Franco T.,
Too Long; Didn't Read
AI projects do not fail because of the technology. They fail because of the wrong question. "How do we implement AI?" leads nowhere. "Which business problem are we solving?" leads to real ROI. Most companies invest in tools, not in strategy. That is expensive.
The Scene That Keeps Repeating Itself
The CEO comes back from a conference. Or has read an article. Or has heard that the competitor is "doing something with AI".
"We urgently need AI."
Budget is approved. A team is assembled. Nine months later, there is a model lying around that predicts something. But nobody uses it.
Why? Because the prediction is not valuable to anyone. Or because the quality is not good enough. Or because there was no plan for how people were actually supposed to use the thing. The pressure not to fall behind is real. But it often leads to the wrong decision.
The Uncomfortable Truth
More than two thirds of major AI initiatives in companies do not deliver the expected business value. In our work with Swiss companies, we see this regularly.
Not because the models are bad. Not because the data scientists can't do it. But because companies start with the wrong question.
"How do we use AI?" That is the wrong question.
The right question: "Which specific business problem can we solve with it?"
That sounds trivial. But it isn't. The first question leads to bottom-up: you gather tech experts, and they look for problems for their solution. That's backwards.
The second question leads to top-down: you ask where it hurts, where costs arise, where customers are frustrated. And then you check whether AI can help.
The Five Reasons Why AI Projects Fail
1. Hype Instead of Strategy
The pressure not to "fall behind" leads to investments without a clear goal. "We need to implement generative AI" is not a strategy. That's panic.
2. Data Is Not What Companies Think
"We have so much data!" Yes. But having data in the system is not the same as having clean, structured data that can be used for machine learning.
Most companies massively underestimate the effort involved in data preparation. It is not the glamorous part. But it determines success or failure.
3. No Business Ownership
Many AI projects are driven by IT. That's a classic mistake. If the business owner is not behind it, if the person who is supposed to use the result is not actively involved, you build a solution that nobody wants.
4. Change Management as an Afterthought
Building a model is the easy part. Getting people to use it and adapt their processes is the hard part.
We have seen highly accurate prediction models that nobody uses. Because the users were not trained. Because integration into their workflows is painful. Because no one explained why it should be valuable.
5. ROI Is Not Defined
In the end, someone asks: "Was that successful?" And nobody knows. Because it was never defined beforehand what "successful" means. Was it time savings? Revenue growth? Error reduction? If that is not clear, you can't measure anything. And the next manager says: "The project was expensive and delivered nothing."
For a deeper look, see our article The AI Readiness Check: Is Your Company Ready for AI?.
What That Means in Practice
A service company from French-speaking Switzerland had invested three months in an automation project for invoice processing. The model was good. Accuracy above 90 percent. Still, after six months nobody used it anymore.
The reason: the accountant who should have been using the system every day had never been involved in development. The interface did not fit into her workday. The exceptions that came up daily could not be handled by the model. And nobody had defined how she should escalate errors.
The project was not a technology failure. It was an organizational failure. And it could have been noticed in week two if someone had asked the right person.
What AI Strategy Actually Means
AI Strategy is not "We implement AI." That's like saying "Our strategy is to use software." That says nothing.
AI strategy means:
A clear goal. Not "we want to use AI," but: "We reduce the average response time in customer service from 24 hours to 2 hours through automatic ticket classification."
Realistic expectations. Generative AI tools are powerful. But they hallucinate. They do not always understand nuanced business logic. They may not work on your specific data.
Honest readiness assessment. Do you have the data? Do you have the skills? Do you have the infrastructure? These are not theoretical questions; they are practical blockers.
Clear metrics. Before you start: what does success mean? How do you measure it? When do you know whether it worked?
Generative AI: The Reality Behind the Hype
Generative AI is not a miracle cure. It is a powerful tool for specific use cases.
Where it works well:
Summaries of long documents
Content generation with human review
Q&A based on knowledge bases
Code generation that is reviewed afterwards
Classification and extraction
Where it works poorly:
Financial or critical decisions where accuracy is crucial
Highly specialized domains where the model could hallucinate
Privacy-critical use cases when data protection is not ensured
What most people underestimate:
Hallucinations are real. Generative models can present invented facts as if they were true. That's not a small problem.
Data protection is complex. If you use ChatGPT directly and not the Enterprise API, your data can potentially be used for training. That's a GDPR problem.
Costs scale. "Generative AI is free" is a myth. At one million API calls per month, that quickly becomes five figures.
The Pragmatic Way
If you want to take AI seriously:
1. Start with the problem, not the technology
What are the three biggest pains in your company? Where do most costs arise? Where do customers lose time? Where are error rates high?
Only once you know that, ask: Can AI help here?
2. Assess your readiness honestly
Do you have the data? Is it clean? Do you have someone who can drive it? Is the business behind it?
If the answer to several of these questions is "No", start there. Not with the AI project.
3. Start small
Not "we transform everything with AI." But: a pilot. One problem. Clear metrics. If it works, scale. If not, learn.
4. Define success in advance
What does it mean if the project is successful? How do you measure that? Write it down before you begin.
5. Take governance seriously
The EU AI Act is not optional. If your model runs in sensitive areas, there are regulatory requirements. Ignoring them is expensive.
What You Can Do
List the three biggest business problems. Not IT problems, business problems. Where are you losing money? Where are customers frustrated? Where do errors occur?
Ask yourself for each one: Could AI help here? Be honest. Sometimes a simple rules-based system is better than machine learning.
Assess your data situation. If you wanted to train a model tomorrow, could you give the data to the data scientist? Is it clean? Structured?
Talk to the business, not just IT. Would someone from the business use the result? Is there real interest there?
The Point
AI is not a magical thing. It is a tool. A very powerful one, but a tool.
AI creates value when you solve the right problems. When you are realistic about opportunities and limitations. When your data is in order. When the business is behind it.
The companies that succeed with AI do not have the best technology. They have the clearest goals. They know what they want to achieve. They measure whether it works.
That is not sexy. That is not hype. But it works.
You want to approach AI strategically, not as a hype project? We help Swiss companies find the right problems and achieve real ROI. Talk to us.
