AI is taking over embedded software, but how do you guarantee reliability?
From programming to software verification: more and more parts of the embedded development process are using AI. This presents organizations with a dilemma, because how do you demonstrate that AI-supported software is safe, reliable, and certifiable? On WoTS 2026, the largest technology trade fair in the Benelux, three experts shared their vision on this issue during the seminar '‘Embedded Software and AI’.
In embedded software, a small error in the code can have far-reaching consequences for security and business continuity. This is certainly true for software in safety-critical systems, such as medical equipment, energy management systems, and defense systems. Industry and government therefore impose strict requirements on these systems, laid down in standards such as ISO 26262, IEC 61508, and IEC 62304.
Who is responsible
Transparency is a prerequisite for meeting legal requirements. Developers must be able to substantiate how design decisions were reached. However, that traceability is not a given with generative AI, according to Walter van der Heiden of VDH Informatica. In his opening presentation, he shows how generative AI can help structure requirements and link them to SysML and UML models.
This makes the development process more efficient, but it also raises questions about who is responsible when AI misinterprets a legal requirement. After all, a minor (interpretation) error at the beginning of an embedded project can later lead to costly design changes. Therefore, human assessment and independent validation remain necessary during every phase of development, concludes Van der Heiden.
Detecting architectural erosion
After this tough legal fare, Neil Langmead, co-founder of CodeClinic, addresses a current theme that many developers struggle with in practice: architectural erosion. This involves the gradual deviation from a software architecture through extensions and modifications of system requirements. Langmead presents an approach that combines Dependency Structure Matrix analysis with Large Language Models and reinforcement learning. In this context, AI acts as a tool to make unwanted dependencies between systems, modules, and functions visible, and to detect and minimize technical debt early on.
Domain knowledge remains essential. After all, architectural choices are closely linked to business processes, technical constraints, and previous design decisions. That context is not always fully available to an AI model. And so Langmead ends his presentation with the same conclusion as Van der Heiden: AI is a useful tool, but final control remains human work.
More AI requires stricter verification
As AI takes over more development tasks, the need for independent oversight grows. Especially with the rise of agentic AI; systems that not only propose code, but can also independently execute changes and start tests. Statistical analysis, runtime checking, Regression testing and compliance with guidelines such as MISRA and CERT are becoming more important, as became apparent during the lecture by Marc Thomas of IAR Systems. In his concluding presentations, he discusses the verification of software developed entirely or partially using AI and the challenges this entails.
The presentations all point in the same direction. AI does not replace the embedded engineer, but his or her role does change. The focus is increasingly on oversight, decision-making, and quality assurance. AI can accelerate the development process, but trust remains dependent on verifiable processes and careful engineering work.
Join the Software Cluster
This seminar is an initiative of the Software Cluster of FHI, an active network of embedded software suppliers. Annually, the members organize thematic meetings on current topics and a seminar on (embedded) software during the tech fairs WoTS and E&A.
Related events: