AI agents in IT operations are reaching a new stage: they analyse a fault, propose a fix and carry it out themselves once a person approves. What determines practical value is less the analytical capability than the design of that approval step.
Pure recommendation systems save little time, because the specialist ends up doing the work anyway. Agents without any approval are not authorisable in most organisations as soon as they write into systems. The middle path works only under three conditions: the approval shows concretely what will change, otherwise it becomes a routine click. Permissions attach to the type of task rather than broadly to a person. And logging captures the reasoning, not just the command.
Benefit typically sits with frequent, uniform cases, where risk is lowest too. Anyone wanting to demonstrate value should measure the effort for the three most common cases before an agent goes live. In business processes the requirements rise considerably: an agent restarting a service causes different damage from one that triggers a purchase order.