# Why most AI projects fail before they begin — and why the use case matters more than the model > The FSE use case filter shows why the best use case matters more than the most capable model. - **Author:** Gunnar Hopfe - **Date:** 07 / 2026 - **Reading time:** 5 min - **URL:** https://fse-group.de/en/thinktank/warum-ki-projekte-scheitern-anwendungsfall-vor-modell The most expensive AI is not the most capable one, but the one nobody can tie to a concrete process. Why the use case decides success or failure of AI projects. ### When good advice comes at a price The most expensive AI is not the most capable model. It is the AI that gets used without anyone being able to say which process it is supposed to improve. That is exactly what we currently see in many companies. Licences are bought. New models tested. Agents built. Teams encouraged to experiment. Activity is high. The measurable benefit often stays modest. Not because the technology is immature, but because the central question is usually asked far too late: which concrete work is supposed to get better after the AI is introduced? The real bottleneck of successful AI projects today is no longer the technology. It is a company’s ability to define the right use cases. ### A pilot is not yet a use case Many companies speak of an AI pilot although in reality they are merely trying out a new technology. A pilot answers the question “what can the technology do?” A use case answers a different question: “which operational problem do we solve with it?” This difference decides whether a demonstration turns into commercial success. An AI pilot without a process metric stays a technology test. Only once it is clear which process is to be improved, which task simplified and which metric moved does a business case emerge. ### The biggest mistake: introducing technology before the process is understood In many organisations AI is anchored organisationally in IT. Technically that is right. Commercially it is not enough. IT provides security, integration and scalability. But only the business units can define the business benefit. That is where it is decided which tasks cost time today, where quality problems arise, which decisions need preparing and which work staff repeat every day. Hence: AI may be anchored technically in IT. Commercially it has to be anchored in the process. Responsibility is clearly divided. The business owns the value. IT owns keeping it under control. Only when both perspectives come together does lasting value emerge. ### Working faster does not automatically mean being more productive Generative AI can produce content in seconds. That does not automatically mean the whole process gets faster. On the contrary. If staff have to rework results, verify facts or rewrite answers, work is merely moved around. The authors Jeff Hancock and Kate Niederhoffer, together with BetterUp Labs, describe this phenomenon as “workslop”: AI produces apparently finished work that then has to be extensively checked or corrected. The local productivity gain of individual employees can thereby even slow down the overall process. The success of an AI application should therefore never be measured by the output it produces but by its effect on the whole process. The central question is not “how fast does the AI write a text?” but “how much faster does our company reach a reliable result?” ### Small use cases beat big visions Many companies start their AI initiatives with ambitious target pictures. Intelligent customer service. The autonomous case handler. The fully automated back office. The vision is right. The first step is often too big. The broader a use case is defined, the more systems have to be integrated, the more edge cases arise, the harder quality control becomes and the blurrier responsibilities get. Hence a simple rule: think big. Cut small. Connect systematically. Bounded use cases reduce complexity. They produce visible results faster. They create trust. And they deliver exactly the experience needed later for larger automation. With AI, drawing boundaries is not a lack of ambition. It is the precondition for scaling. ![The use case matrix: business value against technical complexity. Top left (high value, low complexity): productive, AI changes the process. Top right (high value, high complexity): showcase, impressive but without effect. Bottom left: automation without relevance. Bottom right: complex but little value, high effort, low benefit. Most companies invest on the right; winners invest top left, in the use case that changes the process](https://fse-group.de/media/think-pieces/20260715_thinktank-usecase-gewinnt.png) ### The FSE use case filter Before an AI project starts we answer six questions. If one of them cannot be answered, the use case is not ready for production. - Relevance: which commercial or operational problem is being solved? - Boundaries: where does the task begin and end? - Measurability: which metric improves? - Verifiability: can a human assess the result quickly and reliably? - Connectivity: can the output be integrated into the existing process without additional breaks in the chain? - Ownership: who owns the business benefit and the continuous development? This approach changes the discussion. The model is no longer at the centre; the work is. A good use case does not describe what an AI can do. It describes which work no longer has to be done once it is in place. ### From experiment to value Between a successful prompt and a successful company lie several stages of development. That is why we distinguish three formats. - Exploration: what can the technology do in principle? - Use case pilot: does it work in a clearly defined business process? - Production: does it deliver a measurable commercial benefit over time? Many companies skip this distinction. They confuse successful demonstrations with successful transformation. Yet McKinsey studies keep showing a similar picture: the number of AI initiatives grows far faster than the number of applications scaled into production. Business value arises above all where concrete functions and processes are improved, not where AI is treated as a general innovation project. ### Practice beats theory How big the difference between technology and use case can be is shown by the introduction of Kai Simone in SENEC’s customer service. The question at the centre was not which language model to use. At the centre was taking pressure off customer service. After only six weeks the first-time-fix rate was above 50 percent. At the same time, more than 45 percent of service cases are handled autonomously today. The success did not come from the largest possible AI rollout. It came from a clearly bounded use case with an unambiguous objective. That is exactly where commercial benefit begins. ### The competitive advantage is created before the first line of prompt The next level of maturity in using AI does not come from more capable models. Nor from more licences. It comes from better decisions about which tasks intelligent systems should take over permanently and which they should not. The companies that will prevail in the coming years will not be the ones with the largest AI budgets. They will be the ones that understand their processes best. Because in the end it is not the best AI that wins. It is the best-defined use case. Before you approve the next AI pilot, ask one simple question: which concrete step of work will measurably improve afterwards? If the answer stays general, the use case is not finished. #### Sources - BetterUp Labs & Stanford Social Media Lab (2025): Workslop: The Hidden Costs of AI-Generated Work. - McKinsey & Company (2025): Superagency in the Workplace: Empowering People to Unlock AI's Full Potential. - McKinsey Global Survey (2025): The State of AI in 2025. - FSE brand manual: positioning, market understanding and a vision of hybrid working worlds.