Measure the work left after the AI finishes

The time until an AI output appears is not necessarily the time until the work is usable. For an agency or underwriting team, the practical endpoint should be a reviewed result ready for its next authorized use.

NIST recommends defining business value, demonstrating performance under conditions similar to deployment, and documenting measurement limitations (NIST AI RMF Core). The timing method below is an authored workflow recommendation, not an industry benchmark.

Measure elapsed time from the agreed starting event to the finished result. Separately measure human touch time: minutes a person spends preparing input, operating the tool, reviewing output, correcting it, and handling rework.

Do not add unattended processing time to labor time. Conversely, do not delete waiting time when reporting how quickly a client or colleague receives the result. The two clocks answer different business questions.

Use the same task boundary for the manual and AI-assisted workflows. Note document complexity, user experience, and any assistance. If one process receives a pre-cleaned packet while the other begins with an unorganized inbox, disclose the difference.

A repeated packet can introduce learning effects: the reviewer may remember it on the second pass. As an editorial testing recommendation, alternate the order where practical or use comparable cases, and disclose the approach rather than describing the comparison as perfectly controlled.

Record preparation, active operation, review, correction, and later rework as separate fields. Record failed attempts and manual fallback rather than discarding them from the timing set.

In a synthetic example, a manual workflow takes 30 minutes of human effort. The AI-assisted version takes four minutes of preparation, two minutes of active operation, 12 minutes of review, and seven minutes of correction: 25 minutes total. Human effort falls by five minutes, or about 16.7%, not by the difference between 30 minutes and the generation time.

These numbers illustrate the calculation only. They do not represent a tested tool or a promised result.

Show the number of evaluated cases, their mix, the observation period, and any exclusions. Report the spread of results as well as a typical value when the sample supports it. Put critical defects and fallback counts beside the timing result.

A faster draft that requires an unacceptable release risk is not an accepted workflow. A slower draft that improves another outcome may still be worth evaluating, but name that outcome instead of calling it time saved.

Next: Record both clocks in the pilot worksheet and pair them with the accuracy definitions.