‘It's not the AI's fault’ – the inconvenient truth about AI implementation

This is not an incident, it is the norm. The GenAI Divide: State of AI in Business 2025 MIT study shows that as many as 95% of generative AI pilots fail to deliver a measurable return on investment. While this percentage is widely cited in the marketplace, the crucial question of why these projects fail on such a massive scale often remains unanswered. The uncomfortable truth is that these failures are rarely about the technology; they're about the organization behind it. Companies soon realize that AI isn't a quick fix for organizational issues. In fact, it acts much more as a magnifying glass, exposing the flaws that were already there. 

The persistent misunderstanding about process improvement

The biggest mistake many organizations make today is assuming that AI can fix flawed processes. AI does not improve processes on its own; it merely automates them – albeit at an unprecedented pace. When a chaotic, unstructured process is linked to AI, the immediate result is a chaotic automated process. It becomes faster, more scalable, and (due to the technology's complexity) many times more difficult to reverse. But improving processes requires more.

Automating a poorly functioning customer service department does not resolve customer dissatisfaction. In this scenario, automation merely ensures that more customers receive a poor-quality response more quickly. The technology does exactly what it is told to do, but the input determines the output. Precisely because AI is so powerful and effective, it anchors existing errors and makes them a permanent part of business operations.

Building a chatbot is easy in practice; building the foundation is not

I recently saw how this mechanism works in practice at an organization that wanted to make its internal knowledge base accessible via an AI chatbot. Technically speaking, this was a simple project; the chatbot was up and running within a few weeks. The real challenge, however, only began when the tool actually had to start generating answers. Upon closer inspection, the knowledge base turned out to be far from optimally structured. Articles contradicted each other, outdated documentation was interspersed with new guidelines, and crucial business knowledge was not in the knowledge base but only in the heads of a few experienced specialists. AI is capable of much, but it cannot create logic within an illogical data structure.

Management then decided to rigorously overhaul the project. For six months, the chatbot was used exclusively by the support team internally. Months of intense, behind-the-scenes heavy lifting followed – with zero flashy demos – to improve internal knowledge management:

• Systematically analyzing the questions asked
• Manually correcting incorrect answers
• Thoroughly cleaning up the entire knowledge base
• Closing content gaps and interviewing specialists to explicitly capture implicit knowledge

Only after this foundation was fully restored did the chatbot’s success rate climb to 98% – a score that surpassed the average performance of an individual employee. At that point, confidence was finally high enough to roll out the technology externally, to customers. That confidence proved well-founded: the final results were excellent, and the AI regularly turned heads with the accuracy of its responses. That success, however, was solely due to the fact that the invisible groundwork had been done right – even if it happened retroactively.

The increased risk in vital sectors

The need for a solid foundation applies to any market, but the risks become exponentially greater in sectors where accuracy, traceability, and reliability are core values. Consider healthcare, financial services, publishing, and compliance-driven environments. In these segments, erroneous output can lead to direct legal or financial damage, as well as tarnish your good reputation.

If you apply AI to processes with unclear ownership, inconsistent data, or half-hearted documentation, a specific danger arises: the tool generates output that looks highly professional, grammatically correct, and persuasive, but is factually incorrect. This danger is many times riskier than using a system that produces patently bad results. Because the presentation is flawless, the need (or urge) to check information quickly diminishes, allowing errors to go unnoticed and end up in your workflow.

The four steps to a successful AI implementation

To ensure that AI actually adds value and doesn't result in a failed pilot, organizations must go back to basics. Successful projects almost never fail because of the technology implementation itself, but rather because crucial steps prior to that implementation were skipped.

An effective transformation always follows this fixed sequence:

1. Understand: map out exactly what is currently happening in the specific business process. Where is the human input, where does the friction arise, and how does the data flow?
2. Optimize: eliminate unnecessary process steps and simplify the logic. You shouldn't apply AI to a process that isn't efficient even in its analog form, because that will only amplify the inefficiency.
3. Document: document the optimized process and its data flows. This ensures that the information structure is transferable, unambiguous, and verifiable for both humans and machines.
4. Automate: this is when you actually start applying the AI component. In this phase, the technology is provided with a clean, reliable, and structured context, minimizing the risk of errors and maximizing returns.

Although AI appears to be deployed only in step 4 of the roadmap above, it's important to realize that AI can be extremely useful in all four steps.

The mirror AI holds up to us

AI won't automatically rid organizations of their internal inefficiencies. It's not a shortcut to streamlined operations, but a mirror that ruthlessly exposes the true state of your processes and data.

Although that may seem like a harsh conclusion, it actually carries a positive message for businesses. It means that the vast majority of the work required to make AI a success essentially has nothing to do with the technology itself. It is the fundamental organizational work that has often been on the back burner for some time. AI simply forces organizations to confront the backlog of maintenance in their processes. The effective path to innovation therefore has no technological shortcuts: start by cleaning up, and then automate only what has been proven to work. Only then you'll create the calm and solid foundation needed to make a real, lasting impact with AI.