How to calculate AI workflow ROI without overlooking review costs

Measure the current process before estimating savings

Choose one eligible enquiry type and follow a representative sample through the existing process. Record active staff time for reading, gathering context, drafting, reviewing and updating records. Separately record waiting time, such as a draft sitting overnight for approval.

Use a consistent definition of completion. Comparing an AI-produced first draft with a human-approved final message overstates the benefit. Count all staff work needed to reach the same finished state in each approach. Include awkward cases, not just straightforward enquiries.

• Eligible monthly volume: cases within the proposed workflow scope.
• Baseline handling time: total active staff minutes per eligible case.
• New handling time: review, edits, manual steps and normal exception effort.
• Recurring costs: software, model usage, service fees and maintenance.
• One-time costs: configuration, data preparation, testing and training.

If the workflow covers only half of your enquiries, use that half in the calculation. Work outside scope still needs its existing process.

Calculate net released capacity

For each eligible case, subtract the new active handling time from the baseline. Multiply that difference by eligible monthly volume, then convert minutes to hours. Subtract additional recurring staff maintenance time that you have not already counted.

Illustrative example: 240 eligible enquiries arrive each month. The existing process takes 14 active minutes per enquiry. With the proposed workflow, review and editing take 4 minutes, and exception handling averages another 2 minutes across all eligible enquiries. The net release is 8 minutes per enquiry, or 32 hours per month.

The 2-minute average must include cases with no exception time. For example, if one in five enquiries needs 10 extra minutes, the average is 2 minutes per enquiry. Do not add that average again if your measured 6-minute handling time already includes it.

At an assumed loaded labour cost of USD 35 per hour, 32 hours represents USD 1,120 of monthly capacity value. Loaded cost should reflect a consistent treatment of salary and relevant employment costs. If different roles do the work, calculate their time and rates separately.

Subtract the full ongoing cost

Suppose the example workflow costs USD 450 per month for the combined service and software, and needs another 3 hours of internal maintenance at USD 35 per hour. That gives USD 555 in recurring monthly cost and USD 565 in net monthly capacity value.

The recurring total depends on the actual proposal and tool usage. Check whether model charges, connector subscriptions, support, monitoring and change requests are included. Add costs once; a managed fee that already includes a tool should not be counted again as a separate charge.

NIST’s AI Risk Management Framework treats evaluation and risk management as lifecycle activities. For a business case, that is a reason to budget for work after launch. The precise amount must come from your scope and observed operating effort, rather than a generic percentage.

Include setup in the period you are evaluating

Assume one-time setup in the example is USD 2,400 and you evaluate 12 months of steady-state operation. Annual gross capacity value is USD 13,440. Recurring costs total USD 6,660. Adding setup gives USD 9,060 in total first-year cost, leaving USD 4,380 in net first-year capacity value.

On these assumptions, net value divided by total cost is about 48%. Simple payback is approximately 4.2 steady-state months: USD 2,400 divided by USD 565. Neither figure accounts for a slow ramp, financing costs, tax or changes in volume. Add implementation and ramp time when estimating calendar payback.

These figures describe a capacity-based business case. They do not establish a cash return. If salaries and other spending stay the same, the released time has not automatically produced a bankable saving.

Keep cash savings and sales upside separate

Show a cash-saving scenario only when spending is genuinely avoided or reduced, such as a documented fall in paid overtime. If the benefit is time redirected to proposals, customer work or quality checks, state how that time will be used and who owns the change.

Additional sales need a separate evidence trail. Faster preparation may help, but demand, lead quality, offer and sales follow through also affect outcomes. Compare similar enquiries and use incremental contribution after the cost of delivering extra work, rather than counting total sales revenue as profit.

Do not count the same released hour twice: once as labour value and again as the full profit from work performed in that hour. Present alternative scenarios or explicitly remove the overlap.

Test how easily the business case breaks

In the example, 8 minutes released at USD 35 per hour is about USD 4.67 of capacity value per enquiry. With USD 555 in fixed monthly recurring cost, approximately 119 eligible enquiries are needed to cover that cost. To also recover USD 2,400 of setup over 12 months, the threshold rises to about 162 enquiries a month.

Those break-even thresholds assume the USD 555 cost stays fixed over the relevant volume range. If usage fees or review effort change with volume, recalculate using the actual cost structure.

Now reduce the time release from 8 minutes to 4 minutes. At 240 enquiries, gross capacity value falls to USD 560 a month. That leaves only USD 5 before recovering setup. A workflow can look promising in a demo and become uneconomic if reviewers repeatedly rewrite the output.

• Test lower volume than expected.
• Test higher review and exception time.
• Test higher recurring usage costs.
• Include a realistic ramp period before steady-state performance.

Use the pilot to replace assumptions with observations

Agree which measurements will decide whether to continue. Record eligible volume, total staff handling time, material errors, recurring costs and adoption. Review the results with the people doing the work. A short low-volume pilot can expose operational problems without proving a lasting sales improvement.

Being AI begins with diagnosis and a scoped pilot before deployment and ongoing management. Request a workflow diagnosis to identify the baseline, review burden and cost assumptions your business case needs. The aim is a decision you can defend, including a decision to stop when the evidence does not justify expansion.