A regulated environment such as medical technology poses special challenges for test automation. Anyone developing machines here needs systems and processes that remain traceable and enable innovation.
In this episode, I talk to Alexander Frenzel about AI-supported test case generation in a strictly regulated environment. Sounds like a contradiction. But it's not. Alex shows how an assistance system with a one-click generator, modular models and a RAG system derives tests from requirements. Instead of asking questions, he uses the Hyde principle: make a claim, find evidence from your own documentation. This preserves traceability, logging and the human in the loop.
"First of all, you have to be able to see what your individual process steps are. I need to be able to intervene everywhere. I need traceability." - Alexander Frenzel
Alexander Frenzel has been working at Fresenius Medical Care Deutschland GmbH since 2019 as Expert for Test Management, Global Verification & Validation Lead for several product series and has also been Director Test Management since 2025.He is a member of the QA Day Program Committee and Board Member of the German Testing Board.
Software testing with artificial intelligence? In many industries, this sounds like a dream of the future. In medical technology, however, it's not just about progress, but above all about safety and responsibility. In the podcast Software Testing, Alexander Frenzel from Fresenius reports on how his team tackled precisely this challenge.
Imagine developing systems on which human lives depend. That's exactly what the teams at Fresenius Medical Care do. They build dialysis machines. There, patients' blood flows through systems consisting of software, electronics and hardware. This doesn't work half-heartedly - it has to be safe, reliable and verifiable. All of this is controlled by strict rules and specifications. Innovation projects such as AI in testing quickly reach their limits in such an environment.
Alexander Frenzel describes how the big goal of their proof of concept came about: they wanted to effectively relieve their testers by developing an AI-based assistant. The idea was to automatically generate test cases for software modules. But how do you do that when every step has to be documented, justified and tested?
AI should not replace people in the Fresenius team, but rather help them. Test case generation for complex machines and old systems is particularly time-consuming, explains Alexander Frenzel. Documents are often missing digitally, data is distributed across different systems and requirements change over many years. "We want to bring the time of five days per test case down to a reasonable level."
How does that work? Fresenius relies on an agent principle: an AI system not only creates test cases, but first builds the appropriate prompts itself. A subsystem generates instructions for the actual AI, which then compiles the test case. The goal: to get from the requirement to the test case with a single click.
Safety is the be-all and end-all. The project was clear: no artificial hallucinations, no dangerous fantasies. That's why the architecture was given a special feature: the AI was not allowed to make up new facts, but had to draw everything from existing documentation in a clearly comprehensible way. All requirements, dependencies and data came from a central document system ("rack system"). Instead of asking the AI questions, the system formulated hypotheses, which it then backed up with real documents.
This meant that everything remained verifiable. The knowledge of the experts was incorporated, as was the experience of the testers and system architects.
What was the outcome? A lot of euphoria at first. In internal testing, the system helped to make even hidden documents visible. Suddenly, testers found information that they would never have discovered otherwise. This not only helped with individual tests, but also improved the transfer of knowledge within the team.
The biggest relief: instead of starting from scratch, there was now a usable test case draft that experienced testers only had to adapt. This saved time - and brought quality. "Someone still has to check and approve it, but the basis is there much faster," says Alexander Frenzel.
However, limits also quickly became apparent. AI was not yet sufficient for purely physical or dynamic processes, such as the exact pressure in hoses and pumps. Models and expert knowledge are still needed here.
New questions arise in the regulated environment in particular. How do you validate a system that makes creative suggestions? Alexander Frenzel explains: "The key is that the AI only provides suggestions, never releases them independently. Every step remains verifiable, logged and traceable. The results remain a "draft". They are only released after a human review - with a signature.
This means that classic tool validation is not even necessary. The system is an intelligent tool, not a substitute for human decision-making and responsibility.
The proof of concept is not an experiment. The team plans to integrate the assistant directly into systems that every tester uses on a daily basis. And the next logical step: why not generate automated test scripts right away? This could make the system even faster. Despite all the technology, one thing remains clear: "The man in the middle" is still essential.
AI in the testing of high-risk systems such as dialysis machines is possible - if technology, expertise and responsibility work together. The mixture of a structured approach and clever use of modern technology can set an example. For innovation - and safety.