A common misconception about AI agents, tested against the evidence.
Automation rarely deletes a unit of work cleanly. It changes where the work happens.
An agent may remove the need for a person to execute ten routine steps. But the system may now need someone to define permissions, inspect failures, resolve exceptions, maintain tools, evaluate quality, recover partial actions and decide when the agent should stop.
That can still be an excellent trade. The mistake is measuring only the work that disappeared from the original role.
McKinsey’s 2026 data shows why the distinction matters: widespread individual productivity gains have not translated into equally widespread EBIT impact. Its coordination-tax work argues that enterprise value appears when organizations redesign end-to-end workflows rather than making isolated steps faster.
Long-running agent systems also make the new work visible. Anthropic describes explicit handoff artifacts and persistent state as necessary for agents working across sessions. OpenAI’s Auto-review automates part of the approval burden around coding agents, which is itself evidence that agent adoption creates a review problem worth redesigning.
The question is not how many tasks the agent removed. It is how much total human effort a successful outcome now requires.
Task demos show the visible before-and-after: a person did the task; now an agent does it.
The displaced work is harder to see because it often moves to another team or appears only after deployment.
The economics of agentic automation depend on the full operating system around the agent. Supervision and recovery can shrink with better design, but they do not disappear by definition.
Measure human minutes per successful outcome, including review, exception handling, recovery and maintenance.
If that number falls while quality and risk remain acceptable, the agent removed work in a way that matters. If not, it may only have moved the work somewhere less visible.
As AI moves from assisting work to leading it, hours saved stop telling the whole story. The scarce resource shifts to human supervision: approvals, exceptions, context and judgment.
AI adoption is rising faster than enterprise value because companies keep installing new intelligence inside old operating models.
RealitySaved time is released capacity. It becomes value only when the organization captures that capacity as better output, lower cost, higher quality, more revenue or less risk.
RealityA human checkpoint is only a control if the person has the context, competence, time and authority to detect a problem and stop or reverse the action.
RealityBad governance creates queues. Good governance predefines boundaries so low-risk actions can move faster without waiting for ad hoc approval every time.