AI ROI measurement before you sign an implementation contract

What counts as a return

Start with the decision the business is trying to make. Reducing paid overtime, accepting more work with the same team, and shortening customer response times are different objectives. Each needs its own evidence. A report showing that employees generated more drafts does not by itself establish any of them.

IBM's September 17, 2026 guidance on scaling AI recommends linking initiatives to business measures with a baseline, current result, and target. For a buyer, the practical implication is to agree on the measurement before implementation begins. IBM guidance on measuring business value

Use the table to distinguish what a metric tells you from what it leaves unresolved.

Quality can be a condition for accepting a project even when it is difficult to express in money. Keep that condition visible rather than assigning an unsupported monetary value to it.

How to build a comparable baseline

Define where the task starts and ends. For an invoice workflow, extracting fields is only part of the work if someone must still inspect, correct, approve, and enter the result. Measure the current process through the same end point you will use for the new one.

Record volume, task complexity, staff handling time, elapsed time, and errors over a period that reflects normal work. Note unusual demand, staffing changes, and seasonal effects. The appropriate observation period depends on how often the work occurs and how much it varies; one good demonstration is not a baseline.

Where practical, compare similar groups of work during the trial. If that is not possible, document what changed between the before and after periods. A new employee, cleaner source data, or a simplified approval policy may explain part of the improvement. The project can still be worthwhile, but the report should describe the combined intervention honestly.

Microsoft's September 17, 2026 account of its internal AI transformation distinguishes efficiency signals from wider business outcomes and notes that some benefits emerge later. That supports tracking leading indicators without treating them as already realized returns. Microsoft's account of measuring AI value

Which costs belong in the calculation

A practical AI ROI measurement plan includes implementation, internal preparation, operating charges, and the work needed to keep the system useful. Ask who prepares data, tests outputs, trains users, reviews exceptions, and maintains connections. Their time is part of the investment even when it does not appear on the provider's invoice.

Keep one-time and recurring costs separate. Use the same time horizon for both benefits and costs, and state whether the view measures cash expenditure or broader economic cost. Do not quietly use an economic value for employee time on one side and cash payments alone on the other.

For a simple, undiscounted assessment over a stated period, ROI is the monetized benefit minus the total cost, divided by that total cost. The formula is only as useful as its inputs. Benefits that cannot yet be valued defensibly should sit beside the calculation as operational results, not be forced into it.

Also test a less favorable case: lower usage, more human review, or higher running costs. A proposal that only works under its most optimistic assumptions needs a smaller trial or a clearer explanation of the uncertainty.

Why two promising projects produce different conclusions

Illustrative example: capacity without immediate cash savings. Assume a team handles 300 requests each month. Average staff handling time falls from 10 minutes to 6 minutes, including review and corrections. The hypothetical improvement releases 1,200 minutes, or 20 hours, per month.

If salaries and overtime spending stay unchanged, the calculation has not established a cash saving. The business could use the capacity to reduce a backlog or serve additional customers. It should then measure that result. If the spare minutes are fragmented and no additional work is completed, multiplying them by an hourly rate overstates the demonstrated benefit.

Illustrative example: an avoided recurring expense. Assume a different project demonstrably removes USD 600 per month of external processing charges. Also assume USD 1,800 in initial costs, including internal preparation, and USD 300 per month in total incremental operating costs. These are invented calculation inputs, not supplier prices or a forecast.

Over 12 months, the assumed benefit is USD 7,200 and the total cost is USD 5,400. Net benefit is USD 1,800, giving a simple ROI of about 33%. With costs and savings beginning as assumed, the initial cost is recovered after six months of USD 300 monthly net benefit. If charges are only reduced by USD 400 per month, the same 12-month case has a USD 600 net loss. The conclusion changes with the benefit actually realized.

What to agree with the implementation team

Ask five questions before accepting the business case:

1. What is the primary outcome? Identify the business measure, its owner, and what would count as a meaningful improvement.
2. What is the baseline? Specify the task boundary, observation period, task mix, and evidence the comparison will use.
3. What is included in cost? Separate initial work, recurring charges, internal effort, review, maintenance, and additional usage.
4. How will benefits be attributed? Record other process changes and prevent double counting, such as valuing the same freed time as both labor savings and extra output.
5. What decision follows the review? Define when the business will expand, revise, or stop the project based on measured results.

Keep the worksheet simple enough for the person running the workflow to maintain. A sophisticated forecast cannot compensate for missing records of corrections, adoption, or actual spending.

Make the business case testable

AI ROI measurement is most useful when it turns a broad promise into a decision the business can revisit. Separate realized financial benefit from capacity and service improvements, include the full cost, and retain the assumptions that connect them.

Before your next provider conversation, write down one target outcome and the evidence you already have. Ask the team to identify what must be measured during a trial to confirm or challenge the case. That gives you a reason to invest, and a way to know when the original reasoning no longer holds.