The most cited field experiment on generative AI and knowledge work came out of an academic setting: Noy and Zhang gave professional writing tasks to hundreds of college-educated workers, half with access to ChatGPT.

What they found

MeasureResult
Time to completeAbout 40% less for the AI group
Output quality (blind ratings)Higher for the AI group
Quality gap between workersCompressed
Effect by initial skillLarger relative gain for lower-rated writers

The “compression” result matters: AI helped the weaker writers most, echoing the customer-support finding that gains concentrate among less experienced workers. The average number hides a distribution.

Caveats to carry into a business case

Gains are real but uneven Task time about -40% Quality ratings higher on average Skill gap compresses Not a headcount result: task time ≠ cost saved measure your own workflow

VERDICT

Expect real speed and quality gains on well-specified writing, concentrated among less experienced staff. Do not extrapolate the 40% into a company-wide number without measuring your own tasks.

Sources