AI agent ROI: build a defensible business case

Define the baseline and the unit of value

Start with one workflow and a measurable completed task. Estimate how many tasks are eligible, how much human effort the current process requires and what quality it achieves. Separate eligible volume from total business volume; a feature that handles only some cases should not claim savings across every case.

Choose the value mechanism. The business may redeploy capacity, reduce purchased services, avoid additional hiring or improve throughput. Those mechanisms have different evidence and timing. A spreadsheet that multiplies minutes by salary does not establish that cash will leave the expense base.

A transparent workflow model

adopted tasks = eligible tasks × adoption
released hours = adopted tasks × (baseline minutes − remaining minutes) / 60
realized value = released hours × hourly value × capture fraction
net monthly value = realized value − monthly operating cost
simple payback = initial investment / positive net monthly value

Remaining minutes include review, correction and fallback effort averaged across adopted tasks. Monthly operating cost covers the additional system operation, maintenance and other costs not already counted in that remaining effort. Keep the accounting boundary explicit to avoid charging for the same review labor twice.

The capture fraction represents how much of positive released capacity becomes useful value. If the new workflow consumes more human time than the baseline, treat that added labor as a cost at the full stated hourly value; do not discount it as if it were unrealized savings. The calculator follows this conservative convention.

This model holds volume and cost constant. It does not include discounting, taxes or growth. Quality must meet the defined task requirements; a faster unacceptable result is not an equivalent completed task.

Worked example: capacity is only part of the benefit

Use synthetic assumptions: 10,000 eligible tasks per month, 70% adoption, six baseline minutes and 1.2 remaining minutes per adopted task. The workflow releases 7,000 × 4.8 / 60 = 560 hours. This is capacity, not a cash saving.

Assume only 50% of the released capacity becomes useful value: the equivalent of 280 hours at the organization’s measured hourly value. Subtract additional monthly operating cost from that value, then divide the initial investment by positive net monthly value to estimate simple payback. Monetary assumptions must come from the business.

Test the assumptions that can reverse the decision

The same technical feature produces very different economics when adoption or capture changes. This is why a pilot should measure the actual workflow and the operating change needed to realize value. Do not select only the optimistic case for a proposal.

Also test higher maintenance cost, longer review time, a narrower eligible population and lower accepted-outcome rates. If a modest change makes net value negative, the next phase should reduce uncertainty before expanding the investment.

Design the pilot around the uncertain assumptions

Measure baseline and proposed effort using comparable task classes. Include difficult cases and unsuccessful attempts. Track review and rework rather than timing only the model. Record adoption among eligible users and whether released time was actually redeployed.

• Fix the task definition and quality rubric before comparing effort.
• Measure a representative mix, including exceptions and fallback.
• Document the source of each cost input and which costs are excluded.
• Check whether the proposed operating change is feasible for the team.
• Set a decision date and the conditions for expanding, revising or stopping.

A positive time result is useful evidence, but business value can lag. Training, process changes and integration maintenance may be necessary to achieve adoption and capture. Include that work in the investment rather than treating it as free.

Use the model to stage the investment

A defensible funding decision states the workflow, assumptions, sensitivity range, unresolved risks and next evidence to collect. It can authorize a bounded pilot while withholding a larger rollout until the uncertain inputs are measured. That is often more useful than a single precise ROI percentage.

Recalculate after changes in task mix, model behavior, review requirements or operating cost. Preserve the original assumptions so the organization can learn why the result differed. The goal is a business decision that remains understandable after the initial excitement of the demo.