Most AI ROI claims compare a demo to a guess. A useful measurement compares the same workflow, on the same kind of work, before and after — and counts the review time and running costs the demo left out. These four steps give you a number you can defend, and a clear signal when the answer is “not worth it yet”.

Step 1 — Fix the unit of work

Goal: Decide exactly what one “run” of the workflow is.

  1. Name the unit: one invoice, one lead, one ticket, one report.
  2. Write down where a run starts and where it ends.
  3. Exclude work the AI will not touch, so it does not inflate either side.
  4. Agree the measure with the person who owns the workflow.

You finish with: A one-line definition of a run and its start and end points.

Step 2 — Record the baseline

Goal: Measure the manual process before anything changes.

  1. Time a representative sample of runs, including waiting and rework.
  2. Count errors and anything sent back for correction.
  3. Note the people involved and what an hour of their time costs you.
  4. Keep the sample and the notes — the comparison depends on them.

You finish with: Time, error rate and cost per run for the manual process.

Step 3 — Measure the full cost after

Goal: Count everything the new workflow takes, not just the automated part.

  1. Time the same kind of runs with the workflow in place.
  2. Include human review, exception handling and corrections.
  3. Add running costs: seats, API usage, hosting and maintenance time.
  4. Spread the one-time build cost over a period you would accept for payback.

You finish with: Time, error rate and total cost per run for the new workflow.

Step 4 — Decide and report

Goal: Turn the comparison into a decision someone can act on.

  1. Compare cost per run and error rate side by side, on the same sample size.
  2. Multiply the saving by the real volume, not the hoped-for volume.
  3. State what the number does not include, such as speed or customer experience.
  4. Decide: expand, adjust, or stop — and write down why.

You finish with: A one-page result with the method, the number and the decision.

Common pitfalls

  • Comparing against a vendor’s average instead of your own baseline.
  • Leaving review and exception time out of the “after” measure.
  • Counting hours saved that nobody can reuse.
  • Reporting a single good week as the result.

More guides like this

Related answers

Read as Markdown Automate the repeatable