by Jeanine Schoonemann

In our overview article on Data & AI trends for 2026 we described how enthusiasm about generative AI is undiminished, but the tone has fundamentally changed. The phase of "seeing what can be done" is giving way to the question: what will it deliver, at what risks, and can we scale it up responsibly? The AI hype has given way to a mature reality where results matter, not promise.

That reality confronts many organizations with an uncomfortable gap. Demos are spectacular, success stories inspiring, the technology seemingly ripe. Yet concrete results often fail to materialize. Pilots that don't scale. Investments that don't pay off. Systems that do not perform adequately. And a growing awareness that what works for others cannot simply be copied. Because what to the consumer feels like a simple chat interface - ChatGPT - is in reality a complex AI ecosystem: not a separate model but a complete software package with countless layers, dependencies and preconditions.

In this article, we dive deeper into this reality gap. Through sharp statements, we take you through critical insights that make the difference between AI as an expensive exercise and AI as a valuable investment. With concrete real-life examples and clear frameworks, we'll show you how to get from inspiration to implementation. Because value is not created by copying what impresses, but by making sharp choices about what works in your situation.

Thesis 1: Those who rely on slick demos and successes of others when it comes to generative AI mainly build castles in the air rather than real value.

Management sits in the boardroom watching a demo. A generative AI solution analyzes complex documents within seconds, writes clear summaries, and answers questions as if an expert were at the table. It looks effortless. Colleagues nod enthusiastically. Someone mentions the competitor who is already doing it, too. The conclusion seems obvious: "We should have this, too."

Six months later, the reality is more unruly. The pilot is running, but performing below expectations. The system that seemed so intelligent regularly gives incomplete or incorrect answers. Employees start going through the results completely instead of trusting them. The efficiency gains evaporate into audit time and corrections.

The demo effect

That's the demo effect. Demos show what technology can do under ideal conditions. Cured data, predefined scenarios, no noise. But your reality is not a demo. Your data is chaotic, your systems do not always work happily together, your processes contain exceptions, your employees do not work according to the ideal script.

Perhaps more importantly, a solution that works amazingly well in broad and general context is far from reliable enough in your business (critical) process. An error rate of 5% sounds acceptable , until you realize that at 1,000 transactions per day this means 50 errors that must be corrected manually. Between amazingly good and reliable enough lies a world, and months of work and investment, of difference.

The copy trap

You may be thinking now, but I didn't start from a slick demo. I was smart, I looked at what already works at another organization and copied that. There's nothing wrong with that, is there? This is the copying trap: someone else's success is not your blueprint.

The temptation is understandable. You read that a retailer increased conversions with AI-driven personalization. A bank automated credit reviews and accelerated processing times. An insurer deployed chatbots and reduced service costs. The use cases sound clear, the results measurable, the path laid out. So you copy the approach. Same use case, similar technology, similar ambitions. But what you don't copy - because it's not in the success story - is the context in which it worked.

That retailer had invested five years in data platforms and had clean, structured customer data. That bank was working in a U.S. compliance context with different requirements for explainability. That insurer started with repetitive, low-risk queries where errors did not have a big impact.

Your context is different. Your data is scattered across systems. Your compliance requirements are more stringent. Your processes are more complex. And sometimes, this is where the geographical dimension comes in, the technology they were using is simply not yet available in Europe, or not compliant with GDPR or the AI Act. You also have to deal with cultural acceptance differences. What works there doesn't automatically work here. Not because the technology is flawed, but because the context is not right.

Critical thinking and foundation

Successful deployment of AI is not accomplished with a demo or an inspiring use case. That is inspiration, shows the potential and helps you think about solutions. Successfully deploying AI starts with a concrete problem, a thorough analysis of the process and the pain points. An important step that is sometimes skipped these days: considering a non-AI solution. AI comes into the picture only where it can add value. And when AI comes into the picture, questions such as:

• Do we have the right data and is it of sufficient quality?
• What technology do we have at our disposal and what opportunities and frameworks does it give us?
• If we implement AI in this process step, are we still meeting (our own) compliance requirements?
• What margin of error is realistic and is it acceptable?

These are not easy questions to answer, some answers require a robust conversation with the legal department, some answers may even be found only by doing a pilot. Not easy, but certainly indispensable for a successful project. By having the answers to these questions clear, you are building on a foundation rather than in the air.

Thesis 2: AI assistants will be widely deployable in 2026, but full process automation requires robustness, compliance and explainability that (generative) AI does not provide by default.

By 2026, AI assistants will be everywhere. Developers write code with it, analysts use it for summaries, marketers have ideas and inspiration generated. The value is tangible, the adoption wide, the risks low. But as soon as organizations want to move to full process automation, AI that does not assist but decides autonomously, they encounter a harsh reality: generative AI, in contrast to "regular" AI, does not offer the safeguards needed to do so.

