Short answer: Check who conducted the study, what the baseline was, which metric changed, how many cases or workers were measured, and what work remained. “40% faster,” “15% more issues resolved per hour,” and “one day instead of seven” are different measures and cannot be compared directly.
Three published examples and their limits
- In an experiment on professional writing tasks, Noy and Zhang found participants completed tasks 40% faster on average and had output-quality ratings 18% higher with ChatGPT access. The result applies to the study’s incentivised writing tasks, not every workplace workflow. Science paper
- A field study of 5,172 customer-support agents found a 15% average increase in issues resolved per hour after access to a generative AI assistant, with different effects by experience. The earlier NBER working-paper version reported 14% from 5,179 agents; both figures describe versions of the same study. This measures support productivity in one setting; it does not mean labour hours fell 15%. Published paper · Earlier working-paper version
- UiPath’s customer story about AM/NS India reports invoice processing moving from 7–10 days to one day after an automation project. That is a vendor-published customer case study reporting turnaround time. It is not an independently audited result or a statement about Formula Works. UiPath case study
These sources are useful examples of measured work, but they differ in setting, method, and metric. They should not be combined into a universal “AI saves X hours” estimate.
Questions to ask about any success story
- Is the source a peer-reviewed study, a company case study, or marketing material?
- What does “time saved” mean: active minutes, elapsed cycle time, or output per hour?
- Was quality measured, and did rework or exception handling increase?
- Were one-time setup, training, subscriptions, and maintenance included?
- Is the workflow similar enough to your team’s process to be informative?
- Can the result be reproduced in a small pilot with your data and constraints?
Build your own evidence
Use a baseline period and a limited pilot. Record the same volume, active time, elapsed time, quality, and exceptions before and after. Publish a Formula Works case study only when the customer approves the details and the method is clear. Until then, examples on this site are labeled as fictional illustrations.
For a measurement template, see how to measure time saved by automation and how to calculate automation ROI.
Written by
Formula Works
Published
Thu Nov 12 2026