A paper strip connects a folded blue section to a light-colored bundle and passes beneath a desk magnifier.

Where the savings from automation end

Savings in a task are not the same as savings across a process or a budget. How to follow review, delivery and the use of freed time without denying a real improvement.

Alejandro Mayorga6 min read

In the support service studied by Erik Brynjolfsson, Danielle Li and Lindsey Raymond (opens in a new tab), AI assistance increased average productivity by 15%, measured as issues resolved per hour. The research, published in 2025, examined a specific tool and operation; it does not establish how much any other process would improve. It does show something worth retaining in the discussion: a technology can produce a real operational gain. We do not need to begin by suspecting that every saving will be canceled out by a burden we have not yet seen.

Where the calculation was made matters, however. Resolving a support issue is a different unit of output from generating a response. Likewise, producing the first draft of a report is not the same as making it ready for use. Between those points there may be checks, corrections and integration with other materials. Some already existed before automation; others may change with the new way of producing. To know how much work was saved, we need to compare both sequences, neither charging the technology with everything that happens afterward nor leaving all of it out of the calculation.

Following the report through to completion

Consider a hypothetical example of report preparation. We compare a completed unit, with the same scope and quality requirements, before and after automating draft preparation. The figures below are invented to examine the comparison: they do not estimate the effect of any tool. They represent minutes of human work, not system processing time.

Work per report Before After
Preparing the draft 30 min 5 min
Reviewing and correcting 15 min 20 min
Preparing delivery 5 min 5 min
Total 50 min 30 min

The draft frees up twenty-five minutes, and review requires five more. That leaves twenty fewer minutes of work per report. In this example, the additional checking reduces the saving but does not eliminate it. Counting only the first row would exaggerate the improvement; presenting the twenty minutes of review as an entirely new burden would also distort the comparison. Fifteen minutes were already devoted to that activity before. The change we need to explain is the difference, along with whatever happened to the result.

The quality condition cannot be settled by the table heading. In a real trial, we would have to check that reports still serve the agreed purpose, including corrections that emerge after delivery. If the new version contains more information, handles harder cases or requires the recipient to redo part of the analysis, we are no longer comparing exactly the same work. There may be a valuable expansion of scope or a transfer of burden, but it would need to be named before the entire difference could be attributed to efficiency.

It also matters to keep those receiving the result within the observation. A file may leave the area preparing it sooner and then require more effort from another area to make it usable. If that second part is excluded, the first area’s saving describes a local improvement, not necessarily the effect on the process. The comparison must extend to a recognizable result for whoever needs it, not to the most convenient point for demonstrating that automation worked.

What we can do with the time left over

The twenty minutes in the example are not yet an equivalent reduction in delivery time or a payroll saving. They may be spread across people, arise at different times or be used to handle work previously postponed. Reducing the effort needed for each report expands the options, but turning that improvement into more deliveries, less overload or lower cost requires observing what actually changed in the organization.

That difference in scale appears in Eleanor Dillon and her colleagues’ experiment on generative AI and work patterns (opens in a new tab). The May 2025 working paper, based on randomized access to Microsoft 365 Copilot, found reductions in time spent in Outlook but no statistically significant changes in meeting time recorded in Teams. The authors distinguish behaviors an individual can change alone from those requiring coordination. Their data observe activity in those applications, not content quality or the entire working day. The contrast allows us to retain both an individual improvement and the lack of evidence for an equivalent change elsewhere in the work.

Let us return to the reports. If freed capacity allows more units to be produced, review will also receive more reports, even if each requires less total work. That increased workload does not demonstrate that automation has failed: it could accompany desirable growth in the service. But it does not allow the same deadline to be promised without examining whether the next stage can absorb it. The new production capacity changes decisions about how much work to accept and how to distribute it.

There may be another legitimate use of the time: reducing after-hours work, allowing more careful review or restoring an activity for which there was no room. Those outcomes deserve to be checked in their own right. If personnel spending has not fallen, they should not be presented as though that budget saving had already occurred. And if the improvement consists in maintaining quality within a less overloaded working day, measuring only how many additional reports were sent would leave out precisely what the organization decided to achieve.

The table does not contain the solution’s full cost either. Configuring it, maintaining its integrations, paying for its use and updating controls may require resources that do not recur with every report. The economic assessment needs to include them over an explicit period and workload. That distinction should remain clear: twenty fewer minutes of human intervention per unit and a worthwhile investment are related claims, but the first does not settle the second on its own.

Automating part of the work can yield a defensible improvement even if some checking increases and no job disappears. In our report, the draft saves twenty-five minutes and the completed delivery saves twenty. That difference does not invalidate the improvement. Omitting the five minutes of additional review does not turn them into time saved either.