How to Measure AI ROI: A Practical Scorecard for Business Leaders

The common calculation multiplies minutes saved by salaries and calls the result ROI. That number assumes every saved minute becomes productive capacity, every output is usable and the system carries no review or operating cost. Those assumptions rarely survive contact with a real workflow.

OpenAI’s 2026 investment guidance recommends evaluating useful work per dollar and the full cost of reaching an acceptable result. McKinsey’s 2025 survey supplies the warning: 88% of respondents reported AI use in at least one function, but only 39% reported enterprise-level EBIT impact. Adoption and economic value are different measurements.

Measure the current process for two to four representative weeks. Record completed volume, queue time, hands-on time, first-pass acceptance, error severity, rework, escalation and the customer or employee outcome. Segment simple and difficult cases because AI may help one group and hurt the other.

Use system timestamps where possible. A survey that asks people how long work “usually takes” is useful for orientation, but weak as a financial baseline. Preserve the query or report used to calculate each number so finance and the process owner can reproduce it later.

A practical productivity formula is: annual value equals annual case volume multiplied by minutes saved per accepted case, loaded labour cost per minute and a realization rate. The realization rate reflects how much released capacity becomes useful work. Use a conservative rate when saved time arrives in scattered two-minute intervals or demand is too low to redeploy a role.

Revenue and risk require different logic. For sales, compare conversion or retained revenue against a valid control group. For risk, estimate expected loss using incident probability and impact, then document the uncertainty. Do not add time savings, revenue and avoided loss if they describe the same underlying benefit.

Model usage may be a small part of total cost. Add software licences, retrieval or database services, integration engineering, security review, evaluation design, data preparation, monitoring, support and employee training. Include the human review time that remains after launch and the expected cost of failures.

Separate one-time implementation cost from recurring cost. This lets leaders see payback period as well as steady-state margin. For model selection, compare cost per accepted outcome. A cheaper model that retries twice and sends more work to a reviewer can cost more than a stronger model.

The scorecard should pair financial results with leading indicators. Track task success, first-pass acceptance, correction rate, serious-error rate, human intervention, cycle time, cost per accepted outcome, user adoption and customer impact. Report averages and the worst important cases; an average can conceal a small group of expensive failures.

Adoption needs context. Low use can signal poor training, weak workflow fit or a broken integration. High use can signal value, novelty or mandatory policy. Pair usage with accepted outcomes and short interviews with the people doing the work.

Use a staged rollout, matched teams or alternating time periods when random assignment is impractical. Keep case mix, staffing and service levels visible. If a seasonal demand spike or policy change occurs during the pilot, note it rather than attributing the whole difference to AI.

Decide thresholds in advance. For example: scale when first-pass acceptance exceeds the baseline, serious errors stay below the agreed ceiling and payback remains under 12 months; revise when quality passes but adoption does not; stop when the workflow cannot meet the safety bar after two defined iterations.

Keep the main report to one page per workflow. Show the owner, stage, baseline, current result, quality threshold, monthly benefit, monthly recurring cost, cumulative implementation cost, payback forecast, top failure and next decision date. Put methodology and raw evaluation detail in an appendix.

Portfolio reporting should separate broad employee tools, function-specific workflows and strategic systems built around proprietary company data. Their economics and evidence standards differ. A writing assistant may be judged by adoption and time saved; an agent issuing refunds needs stronger outcome, error and control evidence.