Blog · AI for small and midsize businesses
How to calculate AI project ROI without overstating savings
Time released by AI is not automatically cash saved. A credible return calculation includes review and correction, distinguishes usable capacity from reduced spending and tests assumptions against more than one workload scenario. This example shows the arithmetic and the questions behind it.

In this article
An assistant that drafts a task in seconds can still require several minutes of checking. Multiplying its generation speed by every employee's salary will overstate the benefit. Calculate return on investment from the completed workflow and distinguish financial savings from time that becomes available for other work.
Start with net time released
Use this planning formula: monthly cases multiplied by the difference between current handling time and assisted handling time, divided by sixty. Assisted time includes preparation, reading, checking, correction and the manual work required when the tool fails. Record recurring maintenance separately so it is also counted.
In a hypothetical administrative task, assume 600 monthly cases take eight minutes each today. Assisted preparation takes two minutes and review takes three. The apparent net release is 600 × (8 − 5) ÷ 60 = 30 hours. If some cases need full manual rework, include that effort before accepting the figure.
Distinguish capacity from cash
Suppose the company uses an internal planning rate of MXN 250 per hour. Thirty hours represent MXN 7,500 of theoretical capacity value. No salary expenditure disappears simply because those hours are released. The owner must identify what happens to the capacity: reduced overtime, avoided temporary staffing, smaller backlog or additional productive work.
If the illustrative company expects only 60% of that capacity to be used productively, its planning benefit is MXN 4,500 per month. This realization factor is an assumption to verify. Do not combine it with separate revenue claims for the same hours or count the benefit twice.
Include the cost of obtaining the benefit
Assume an illustrative one-time implementation cost of MXN 30,000 and monthly operating costs of MXN 2,500, including support and source maintenance. Over twelve months, the modeled benefit is MXN 54,000; total cost is MXN 60,000. Using (benefit − cost) ÷ cost, the first-year planning return is −10%. These figures are invented arithmetic inputs, not market prices or expected results.
| Assisted minutes per case | Monthly hours released | Realized monthly capacity value |
|---|---|---|
| 4 | 40 | MXN 6,000 |
| 5 | 30 | MXN 4,500 |
| 7 | 10 | MXN 1,500 |
Review time matters more than a quick drafting demonstration. Check lower volume, higher support demand and slower adoption as additional scenarios. Keep nonfinancial outcomes, such as more consistent answers, visible but separate unless the business has a defensible valuation method.
Assign ownership of the assumptions
The process lead measures task time and quality. Finance confirms the planning rate, cost categories and which benefits are actual spending changes. The technical partner estimates usage and maintenance. The sponsor decides whether capacity improvements justify a project even when the first-year model is negative.
Deloitte's use-case framework includes implementation and maintenance economics. Use that perspective without importing enterprise savings into your SME forecast. The NIST profile adds the need for continuing evaluation as conditions change.
Bring measured cases and your own cost assumptions to an AI project assessment to build a calculation the decision makers can inspect.