The ROI from investments in AI projects is seemingly calculated the same way as the return on any other investment. Seemingly, because if it were that simple, reports, data, and analyses would speak of successful implementations, not failures. What makes this so deceptive? AI technology is new to many organizations, and no one has previously evaluated its effects. The lack of experience in calculating ROI and the absence of typical reference points mean that many companies don't even start calculating when an investment will pay off.
- What is AI ROI and why is it so hard to calculate?
- Why do most AI implementations not pay off?
- How to calculate the return on investment in AI solutions?
- When does AI truly pay off?
- How to increase the chances of return: pilot and stages
- Summary
What is AI ROI and why is it so hard to calculate?
The formula and method for calculating AI ROI are typical: we compare costs and benefits, add the investment value, the cost of money, and it's done. For a financial controller or CFO, this should be a simple task. It should be, but for some reason, it isn't.
Firstly, because while it's easy to calculate the cost of a license, in the case of more complex projects requiring a pilot phase and burdened with the risk of failure and uncertainty regarding outcomes, there are many variables. The second issue is benefits. The goal of implementing AI-based tools might be, for example, to increase analysis precision or accelerate decision-making. Of course, everything can be calculated somehow. But... that's the point. Somehow.
Calculating AI ROI, just like any other project, only makes sense if the result is realistic and achievable. ROI calculated on "from the sky" assumptions, meaning "let's assume a better decision means 100,000 annually," makes no sense.
Why do most AI implementations not pay off?
Most AI implementations do not pay off not because the technology is bad or the CFO miscalculated the ROI, but because companies start with the tool, not the problem. They buy AI because "everyone is implementing it," and then they look for what they could use it for. This is not an investment. This is testing, and testing is for checking, not for paying off.
The scale of this phenomenon is impressive. In 2025, MIT, in its report "The GenAI Divide: State of AI in Business 2025," found that approximately 95% of generative AI pilot programs did not yield measurable returns for companies. Another statistic: according to S&P Global Market Intelligence (report from March 14, 2025), 42% of companies abandoned most of their AI implementation initiatives. These, in my opinion, are the two main reasons why organizations do not see a return: the large scale of solution testing and the uncertainty of assumptions.
How to calculate the return on investment in AI solutions?
Theoretically, the method for calculating ROI is the same: we compare benefits and costs. The math is simple. The difficulty lies in the assumptions. Benefits from change can mean increased productivity, increased revenue, but also better decisions. And while productivity and revenue are easy to calculate, how do you value the assumption that AI will make company decisions faster or more accurate?
On the cost side, we have obvious expenses for licenses, hardware, project teams, and employee training. These are "hard" expenses, easy to calculate. Additionally, there's the probability of pilot success or failure, risks associated with AI, cybersecurity, and so on. As you can see, the added value from AI is not just easily quantifiable cost reduction. Many variables are involved, for which calculating the probability of occurrence is difficult or even impossible. There are no benchmarks or histories of similar projects, and in most cases, there aren't even good practices, because the entire concept of artificial intelligence on such a scale is new.
When does AI truly pay off?
For an AI investment to pay off, in addition to positive figures (because, as we know, Excel will accept everything), four criteria must be met.
-
The AI-based solution must be closely linked to an existing business problem. Change management is about making things better: a more satisfied customer, higher sales, lower costs.
-
Another condition is data. As we know, data is often a greater company asset than fixed assets or even loyal employees. Starbucks has been collecting data for years: order history, purchase times, and customer preferences. Based on this, they built a solution that personalizes offers and helps manage logistics and employee schedules. Thanks to this, revenues grow, and costs decrease. Cash is king, but data is queen.
-
Do you know the saying: using a cannon to kill an ant? It's about scale, which is very important for AI projects. AI is expensive technology. I'm not talking about buying a GPT subscription, but about comprehensive, well-thought-out solutions. Since it's expensive, it will only pay off if it solves a large-scale problem.
-
And the last point is repeatability. Let's remember that AI is about probability. AI is not your advisor or coach. It's an algorithm that, with a certain probability, selects the most expected "solution." If you use AI daily, you probably initially felt that Claude could read your mind. But then frustration comes, because you want a constructive answer, and you get the same "gibberish." This happens because AI excels in repetitive processes where the probability of a correct answer is high.
How to increase the chances of return: pilot and stages
The chance of return on investment lies in good change management. Simply calculating the rate of return and monitoring costs is only part of the success. Since AI projects are burdened with uncertainty resulting from technology, organizational culture, and legal changes that may affect the implementation, it is reasonable to prepare well and divide the project into stages.
Where to start? Besides preparing the organization, i.e., ensuring appropriate maturity, the next step is to select use cases, which means identifying areas for change. Use case selection criteria is a set of criteria based on which the priority of changes is established. These include:
-
how important the business problem is that AI will help solve
-
the scale and repeatability of the process
-
measurability of benefits (ROI)
-
risks
-
availability of resources (including competencies)
-
availability of technology
The criteria may vary depending on the project's complexity and may have different weights for different industries or products. It is important that they are defined and consistent with the company's priorities.
Summary
AI ROI doesn't come from the implementation itself, but from AI solving a specific, repeatable problem for which there is appropriate data and potential for scaling. The rest is discipline: calculating costs, valuing benefits – hard ones directly, and soft ones cautiously, choosing the right use case for a pilot. Most implementations don't pay off not because AI doesn't work, but because they start with the tool, not the problem and the calculation. In short: AI pays off when you treat it as an investment, not a fad.
If you want to calculate whether a specific AI implementation will pay off, or organize it before you spend your first zloty, check out the CFO for hire service.
