AI ROI Calculation: A Step-by-Step Worked Example

Calculating the ROI of an AI use case means comparing the benefits measured against a baseline with all the costs of the use case, using the formula: ROI = (benefits − costs) ÷ costs. Payback time shows after how many months the initial investment is recovered by the monthly net benefit.

Below is a complete example on a bounded case. All numbers are hypothetical and for illustration only: they show the method, not expected results. For the general rules of measurement, see how to measure AI ROI.

What is the use case in the example?

Handling recurring internal requests (for example questions about leave, expenses, procedures) by a team of 8 people. The case is suitable because it has repeated volumes, a start and an end for each request and an owner. Claude or another assistant helps draft the reply from an internal knowledge base; the person reviews and sends.

A fully loaded hourly staff cost of €35 is assumed.

Before starting, two methodological choices. Work over a defined period (12 and 24 months) and always separate one-off costs from recurring ones, because only the latter repeat every year. In addition, value only what was measured in the baseline, leaving out any benefit nobody could verify.

Step 1: what is the baseline before AI?

After three months of measurement, the monthly averages are:

• 1,000 requests received;
• 12 minutes of average work per request;
• 10% of requests reopened (the requester did not get the right answer), meaning 100 requests, each reworked in 12 minutes.

Monthly work is: 1,000 × 12 = 12,000 minutes, plus 100 × 12 = 1,200 minutes, for a total of 13,200 minutes, equal to 13,200 ÷ 60 = 220 hours a month.

Step 2: what is the situation after?

After the ramp-up, measured at steady state (excluding the first weeks):

• the same 1,000 requests;
• 9 minutes per request, including the time to review the proposed reply;
• 6% of requests reopened, meaning 60, each taking 9 minutes.

Monthly work is: 1,000 × 9 = 9,000 minutes, plus 60 × 9 = 540 minutes, for a total of 9,540 minutes, equal to 9,540 ÷ 60 = 159 hours a month.

Step 3: which net benefit can be valued?

The hours freed are 220 − 159 = 61 hours a month, or 61 × 12 = 732 hours a year.

Converting all of them into euros would be optimistic, because not all free time becomes useful work. Prudently, assume that 50% is reused for other team activities: 732 × 0.5 = 366 hours. Valued at €35: 366 × 35 = €12,810 a year (€1,067.50 a month).

This benefit already includes the quality effect: reopened requests fell from 10% to 6% and their time is within the 159 hours. Unmeasured benefits, such as greater requester satisfaction, are not counted.

Step 4: what does the use case cost in full?

Hypothetical values, to be replaced with those from your own contracts.

• Licenses: 8 users × €25 a month × 12 = €2,400 a year.
• Knowledge base preparation and configuration: €6,000 one-off.
• Training: 8 people × 3 hours × €35 = €840 one-off.
• Governance and maintenance of the knowledge base: 4 hours a month × €35 × 12 = €1,680 a year.

First-year cost: €2,400 + €6,000 + €840 + €1,680 = €10,920. Recurring cost from later years: €2,400 + €1,680 = €4,080. One-off costs: €6,000 + €840 = €6,840.

For how to structure the items, see the hidden costs in enterprise AI TCO.

Step 5: how do you apply the formula and payback time?

First-year ROI

• Net benefit: €12,810 − €10,920 = €1,890.
• ROI = €1,890 ÷ €10,920 = 17.3%.

Second-year ROI (without one-off costs)

• Net benefit: €12,810 − €4,080 = €8,730.
• ROI = €8,730 ÷ €4,080 = 214%.

Cumulative ROI at 24 months

• Benefits: €12,810 × 2 = €25,620. Costs: €10,920 + €4,080 = €15,000.
• ROI = (€25,620 − €15,000) ÷ €15,000 = €10,620 ÷ €15,000 = 70.8%.

Payback time

• Monthly benefit: €12,810 ÷ 12 = €1,067.50. Monthly recurring cost: €4,080 ÷ 12 = €340.
• Monthly net benefit: €1,067.50 − €340 = €727.50.
• Payback = €6,840 ÷ €727.50 = 9.4 months, or about 10 months.

Cross-check: in 9 months you recover 9 × €727.50 = €6,547.50, which leaves €292.50 uncovered, equal to 0.4 months. The numbers add up.

What happens if the assumptions change?

ROI depends heavily on the share of time reused. With 25% instead of 50%: 732 × 0.25 = 183 hours × €35 = €6,405 a year. The first year would give €6,405 − €10,920 = −€4,515, which is an ROI of −41.3% (4,515 ÷ 10,920). The use case would not pay back in the first year.

The comparison shows why the result should be presented as a range and with stated assumptions. To set scenarios and stop thresholds in advance, see how to build an AI business case.

Payback time deserves careful reading: payback in 10 months over a 24-month horizon is acceptable for many organizations, but less so if the vendor or contract provides for an annual renewal with price review. That is why payback should always be compared with the length of the contractual commitment, and not only with the project calendar.

Which steps should you repeat on a real case?

1. Choose a bounded process with repeated volumes.
2. Measure the baseline: volume, average time, reopenings or errors, over a sufficient period.
3. Introduce the tool with training and rules of use.
4. Measure again at steady state, including review time.
5. Convert the freed hours into value with an explicit reuse share.
6. Add up all costs, separating one-off and recurring.
7. Calculate ROI and payback, then repeat with worse assumptions.

What are the limits of the example?

• Invented numbers. None of the values comes from real measurements or vendor prices.
• Benefit at steady state from month one. In a real case the ramp-up is gradual and the first year is less favorable.
• Causality. Part of the improvement may depend on other factors, such as changes in volumes or headcount. A comparison group would make the estimate more credible.
• Reuse share. The 50% is an assumption, not a measurement.
• Uniform hourly cost. Different roles have different costs.
• No value for qualitative benefits. Requester satisfaction and the team's perceived workload stay outside the calculation and should be read with separate indicators, as in enterprise AI KPIs.
• A single use case. The result does not transfer automatically to other processes.

Where Delphis fits

Delphis is a platform, independent of AI vendors, that connects spend, adoption, impact and regulatory obligations. It reconciles costs with actual usage, simulates the possible saving and compares expected and realized impact on use cases, with before-and-after baselines on bounded processes. See the cost page.