AI does not simply remove work. More often, it removes the predictable portion of work first—and leaves people with the cases that are unclear, risky, emotional, incomplete or expensive to get wrong.
That is the central reality of AI exception handling. A system may answer standard customer questions, flag apparently fraudulent claims, summarize medical records, rank job applicants or route deliveries. But when the information conflicts, the customer disputes the result, a policy has competing interpretations or the model is uncertain, someone still has to take responsibility. That person may have less routine work than before. They may also have a more demanding job.
The result is an exception economy: an organizational layer of review, escalation, correction and accountability that grows around automated systems. It is easy to miss because it is rarely the headline metric in an automation project. Leaders can count transactions processed without human intervention. It is much harder to see the time spent untangling the transactions that did not fit.
This matters for the future of work because a reduction in visible labor is not necessarily a reduction in total labor. It may be a transfer of labor—from frontline staff to specialists, from routine handling to quality assurance, from customer-facing teams to back-office reviewers, or from employees to customers asked to navigate a self-service system. Whether that transfer produces genuine productivity depends on the quality of the outcome, not merely on the percentage of cases an AI can process alone.