Innovation at Pulse Rate – AI as an Accelerator in Embedded Development for MedTech

Relevance

Medical devices such as surgical robots, insulin pens, and intelligent imaging systems are becoming increasingly complex, while regulatory requirements (e.g., IEC 62304, ISO 14971, and ISO 13485) impose higher hurdles. Developers face the challenge of balancing innovation speed with compliance. This is precisely where AI-supported tools come into play: they provide initial drafts, formulate testable requirements, and generate structured source code. However, the responsibility for architecture, safety cases, and regulatory approvals remains firmly in the hands of experienced experts.

Study Design

The study compared traditionally working teams with those using AI tools such as ChatGPT Enterprise, the internal ERNI LLM "AIDA," GitHub Copilot, and domain-specific prompt libraries. All artifacts were created in a revision-secure manner and evaluated anonymously.

Results in Detail

In Requirements Engineering, the AI-supported teams generated a broad and clearly structured requirements base from a few functional descriptions, complete with Gherkin acceptance criteria. In contrast, the manually working team focused on typical user flows and essential requirements—a strategy that was more user-centered but less comprehensive. The AI demonstrated its strength through a comprehensive system perspective and clear structure, but occasionally tended toward "overengineering."

In software development, the use of AI tools led to significantly faster results. Structured code with clear module separation and syntax was produced in half the time. The generated proposals showed good architectural approaches (e.g., sensor abstractions, dependency injection) that the developers further refined. Human post-processing was still necessary to ensure, for example, memory consumption and real-time capabilities.

The differences in test case creation were particularly pronounced: the AI not only covered functional positive cases but also automatically supplemented negative scenarios, security aspects, and load tests. Each test was linked to the corresponding requirement IDs, greatly simplifying traceability. Interestingly, the AI also suggested tests that the manual group had not considered—such as in exceptional situations or unexpected user interactions. While the manual team developed targeted, practical tests through routine and experience, longer working hours led to a decline in concentration—an effect that AI does not experience. However, its tireless working method can also lead to the overproduction of irrelevant tests.

These contrasts reveal a central finding: while the AI works consistently structured, comprehensively, and tirelessly, it requires guidance. The human team, on the other hand, knows the domain, understands real user needs, and recognizes when a test case may be theoretically sensible but practically irrelevant.

Lessons for Practice

AI is not a replacement for human engineering but a powerful partner. Its strengths are particularly evident in systematic structuring, rapid generation, and resistance to fatigue. However, human expertise remains irreplaceable for assessing relevance, identifying risks, and interpreting regulatory requirements.

AI can be especially valuable as a pair programmer and assistant: junior engineers can learn faster and adopt structured working methods through AI suggestions. Senior experts could then shift their focus more toward architecture, risk analysis, and reviews. At the same time, it became clear that automatically generated artifacts require thorough reviews to be eligible for approval. Prompt engineering is emerging as a new key competency.

Key Findings at a Glance

• AI reduces processing time by up to 60% and increases test coverage by more than double.
• Complete trace chains significantly reduce maintenance effort.
• Reviews remain essential but enhance overall velocity when scheduled early and frequently.