How companies prevent AI projects from failing
Is it a technology gap or a problem with execution? Let’s take a closer look. A typical example highlights the challenges. A mid-sized machinery manufacturer wanted to predict quality issues. Competitors reported success with predictive quality, and management did not want to fall behind. A budget was approved and a development team was set up. The goal, however, remained vaguely defined: reduce defects and save costs.
The first few months looked very promising. A pilot project for a single machine type delivered impressive results. Seventy-five percent of quality issues were detected. The solution was based on manually prepared data from a three-month period. Excited, management demanded that the solution be rolled out across all production lines.
However, when the project was scaled to all production lines, reality struck. The system only worked for the machine types tested and could not be applied to other machines. The manual data extracts could not keep up with daily operations. Automated pipelines were missing, and standards did not exist. After twelve months, management pulled the plug.
Unfortunately, this example is not an isolated case. Almost half of all pilot projects fail before they reach production. The problem is not the technology, but the execution.
The difficulties in this example and the ultimate failure can be traced back to four key levels:
1. Strategy First, Technology Second
AI projects often start technology-driven. Companies implement systems because the technology is trending or competitors are already using it. But often, the focus on a specific problem is missing. An AI solution can be technically perfect, but if its results are not used, it creates no real value.
In successful AI projects, the process is different. Companies first identify the problem, define the target outcome and how AI can help, and thoroughly evaluate the business case. Technology is a means to an end. Business processes are analyzed, suitable AI solutions are developed, and the process is then adapted accordingly.
2. Data is the Foundation
The pilot project worked because a small dataset was enough. When scaling up, reality hit. The data was spread across different systems, standards were missing, and quality varied. Without a central platform and automated pipelines, rollout was impossible.
Data quality remains the biggest challenge for AI project success. Often, 80 percent of project time is spent preparing data. Fragmented data landscapes are the norm. ERP systems, MES, IoT sensors, quality management systems, or Excel spreadsheets each have their own structures and logic. Without consistent definitions and standards across systems, inconsistencies and inefficiencies arise when trying to harmonize data manually.
A solid data foundation is essential for successful AI. Companies that invest here reduce effort in later project phases and create the basis for scalable solutions.
3. Clarify Skills and Governance
Lack of know-how is another key reason projects fail. AI developers are scarce, but just as important is understanding within the business units. Who monitors data quality? Who approves models? Who defines KPIs? Without clear roles and processes, chaos and frustration arise.
A governance framework ensures responsibilities are defined and data competence grows within the company. When teams understand the capabilities and limits of AI, they can make informed decisions and secure the value of their projects.
4. Take an Iterative Approach
Many companies claim to be agile on paper, but in practice they follow a more waterfall-like approach. This shows up in long pilot phases, with the expectation that the solution will then be rolled out across all production lines. Without a step-by-step approach, even very good models quickly reach their limits, because technical metrics alone say little about actual value.
A high detection rate (in our example, 75 percent) sounds impressive, but it does not tell us how much waste was actually avoided or whether these savings justify the development and operating costs of the new solutions.
Successful projects rely on short cycles. A single production line is automated first, observed, and improved. Only then is the next line addressed. Incremental implementation reduces risks and ensures that solutions work sustainably in day-to-day operations.
Conclusion
In the end, the success of AI projects does not depend on technology alone. Clear goals and a solid data foundation are crucial. Defined roles, clear processes, and short, iterative cycles are also essential. Only when all of these elements are in place will a project work sustainably. AI then becomes more than a trend, it actively helps improve processes, reduce waste, and control costs.
In my next blog post, I will explore how AI in companies is much more than just ChatGPT. It will cover topics such as:
- Analytics: How companies understand what is happening and why. This includes dashboards, root cause analyses, and pattern recognition.
- Machine Learning: How companies make predictions, for example predictive maintenance, demand forecasting, or quality predictions.
- Generative AI and Autonomous Agents: How AI handles repetitive tasks on its own, such as checking invoices, reconciling deliveries, or triggering processes automatically.
Sources: S&P Global, Informatica
