What the study reports
The November 2024 revision of Generative AI at Work, by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, studies an AI assistant used by 5,172 support agents. Its abstract reports an average 15% increase in issues resolved per hour, with differing effects by worker experience. Less experienced workers benefited more; the most experienced workers had small speed gains and small quality declines.
Read the version-specific abstract. This is evidence about an assisted support workflow. It does not measure an independent business’s profit or establish a number of jobs it can remove.
Three conversions to avoid
Output and time use different denominators. In an illustrative calculation, increasing throughput from 10 to 11.5 issues per hour reduces the time for 100 issues from 10 to about 8.70 hours: roughly 13.0% less time, not 15%. This is arithmetic under fixed-workload assumptions, not an observed result from the study.
Released time is not automatically cash saved. An owner may spend that time on acquisition, product work or rest. Record cash savings only when an expense actually falls. Keep estimated time value separate.
An average is not your forecast. Your knowledge materials, case mix and exception rates may differ. Multiplying a study’s percentage by your entire payroll would exceed the evidence.
Run a bounded pilot
The following is an editorial proposal, not the study’s protocol. Select a repeatable class of product questions. Exclude disputes and exceptional commitments. Record the existing process, then trial AI drafts on comparable issues, with the same accountable operator reviewing before sending.
Track preparation, generation, review and rework time. Also record whether the issue was resolved, reopened or escalated. Measuring only generation speed misses the work required to make a draft usable.
Download the blank pilot worksheet. It contains no assumed benefit or invented business outcome.
Decide before expanding
Before the pilot, state the smallest useful improvement and unacceptable errors, such as fabricated product terms. Compare total labor time and resolution quality. A small sample or changing case mix calls for more evidence, not a stronger claim.
If review consumes more time than drafting saves, narrow the question category, improve source materials or stop. If repeatable work falls without a quality loss, decide how the released time can actually be used.
Include software charges, knowledge maintenance, review and exception handling in the cost record. If you value time at an hourly rate, label that rate as an assumption. A useful business analysis states the task, population and conditions behind its result; it does not fill missing operating data with a revenue story.