Short answer: Generative AI has reduced task time in some controlled experiments and specific workplace settings, but the results do not predict how much time every business will save. The task, user experience, quality checks, and integration into the real workflow all matter.
What two studies found
In an experiment on professional writing tasks, Noy and Zhang reported that participants with ChatGPT access completed tasks 40% faster on average and received output-quality ratings 18% higher. The study tested incentivised writing tasks with college-educated professionals; it does not establish a 40% saving for every job, company, or AI rollout. Science paper
Brynjolfsson, Li, and Raymond studied a generative AI assistant deployed among customer-support agents. Their final published paper reports a 15% average increase in issues resolved per hour across 5,172 agents, with substantially different effects by experience and skill. An earlier NBER working-paper version reported 14% across 5,179 agents; these are two versions of the same study, not independent results. This is evidence from one customer-support setting, not a forecast for Indian businesses generally. Quarterly Journal of Economics paper · NBER working-paper version
The studies are useful because they measure work rather than tool sign-ups. Their differences also show why “AI saves X%” is too broad to be a responsible promise.
Why a company may see a different result
An AI assistant may shorten drafting but add review work. Poor source data may create more corrections. A model may help new employees more than experts, or help one task while leaving the wider process unchanged. Time saved can also become more customer conversations or backlog reduction rather than fewer paid hours.
How to test it in your business
- Choose one narrow, frequent task and define its output.
- Measure baseline active time, elapsed turnaround, quality, rework, and exceptions.
- Pilot with a representative group and the same case types.
- Include prompt preparation, checking, and failure handling in the new time.
- Compare quality and cost as well as speed; stop or adjust if error or rework rises.
Fictional illustration: Neha’s 8-person support team drafts 120 routine replies weekly. A 4-week pilot measures 3 minutes less drafting per accepted reply, but review adds 1 minute and 10% of cases need full manual handling. Net hours are calculated from accepted replies only, with the exception work subtracted. Those figures are made up to illustrate the measurement method, not a Formula Works project or an expected result.
See how to measure time saved by automation and how to measure AI accuracy before using a productivity claim.
Written by
Formula Works
Published
Sat Oct 17 2026