The Unallocated Hour

It is worth pausing on why this study outweighs the survey headlines it contradicts. Almost everything written about AI productivity rests on self-report, people telling researchers they feel faster. Register data removes the feeling: earnings, hours and job moves come straight from administrative records, worker by worker, across the eleven occupations most exposed to chatbots. The savings themselves are unspectacular but solid. Adopters in the Danish sample report saving about 3% of their work hours, consistent with Bick, Blandin and Deming's representative US surveys, which put the economy-wide figure at roughly 1.4% of all work hours. And yet even the heaviest users, the people saving more than an hour every day, show no detectable movement in what they earn or the hours they record. The tool works. The economics do not move. Something between the saving and the balance sheet is absorbing the hour.

The paper contains the number that explains it. Asked what they do with the time chatbots save them, 85% of users say they reallocate it to other tasks in the same job. Only 8% report that AI has given them entirely new work, a figure that roughly doubles, to about 18%, where employers actively push adoption. Read those shares together and the null result stops being mysterious. The freed hour is not banked, priced or redeployed. It dissolves back into the job it came from, spread thinly across the same tasks, the same inbox, the same meetings. Even in best-practice workplaces, where 19% of workers save more than an hour a day, earnings and hours stay flat. Nobody decided where the hour should go, so it went where hours go when nobody decides.

There is a second leak, and it sits upstream of the allocation problem. In January 2026 Workday published research conducted with Hanover Research across 3,200 employees at firms above $100 million in revenue. It is vendor-commissioned work and should be read as such, though the finding runs against the vendor's interest: 85% of employees report saving between one and seven hours a week with AI, and nearly 40% of that reported saving burns off in correcting, rewriting and verifying the output. Only 14% consistently see a clear, positive net outcome, and 77% of daily users review AI output as carefully as human work or more carefully. None of this is an argument against the tools. Verification is simply the cost of using a probabilistic instrument on work that matters. The problem is that almost nobody nets that cost out of the business case, so the hour that survives is smaller than the hour on the slide, and the allocation decision never happens anyway.

So here is the reframe I would put in front of any leadership team. The freed hour is now the most valuable unallocated asset in your company. Every other asset of comparable value has an owner, a budget line and an approval path. This one has none. It is being spent, invisibly and continuously, by whoever happens to be holding it, on whatever work happens to be in front of them, and the Danish evidence says that in 85 cases out of 100 the default wins. The difference between the firms that will eventually show AI in their operating results and the firms producing these null effects is not model choice, spend or adoption rate. Adoption is nearly solved. The difference is whether someone with authority decided, at the level of a specific person and a specific workflow, where the recovered hours go. That is not a tooling question. It is a judgment question, and it is going unanswered in almost every organisation using AI today.