Insights
The Myth of AI Productivity and Idle Time
A COO told me last month that her AI rollout had “leaked.” Her word. The tools worked, the hours came back, and nothing showed up at the other end. She asked me where the time went. I think we mislabel the answer, and the mislabelling is expensive. A Bank of Korea-funded survey of Korean workers, released in February, found that 51.8% now use generative AI at work and that it reduces their working time by about 3.8%. Then the uncomfortable number: the correlation between an individual’s time savings and their reported change in output was 0.008. Effectively zero. The authors conclude the recovered time is absorbed largely as what they call on-the-job leisure.
A vocabulary problem
Most executives read that as shirking. I read it as a vocabulary problem. Our accounting has exactly two categories for a human hour: output, or waste. Anything that is not visibly producing gets filed under the second one. Now set that against a natural experiment published in Organization Science. A supply-chain shortage forced unexpected stops in production plants. Employees who sat through that unplanned idle time went on to produce 58% more ideas than uninterrupted colleagues in the following three weeks. And the details matter more than the headline. Interruptions without idle time — the intrusions, the pings, the reassignments — reduced creative performance. Planned breaks produced no creative effect at all, because people disengaged cleanly and turned to leisure goals. Read those two findings together and you get something strange and useful.
Unstructured, unexpected, slightly uncomfortable emptiness
Unstructured, unexpected, slightly uncomfortable emptiness is generative. Scheduled emptiness is not. The mind does its recombining while it is still half-attached to the problem it was pulled away from — the researchers land on attention residue as the mechanism. Which means the innovation hour cannot be booked in a calendar. It arrives unannounced, looks exactly like idleness, and gets closed as such. R&D research makes the same point from the other side: most corporate slack time is rhetorical rather than genuine. It exists in the policy, then gets consumed by overcommitment and by people’s own sense that thinking is a luxury they will be judged for. So we are in a peculiar position. AI has just handed organizations the largest supply of unplanned idle time in modern working life, and we have given our managers no language for it other than capacity. I do not think the answer is to schedule reflection. The evidence suggests scheduling it removes the property that makes it work. Two things seem more promising. The first is to stop treating the hour as recovered capacity by default. If every freed hour is immediately refilled with throughput, the organization has quietly converted its only new source of variation into more of what it already had. The second is to lower the cost of acting on whatever surfaces during that hour. Idle time produces ideas; almost nothing in a normal company is built to receive them. There is a route for finished work and no route for an unfinished thought. That is a design gap, not a motivation gap.
Saying yes to mess
Underneath both is an older argument I keep returning to. Two management scholars, Herath and Harrington, published a paper with a title I admire: “Saying yes to mess.” Their claim is that disorganization is not a failure of design but a form of it — the condition under which firms sense and adapt. Optimization removes variation. Variation is where discovery lives. We have just deployed the most powerful optimization technology ever built into organizations that were already over-optimized. I told the COO her time had not leaked. It had gone somewhere her instruments cannot see, and some of it was probably the most valuable thing that happened in her company that week. What I could not tell her is how she would know.