AI Finished Your Work. Whose Work Did It Start?

Imagine a company that uses AI to prepare proposals faster, produce reports in minutes, and answer customer questions instantly. Department presentations show the hours saved. Everyone appears more productive.

Yet the same people are still in the office in the evening. The finance manager is correcting assumptions in the report, the experienced engineer is reviewing the code, and the customer service representative is sorting out the issue the automated reply could not resolve.

This raises a question: if everyone is saving time, where is these people’s time going?

There is an easily overlooked item in the calculation of AI investments: the time one employee saves becoming another employee’s workload. This transfer often goes unseen in productivity reports. The first employee’s delivery time falls; the second employee’s corrective work is treated as part of their job.

A September 2025 survey by BetterUp Labs and Stanford Social Media Lab of 1,150 US desk workers examines this problem through the concept of workslop: AI output that looks professional but requires someone else to think again and do more work before it can be used. Some 40% of respondents reported encountering such content in the previous month, with each incident taking about two hours on average to resolve. These self-reported findings should not be read as a definitive loss calculation applicable to every company. But they are valuable in showing how costs can change hands. Study⁠

The management issue is bigger than a faulty report.

The employee who sends the report quickly is praised, while the employee who explains why it cannot be used may appear to be delaying delivery. Under such a performance system, the conscientious employee pays the price of careful work. As long as the cost of sending unchecked work can be passed to someone else, the company unintentionally encourages that behavior.

This resembles an internal version of what economics calls an externality: the benefit and cost of a decision fall on different people. A department may look efficient on its own terms without improving the company’s overall outcome.

AI can therefore amplify an old flaw in an organization’s performance system. A management team that counts delivered files can now count many more of them.

Are employees aware of this?

Not always. In a controlled experiment conducted by METR in 2025, 16 experienced developers worked on 246 tasks in projects they knew well. Tasks where AI use was permitted took 19% longer on average. Yet at the end of the experiment, participants believed the tools had made them about 20% faster. Research paper⁠

It would be wrong to apply this figure to all of today’s tools. In its February 2026 update, METR said newer tools might be producing speed gains, but selection and measurement problems in its follow-up study prevented a reliable estimate of the effect’s size. Still, the original experiment offers managers an important warning: work feeling easier does not, by itself, show that it consumes fewer resources. Update⁠

There is also strong evidence of productivity gains. Research published in the Quarterly Journal of Economics, covering 5,172 customer support employees, found that AI assistance increased the number of customer issues resolved per hour by 15% on average. The outcome measured matters: not the number of answers produced, but the number of problems solved. Study⁠

A company’s investment calculation should begin there too. Report counts are easy to measure. Measuring how much a report contributes to a decision takes more effort. When we treat what is easy to measure as success, technology can produce impressive results on that indicator while leaving the company’s actual need unmet.

There is also a cost to expertise.

If an experienced finance manager spends all day checking sources, reporting periods, and assumptions, what work is not being done during those hours? Is a financing option going unexplored, a collections problem being spotted late, or an investment decision waiting?

The cost of correction is not just the hours spent. It includes the value of the work that could not be done in those hours. A company that turns senior employees into perpetual output cleaners may be saving money through AI while consuming its scarcest resource.

That is why the point at which expertise enters the process matters. An expert who searches for errors at the end of hundreds of outputs soon reaches capacity. If that expert defines the questions, acceptable sources, and results that require review at the outset, the same knowledge can be applied to much more work.

My conclusion is not to add human checks indiscriminately at every stage. Unnecessary review creates costs too. A low-consequence draft and an analysis affecting an investment decision should not pass through the same process.

The starting point for measurement is clear: who spends how much time between an output’s initial creation and its becoming usable, how often it is sent back, and what it ultimately resolves.

Viewed this way, a more meaningful distinction emerges than employees who use AI versus those who do not: people who complete their work, and people who hand off work that merely looks complete.

When we see the hours saved in the next productivity presentation, let us also ask whose hours were saved.

Because a company’s fastest employee may sometimes be not the person who finishes their work quickest, but the person who passes the unfinished part to someone else quickest.