and what is the protocol when it gets it wrong? These sectors are not sharing a problem because they use the same technology. They are sharing a problem because organizations across every sector are making the same architectural omission – and they are making it because the frameworks for agentic governance were never built into the deployment, but only applied afterward, in some cases.
The answer is organizational redesign, and it belongs in a specific place. Not in IT or procurement, but in strategy and design. It is about accountability mapping that assigns explicit ownership of each agent's outputs; oversight protocols that specify where human judgment is required before an AI-initiated action proceeds; human-agent workflow architecture that distinguished between tasks the agent executes autonomously, tasks it executes with human review, and tasks it should not execute at all. These are design decisions, not compliance checkboxes. They require systems thinking at the level of the whole workflow – not at the level of the individual AI tool.
Implementing and establishing a resilient AI stack in today's volatile and uncertain context is challenging, for sure. But organizations that treat this as a technology procurement question will continue producing failures at the pace of their deployments. The ones that treat it as an organizational design question will build the structural conditions in which agentic AI actually performs as intended. The window to do this by deliberate design – before an incident, a board mandate, or an EU AI Act-style regulatory framework imposed from outside forces the architecture – is open. But it probably won't stay open indefinitely.
