ROI

ROI stands for “Return on Investment,” roughly translated as “yield on investment.” The metric answers a simple question: What do I get back for the money I spent? To calculate it, you take the profit from an expenditure and divide it by the expenditure itself. The result is usually given as a percentage. An example: A company spends 100,000 euros on new software and thereby saves 130,000 euros in personnel costs. The profit is 30,000 euros, so the ROI is 30 percent.

The metric that AI projects fail on

In the years after 2022, companies poured enormous sums into artificial intelligence. Large language models—programs like ChatGPT that can write text and answer questions—were seen as a future technology. The trouble was: the benefit was often hard to quantify. That’s exactly why ROI shows up in almost every piece of news about AI. It’s the point where enthusiasm meets accounting.

A widely cited MIT study concluded in 2025 that around 95 percent of the AI pilot projects examined delivered no measurable return. Such figures are disputed, but they point to a real problem. A chatbot that gives impressive answers doesn’t automatically save costs. If employees have to check every output, the effort involved can even increase.

For investors, ROI is also an early warning signal. When corporations invest billions in data centers without corresponding revenue following, concerns about a bubble grow. The term therefore comes up often in connection with the quarterly earnings of Nvidia, Microsoft, or Meta.

Calculation, time frame, and the hidden costs

The basic formula is: ROI equals profit divided by investment, times 100. Profit here is revenue minus costs. A negative ROI simply means: less came back than was put in. The time frame always matters. 20 percent in one year is something entirely different from 20 percent over ten years.

In technology projects, the hardest task is plugging in the right numbers. The investment includes not just the license fees for an AI tool. There’s also training, integration with existing systems, computing costs for every single query, and the working hours spent on quality control. Anyone who only counts the subscription price ends up with a far too rosy figure.

The revenue side is just as tricky. Some benefits can be measured well, such as processing time saved per task. Others cannot: happier customers, better decisions, a more modern image. A common mistake is to generously convert such soft effects into euros until the result looks right. Serious calculations separate hard savings from estimates. Related but not identical is the payback period: it only tells you when the investment has been recouped, not how much is left over afterward.

From school projects to quarterly earnings calls

ROI originates from business economics and is over a hundred years old. You encounter it anywhere someone puts money into something and wants to know whether it paid off. That applies just as much to an advertising campaign as to buying a machine or installing solar panels on a roof.

In tech news, you hear the term mainly in two situations. First, at earnings calls, when analysts ask when spending on AI data centers will pay off. Second, in reports about companies shutting down AI projects again. The reason given is almost always: no discernible ROI.

Pay attention to how flexibly the number is used. Whoever states a figure has decided which costs to include and over what time frame to calculate. Both can be tilted in a convenient direction. ROI is therefore a useful metric, but not neutral proof. The interesting question is usually not how high it is, but how it was arrived at.