Ask people how much time AI saves them and the answer is encouraging. Glean's Work AI Institute found workers estimate roughly 11 hours a week. Ask what they do with those hours and the picture changes: 6.4 hours get spent managing the tools themselves.
The time goes to supplying context the tool didn't have, checking what came back, rerunning prompts, and cleaning up answers that read well and weren't right. The gain in capability for teams with AI tools is real, but it's worth asking whether your team is spending the time it saves managing the tools instead of using it.
If the tools keep getting more capable, shouldn't the cleanup shrink on its own? Some of it will. But no model decides what good output looks like before anything is generated, and none confirms that what comes back meets that standard.
That pattern isn't confined to the people using the tools every day. Just over half of Canadian professionals say their employer has put most or all of its AI resources into new technology, while only 11% say the focus has been on developing people's skills to use it well, according to a Capital One survey of workers in technology and business roles this spring. Go up a level and the picture doesn't improve: Canadian CIOs and CTOs told IBM they expect to be running an average of 1,189 AI agents by 2027, a 36% increase from today, and only 9% of them say they feel fully prepared for that volume. In both cases, the investment went into capability, not into the decisions that make it pay off.
There's reasonable evidence that setting that standard and checking against it are managerial skills rather than technical ones. An NBER working paper by Weidmann, Xu, and Deming put 249 people through collaborative problem-solving tasks, leading both AI agents and human teams, and found their performance across the two correlated at 0.81, holding at 0.69 after controlling for task-specific ability and fluid intelligence. The people who did well in both settings asked more questions and encouraged more back-and-forth, rather than simply issuing more instructions.
What makes someone good at directing AI work looks a lot like what makes them good at directing people's work: set expectations clearly, give the context that's genuinely needed, review what comes back on its merits.
Which means most organizations already have this capability. Among the business leaders we work with, the ones getting real value from AI are the ones who assigned both roles deliberately. Someone decided what "good enough to pass along" means for a specific kind of work, wrote it down, and someone else owns the check. So before your next rollout widens, whether that means a new department or just your own workflow, it's worth confirming both of those have an owner.
If your team is spending more time managing AI than using it, that's the kind of problem we help organizations work through.