AI workflows · measuring a pilot
Measure the whole AI workflow.
A faster draft is one part of the result. Count preparation, checking and corrections against the time it takes to produce an accepted output.
Fictional document task · minutes per case
The draft takes six minutes.
The accepted result takes eleven.
*Expected rework: one quarter of cases × eight minutes. Illustrative assumptions.
22 September 2026
A projection needs a local test.
Hong Kong’s Labour and Welfare Bureau published its mid-term manpower update on 22 September. Page 23 projects a 20–30% individual productivity improvement from AI in applicable roles. It also discusses redeploying people to higher-value work.
That is a labour-market projection. For an operations manager choosing a first workflow, the practical question is how much effort an accepted case takes in their own process. The report’s percentage is not a measured saving for your team.
The example below treats document preparation as one part of a larger job. Its timings are fictional. It shows how the result changes when a reviewer needs longer, when more cases need correction, or when the rest of the job takes most of the time.
Source: Mid-term Update of Manpower Projection, page 23.
Try the assumptions
What reaches the end of the workflow?
Compare the same accepted output and quality standard. This browser calculator uses fixed arithmetic. It runs no AI model and connects to no business system. A batch means the number of cases you choose to compare.
Keep effort, throughput and money separate.
In the starting example, the task falls from 12 to 11 minutes. Add 18 minutes of unchanged work and the whole job falls from 30 to 29 minutes. Across 100 cases, that releases 100 minutes of effort. Whether that becomes useful capacity depends on how the work is scheduled and where the next bottleneck sits.
A 20% reduction in effort and a 20% increase in output per hour are different calculations. At unchanged quality and full use of the released capacity, a task falling from 10 minutes to 8 minutes permits 25% more output per hour. This calculator reports effort reduction; it does not convert released hours into payroll savings.
The model includes one average allowance for rework. Use that allowance for all correction and re-checking effort per affected case. Implementation, training, software charges and ongoing maintenance sit outside the calculation. Waiting time also needs separate measurement if elapsed turnaround matters.
Record what a reviewer accepts.
Choose a repeatable output with an owner and an agreed quality check. Compare cases of similar difficulty, record the manual baseline, then time assisted preparation, review and corrections separately. Keep rejected outputs in the record; removing them would flatter the result.
Record errors alongside time. A faster workflow that changes an amount, misses a condition or sends something without approval needs correction before it is expanded. The person responsible for the process decides whether the observed results justify the next step.
Explore more AI workflow examples in A closer look.
Start with one repeated task.
Bring the current steps, a representative example and the checks an acceptable result must pass.
Discuss an AI workflow · How Enzwa approaches AI workflowsSources and method
LWB announcement, 22 September 2026
Full report, page 23
Sources checked 23 September 2026. All calculator timings and scenarios are fictional Enzwa examples. No client result or measured AI performance is claimed.