Deutsch (DE) 
English (EN) 
In recent years, AI has become a buzzword in machine vision, which has led to it being viewed in many projects as almost the only possible solution. It sounds modern, powerful, and often like a shortcut to a complex goal. In reality, however, a different picture emerges time and again: expectations are high, promises are bold, but the actual results often fall far short of what was originally presented.
This is exactly where projects often run into what could casually be described as “scorched earth.” Investments have already been made, expectations have been raised, and in the end, what remains is mainly disappointment. Trust in AI-based approaches is damaged, even though the real cause often lies not in the technology itself, but in how it is applied. In such cases, AI does not fail because of a lack of potential, but because of unsuitable conditions, an insufficient system design, or unrealistic expectations.
Even the best AI can only analyse information that is really present in an image
Quality starts before the model
For that reason, the success of a machine vision system does not begin with training, but with image acquisition itself. Choosing the right camera, the right lens, and an appropriate resolution determines whether the relevant features can be captured with sufficient clarity and repeatability in the first place. If contrast is missing, details are not resolved, or the field of view does not match the task, this can only be corrected to a very limited extent in software later on.
Lighting is just as critical. In industrial applications in particular, it often determines whether a feature becomes reliably visible or gets lost in process noise. Reflections, shadows, gloss, material inconsistencies, or changing surface conditions can confuse an AI system just as much as they can confuse classical algorithms. Add to that the geometric setup of the system, the object position, and the repeatability of the process. Only when these fundamentals are properly aligned can AI actually play to its strengths.
Another often underestimated point is synchronization with the object under inspection. In many applications, image capture and object motion must be precisely coordinated, for example via trigger signals or encoder feedback. Motion blur must also not be dismissed as a side effect. It not only reduces image quality, but often directly affects the reliability of the final evaluation.
Bad image data will not lead to a good result only thanks to AI, a robust result does not start with training
The right architecture
Not every task needs to be solved with AI. In many cases, a solution based on rule-based methods may be the more suitable choice.
Especially for clearly defined geometries, simple measurement tasks, completeness checks, or fixed tolerance criteria, classical methods are often more transparent, more resource-efficient, and easier to validate. AI tends to show its advantages where there is high variability, complex patterns, or feature characteristics that are difficult to describe. The best solution is therefore often not the most “modern” one, but the one that makes the most architectural sense. Combining both worlds intelligently often leads to more robust, more explainable, and more economical systems.
It is also often the combination of classical rule-based methods and AI-based components that delivers the best results.
Not the most modern, but the best fitting architecture is the best solution.
Why experience matters
For that reason, choosing an experienced system integrator is often a key success factor. A good integrator does not only look at the AI model, but at the entire system: optics, camera, lighting, mechanics, process integration, data foundation, software architecture, and long-term operability. Only this interaction creates an industrial solution that performs reliably in day-to-day operation.
An experienced partner also brings something particularly valuable in the early project phases: the ability to identify risks early. These include unclear requirements, unsuitable data, insufficient feature distinctiveness, unstable process conditions, or unrealistic expectations regarding accuracy and generalization. Addressing these issues early helps avoid later disappointment and reduces the risk of creating “scorched earth” once again.
There are also other aspects that are often underestimated in practice: maintainability, retrainability, extensibility, and documented validation. A machine vision system must not only work in a demo, but also in live operation, under real cycle times, and over a longer lifecycle. Changes to products, batches, variants, or environmental conditions are usually the rule rather than the exception. An experienced system integrator plans for such changes from the outset.
Regulatory and security-related requirements are also becoming increasingly important. In networked industrial systems in particular, taking the CRA, the Cyber Resilience Act, into account can be a relevant part of the project implementation. An experienced integrator can not only implement the technical solution, but also systematically address documentation, responsibilities, updateability, and security requirements. For companies, this is an important difference between a short-term impressive standalone solution and a long-term viable system.
