Generation time is easy to notice because a tool displays it. Review time is often less visible because it appears later in another person's day. A useful workflow model includes both, plus the time spent correcting or replacing unsatisfactory outputs.
Follow the work through four stages
Intake checks whether the task is ready. Generation creates a draft. Review checks every output against the acceptance criteria. Resolution handles corrections and escalation. These stages can overlap in practice, but the effort at each stage still needs to be counted.
The Precogniz simulator models generation, review and additional rework. It does not model preparation, queues or parallel scheduling. That boundary matters: a total of ten effort-hours does not mean the work finishes ten hours after it starts.
Calculate expected effort
For each task, expected assisted minutes equal generation minutes plus review minutes plus error probability multiplied by additional rework minutes. For 100 tasks with 2 minutes of generation, 5 minutes of review, a 10% rework rate and 10 extra minutes per reworked task, the result is 100 × (2 + 5 + 0.10 × 10) = 800 minutes.
If manual effort was 20 minutes per task, the baseline is 2,000 minutes. The difference is 1,200 minutes, or 20 hours. The modeled effort reduction is 60%. These are constructed inputs, not measured performance of an AI provider.
Stress-test the review stage
If review rises to 15 minutes and 40% of outputs need ten minutes of rework, assisted effort becomes 100 × (2 + 15 + 0.40 × 10) = 2,100 minutes. The workflow now uses 100 more minutes than the manual baseline. Faster drafting has not produced a net effort benefit.
Possible responses include narrowing the task, improving input quality, changing the output format or keeping the task manual. Removing review merely to improve the modeled result is not a sound response when checking remains necessary.
Distinguish expectation from a real batch
A probability model can produce a fractional expected rework count. Ten tasks at a 15% rework rate imply 1.5 expected reworked tasks across repeated comparable batches, not half of a physical task in this particular batch. Actual outcomes vary.
Use observed error categories to improve the process. Repeated missing fields may suggest a structured output format. Unsupported factual claims may require stronger evidence constraints. A problem involving missing source information cannot be solved reliably by asking the system to sound more confident.