AI can help you win the task and lose the system. That is the automation recovery tax.
Imagine this: a team automates a three-hour task down to twenty minutes. That looks like a win. But then someone has to review the output, handle the exceptions, investigate unexpected behavior, revert bad changes, repair broken data, or explain the incident afterward.
The original task got cheaper. The surrounding system got more expensive. This is a familiar strategic failure: you can win the battle but still lose the war.
Automation ROI is often calculated battle by battle. How much faster was this task? How much more code did we produce? How many tickets did the agent close? How many documents did it generate?
But organizations do not operate as collections of isolated tasks. They operate as systems.
Reuters recently reported that as Meta dramatically increased technical output using AI, major technical and security incidents rose 40 percent, while employee time spent firefighting those incidents rose as much as 70 percent.
That is the part that should set off alarms for managers.
More output did not simply produce more value. It also produced more failure, and substantially more human time spent cleaning up after it.
That firefighting is only the visible part of the bill. The recovery tax also includes reviewing generated code, verifying AI output, handling exceptions, investigating unexpected agent actions, responding to incidents, remediating failures, and restoring systems when automation goes sideways.
Those hours are easy to hide because they rarely arrive labeled “cost of AI.” They show up on somebody else’s charge code.
I keep coming back to this because I work inside several of the systems that absorb that hidden cost: AI governance, security, systems thinking, documentation, and project delivery. Security sees the incident load. Governance sees the oversight burden. Delivery sees the rework and exceptions. Documentation sees the verification problem. Systems thinking asks the bigger question: did making one task faster actually improve the system?
That is the question an AI business case has to answer.
Not just:
What human work did automation remove?
But also:
What human work did automation create somewhere else?
Because saving hours on one task is not much of a productivity win if those hours simply reappear somewhere else under recovery, remediation, and response.