Capacity / What limited it?
Look at what capped the useful gain.
Demand, delivery capacity, quality, adoption and cost can limit what follows a faster task. Look for the next actual constraint before adding more complexity.

- Useful demand
There must be work worth doing with the capacity.
- Accepted quality
Review and rework can consume a seeming gain.
- The next handoff
The rest of the business has to carry the extra work.
Open the detail
Keep the setup, assumptions and checks in view.
What limited the upside?
What limited the upside?
What stopped the improvement from becoming larger? Free hours cannot pass the tightest link. This check is useful because it shows what to fix next.
The third client waits however many hours are free. More released time alone would not change this week’s result.
Check each link in order
1Hours releasedRoom to work, not yet a result
This is where the chain starts. See how released hours are measured in What changed?
2Output completedDelivery · where this example stops
Next question: how much can the team actually start and finish this week, and what sets that number?
If delivery is the limit, more hours elsewhere will not move the result. The next improvement belongs here.
3Quality acceptedQuality · does it stay accepted?
Next question: how much work comes back? Track rework rate so faster work stays accepted work.
4Demand servedDemand · is real work waiting?
Next question: how much real client work is waiting? Count it before assigning freed hours. With none waiting, the time back has nowhere to go yet.
5Margin realizedCost · what is left after costs?
Next question: what remains after costs? Subtract AI setup and running cost before calling anything a return.
Then repeat: fix the tightest link and check again, because the limit can move to another link. A single percentage does not show that.
Original checks and two published studiesDifferent work and different measures: sourced support, not a promised return
What limited the upside?
Demand, delivery, quality, cost
- Demand · Count the real client work waiting before assigning freed hours to it.
- Delivery · If 3 new clients are waiting and the team can start 2 this week, the third waits however many hours are free.
- Quality · Track rework rate so faster work stays accepted work.
- Cost · Subtract AI setup and running cost before calling anything a return.
- Hours released
- Output completed
- Quality accepted
- Demand served
- Margin realized
Two published examples · different work, different measures
Writing tasks · In a controlled test on selected professional writing tasks, people using ChatGPT took 40% less time on average, and graders scored their work 18% higher. Noy and Zhang, Science, 2023
Customer support · At one software company's support team, AI suggestions helped agents solve 15% more customer problems in the same staff time on average, across 5,172 agents. Results varied by person. Brynjolfsson, Li and Raymond, QJE, 2025
These studies measured different things in different settings. Your own numbers come from your own task logs.
For the setup examples · count the client work waiting each week before assigning freed hours. With none waiting, the time back has nowhere to go yet.