A hospital wants to use AI for diagnostic suggestions. A bank is considering automated credit assessments. An insurer is considering AI-driven claims processing. The efficiency gains seem obvious. Until you ask the critical questions:

Is it reproducible? In credit assessment, the same application should lead to the same outcome. Generative AI repeatedly gives different answers. This is creative in writing, but unacceptable in loan decisions.

Can you explain? When a patient asks why this diagnosis was suggested, you have to be able to say more than "the model thought this." You have to be able to state what symptoms, what medical history, what factors weighed in. Generative AI is a black box - you don't know exactly why it gave this output.

Does it contain bias? AI trained on historical data reproduces historical patterns. If women were less likely to get loans in the past, the model learns that. Not because women are less creditworthy, but because the pattern is in the data. How do you check for that? How do you prevent it?

Is it auditable? In the event of a wrong claim settlement, you need to be able to reconstruct in retrospect how that decision came about. For compliance, for legal accountability, for error recovery. With generative AI, this is complex and requires extra attention (and time) in the design phase.

Note: These are all points that we are working with rule-based systems and interpretable predictive models (i.e. "traditonal" AI) have already found a solution to this, the deployment of such more traditional AI will therefore often be the best choice in these types of cases as well.

In many industries, explainability is not a nice-to-have but a legal requirement. The AVG gives people the right to receive explanations for automated decisions. The AI Act has explicit requirements for transparency and bias control in high-risk uses. Generative AI does not meet these. Not because the technology is bad, but because it is designed for other purposes: flexibility and creativity, not consistency and accountability.

Fortunately, this does not mean that generative AI cannot necessarily be deployed, but that it does not involve complete automation, there will always remain a "human-in-the-loop." Thus, a significant gain in time can still be achieved, without handing over the real decisions to generative AI. Think of summaries of (medical) letters that are still edited by a content expert. Supporting a decision maker by referring directly to relevant files and chapters based on the query and background information. Providing documents with relevant tags created by generative AI that makes it easier for analysts to surface the right information. In addition, a solution with generative AI can also consist of a combination with other techniques, so that as a whole it does meet our requirements. We need to use generative AI where it is so powerful and humans or other techniques where they are better.

Thesis 3: Black-and-white thinking about AI leads to paralysis; successful organizations embrace the shades of gray and actively steer the ethical debate.

When organizations think about implementing AI and the ethical considerations involved, a split often arises. On the one hand, there are legitimate concerns: what if the system discriminates? What if we can't explain why a decision was made? Who is liable if things go wrong? These questions are important, but often lead to paralysis.

Organizations decide that it is all too risky, that there are too many unknown factors. They put off AI projects until all the ethical dilemmas are resolved, until the laws and regulations are completely clear, until they have absolute certainty that nothing can go wrong. It sounds like a prudent, responsible approach. But in reality, it means that nothing happens for years. And all the while, the expectation that these techniques will leave the field again seems as unrealistic as the hope that the ethical dilemmas will resolve themselves. Because that absolute clarity will not come - ethical issues, by definition, remain complex and context-dependent.

There are also organizations, which in a kind of counter-reaction to paralysis, are too little concerned with ethics and focus very strongly on innovation just to keep from falling behind. The objections are dismissed with: that will come later, it should not slow us down now.

Between these extremes operate the organizations that do succeed. They accept that perfect systems do not exist and that ethical questions remain complex. But they do not use that as an excuse to stand still or run on without safeguards. Instead, they consciously choose transparent models, conduct regular checks for discrimination, involve stakeholders in ethical considerations, start pilots, and build in the ability to learn and adjust.

This approach requires that ethics not be a final check, but a design principle. With any AI project: do you check fairness? Can you explain how it works? Who bears responsibility? Can you adjust? Ethical complexity is not a reason for inaction or blind action. It is a reason for careful movement. You can read more about ethics and AI here.

From experimentation to targeted and responsible deployment of generative AI

Generative AI offers enormous potential. The demos don't lie, the success stories are real, the technology is impressive. But between that promise and lasting value lies a reality gap that can only be bridged with conscious choices.

Choices to start with your own problem, not someone else's solution. To be inspired by what is possible, but remain critical of what is feasible in your context.

Choices to understand the difference between generative AI as assistant or autonomous decision maker. To build safeguards where they are needed.

Choices to use ethical complexity not as an excuse for inaction, but as a reason for diligence. To experiment with transparent models or processes, with pilots that provide space for learning and adjustment.

Value is not created by the smartest model or the latest technology. It comes from the best-designed chain of human, data, model and process, with sharp choices about where AI adds value and what safeguards are needed.

This article focused on one of the three developments from our editorial: from experimentation to targeted and responsible deployment of generative AI. In other articles, we explore strategic choices around data sovereignty and cloud costs, and the organizational maturity required to embed Data & AI sustainably.

Curious about going deeper or what (generative) AI can do for you? Follow our series by signing up below or contact us.