FREE WORKSHEET · NO ACCOUNT

Did AI save time—or move the work?

Measure a completed pilot from input preparation to an acceptable result. Include the output you rejected, not just the draft that looked good.

Published 22 November 2026 · Prepared for InriGPT by Inri Inc · Fictional worked example and a reproducible local calculation.

Agree on a usable result before starting

“The AI answered” is too weak a pass rule for business work. A proposal response might need a source for every compliance assertion. A procedure answer must distinguish a documented rule from a missing exception. A content draft must preserve the facts supplied to it. Choose one rule your reviewer can actually check.

Compare the same completed task under each method. Use a measured manual baseline from comparable work, or mark a rough estimate in the task label and treat the comparison as provisional. Keep the source length, expected quality and final deliverable similar. A rough draft is not equivalent to a reviewed response ready to send.

This worksheet compares active human effort. It does not measure model latency, calendar turnaround, financial return or output quality automatically. Record waiting time separately if that matters to your operation.

Run the effort worksheet

Edit the sample below or clear it for your own records. All times are minutes. Preparation includes finding and redacting sources; review means checking the response; correction means making the draft usable. If you reject the response and finish manually, include that manual completion time in addition to the failed attempt.

The interactive worksheet requires JavaScript. The worked calculation and downloadable sample below remain available without it.

Worked example: four tasks, one rejection

The four rows are invented teaching cases. They are not customer observations, measured InriGPT results or a prediction of your savings. A team spends 60 minutes preparing instructions and a test set. It then completes four comparable tasks.

Fictional pilot, all values in active minutes
TaskManual baselinePrepare + review + correctManual completionAI-assisted total
R1: support response406 + 8 + 4018
R2: content draft305 + 7 + 3015
R3: rejected proposal508 + 10 + 55073
R4: procedure summary254 + 6 + 2012

The manual baseline totals 145 minutes. AI-assisted task effort is 18 + 15 + 73 + 12 = 118 minutes. Add 60 minutes of setup and the pilot takes 178 minutes: 33 minutes more active effort than the baseline. Three of four AI outputs pass the chosen review rule.

Ignoring the rejected task would make the pilot look much better while hiding the most important failure. The response that cannot be used still consumed preparation and review time. Its manual replacement is not free.

What the calculation includes

  • Per task: preparation + review + correction + any manual completion after using AI.
  • Pilot total: all task effort + one-time setup effort.
  • Difference: manual baseline − pilot total. A negative value means more effort, not a saving.
  • Accepted outputs: the number you explicitly mark as meeting the same pre-agreed rule. The worksheet does not inspect or grade the output.

Do not compare unfinished AI attempts with completed manual work. Finish the job manually and record the extra effort, or report that task as incomplete outside this completed-task comparison. Do not quietly remove failures to improve the result. Keep the original records so a colleague can reproduce your calculation.

Decide what to change before expanding

In the example, recurring task effort is lower than the baseline, but setup removes the benefit during this small pilot. Repeating the same result might eventually recover that effort; four fictional cases cannot establish that it will. Document quality, task difficulty and reviewer familiarity can change the outcome.

Inspect R3 before buying more capacity. Was required evidence absent? Did the prompt ask the model to promise something it could not verify? Would a source-preparation step prevent the failure, and how much time would that step add? Try one change on the same types of task and keep the first run for comparison.

Record errors separately from minutes. A fast answer with an invented certification fails even if it is easy to read. The captured two-draft review shows why a confident sentence still needs a source check. For a knowledge pilot, include a question the source cannot answer using the document-question practice set.

Download, repeat and review

Download the fictional pilot sample and totals to reproduce the arithmetic in a spreadsheet. The interactive worksheet can export your edited rows. Keep the acceptance rule, source version and date alongside that CSV; those details explain what the numbers mean.

Run a small second batch with routine, ambiguous and missing-source cases. Ask another reviewer to inspect at least the rejected outputs and borderline passes. Expand only when the result is useful for the actual task and someone owns ongoing checks. A positive time comparison alone does not establish legal suitability, data permission or reliable quality.

Report a reproducible calculation or documentation issue through our contact page. See how examples and corrections are handled, or return to the practical AI resource library.