To calculate the ROI of an AI project, subtract the annual run cost from the annual benefit (safest to measure: hours saved × loaded cost/hour), then divide by the total investment. Add the payback period. Set the baseline number before the project, or you cannot prove the gain.
- AI-project ROI = (annual benefit − annual run cost) ÷ total project cost, expressed as a percentage.
- The safest benefit to quantify is time saved on a repetitive task, multiplied by the loaded hourly cost.
- Also calculate the payback period: how many months it takes for the net benefit to cover the upfront investment.
- Measuring the benefit is reported as a major barrier to AI adoption — so you set the baseline number before you start, not after.
What does the ROI of an AI project actually mean?
The ROI (return on investment) of an AI project is the ratio between the net gain the solution produces and what it cost you to build and keep running. Unlike a classic software project, an AI project has two kinds of cost: the upfront investment (design, build, deployment) and a recurring run cost (model calls, infrastructure, maintenance, human validation). The benefit, in turn, rarely comes from one place: time saved, errors avoided, extra capacity without new hires. To get a number you can trust, you isolate one benefit you can measure directly and set the rest aside. At Sapio, in early conversations we almost always start from time saved on a repetitive task, because it is the only benefit you can verify with a stopwatch and a payroll line, not with an optimistic estimate.
What is the simple ROI formula for an AI project?
Use a formula with three inputs you can fill in yourself without a complicated financial model. First you compute the gross annual benefit, then subtract the annual run cost to get the net benefit, then divide it by the total investment. The three inputs are: hours saved per month, the loaded hourly cost (salary plus taxes and overhead), and the run cost of the solution.
- Gross annual benefit = hours saved/month × loaded cost/hour × 12.
- Net annual benefit = gross annual benefit − annual run cost (model, infrastructure, maintenance).
- ROI (year 1) = (net annual benefit − upfront investment) ÷ upfront investment × 100.
- Payback period (months) = upfront investment ÷ (net annual benefit ÷ 12).
What does a worked ROI calculation look like?
The numbers below are hypothetical, chosen so you can see the mechanics of the formula — they are not the results of a Sapio project. Assume a team that manually processes repetitive documents and an AI assistant that takes over extraction and first-pass triage, with human validation at the end.
| Input | Value (example) | Result |
|---|---|---|
| Hours saved / month | 120 hours | — |
| Loaded cost / hour | RON 90 | — |
| Gross annual benefit | 120 × 90 × 12 | RON 129,600 |
| Annual run cost | model + infrastructure + maintenance | RON 24,000 |
| Net annual benefit | 129,600 − 24,000 | RON 105,600 |
| Upfront investment | design + build + deployment | RON 150,000 |
| Year-1 ROI | (105,600 − 150,000) ÷ 150,000 | −30% |
| Payback period | 150,000 ÷ (105,600 ÷ 12) | ~17 months |
The lesson from the example: in year 1 the ROI is negative because the upfront investment is not yet amortised, but the payback period is under a year and a half, and from year 2 the ~RON 105,600 net benefit flows without the upfront cost. That is why an AI project is evaluated over 2–3 years, not over 12 months. If you only look at the first year, you reject projects that pay for themselves quickly afterwards.
Why do AI ROI calculations most often fail?
The most common mistake is not in the formula, it is the missing baseline. If you do not know how long the task takes today and what it costs today, you have no way to prove the AI improved anything. The business press (business24.ro, citing Horváth-area data) reports that measuring the benefit is one of the main barriers to AI adoption in companies: firms cannot quantify what they gained, so projects stall in the "seems useful" phase. The fix is simple and applies before any line of code: measure the current state (time, volume, error rate) for two or three weeks, fix that as the baseline, then compare against it after launch. The other traps are underestimating the run cost (model calls and human validation are not free) and counting benefits you cannot measure, like "better decisions", instead of hours and errors.
The upfront investment depends directly on the complexity of the solution, and that is a separate conversation: we broke down what actually moves the price in our article on how much a custom AI solution costs. Read it before you fill in the "upfront investment" number in the formula, so you do not start from too optimistic an estimate.
How do you move from the formula to an investment decision?
Fill in the formula with your real numbers, then ask it three questions: is the payback period under 18–24 months? Does the benefit hold up after the initial enthusiasm fades? Does the run cost stay stable if volume grows? If the answer to all three is yes, you have a solid business case. If you are unsure about any of the numbers, the cheapest way to find out is a small pilot, with one process and one metric, before a large project. If you want to check your numbers together, book a free discovery call with the Sapio team; in that call we work out what is measurable in your case and whether a pilot makes sense before a larger investment.
Measuring the benefit is reported as one of the main barriers to AI adoption among companies in Romania — Horváth-area data, cited by business24.ro.
Frequently asked questions
What is the ROI formula for an AI project?
Year-1 ROI = (net annual benefit − upfront investment) ÷ upfront investment × 100, where net benefit = hours saved/month × loaded cost/hour × 12, minus the annual run cost. Compute the payback period separately: upfront investment ÷ (net annual benefit ÷ 12). Three inputs you can fill in yourself.
Over what period should I evaluate an AI solution's ROI?
Over 2–3 years, not 12 months. In year 1 the upfront investment weighs on the number and ROI can be negative even for a good project. From year 2 the net benefit flows without the upfront cost. If you only look at the first year, you reject projects that pay for themselves quickly afterwards.
Why is it hard to measure an AI project's benefit?
Because most firms do not fix a baseline beforehand. The business press (business24.ro) reports measuring the benefit as a major barrier to AI adoption. The fix: measure current time, volume and error rate for two or three weeks, set them as the baseline, then compare against them after launch.
What run cost should I include in the ROI calculation?
All recurring costs: model calls, infrastructure, maintenance, and human validation where it stays necessary. These are the most commonly underestimated, because they do not show up in the upfront investment. Omit them and your calculated ROI will be more optimistic than reality, especially as volume grows.
Is ROI tied to how much the AI solution costs?
Yes, directly — the upfront investment is the denominator of the formula. Cost depends on complexity: data preparation, model approach, number of integrations, and the accuracy bar. We broke down what moves the price in our article on how much a custom AI solution costs; read it before filling in the upfront investment figure.
Want to discuss a project?
Book a free discovery call with the Sapio team.