When AI writes the code

Two years ago, generative AI still only completed individual lines of code through autocomplete. Today, AI agents independently take over entire components, suggest test cases, and design architecture building blocks. According to Google Cloud's 2025 DORA Report, around 90 percent of engineering teams now use generative AI in their daily development work (Harvey & DeBellis, 2025).

For teams in medical technology, automotive, or finance, this changes day-to-day work considerably, because speed alone isn't what matters here. Every line of code must remain traceable, verified, and fit for approval before it can flow into a product. This is exactly where the question that occupies many engineering teams right now comes in: Which tasks can reasonably be delegated to AI, and where does human judgment remain indispensable?

Why traceability becomes the real bottleneck

In regulated industries, this changes the calculation. Standards such as IEC 62304 (International Electrotechnical Commission 62304) in medical technology or ISO 26262 in automotive require seamless traceability across the entire software lifecycle, regardless of who or what wrote the code. The upcoming second edition of IEC 62304 will explicitly address AI process planning for the first time (NSF, 2026).

Generative AI therefore shifts the bottleneck from writing code to reviewing and proving it. At the same time, who performs that review is changing too: classic developer roles are turning into reviewers and orchestrators, while newcomers in particular need to build new skills for working with AI output so they don't fall behind. Anyone modernizing legacy software, for example in medical imaging, gains real speed from AI mainly when human expertise continues to own architecture, risk assessment, and verification. This applies especially to systems that have been running in the field for many years, with documentation that has grown over decades.

Human-AI collaboration as the biggest lever

The consensus among experts is clear: it isn't full automation but the deliberate interplay of human expertise and AI-driven scaling that delivers real progress in safety-critical systems. This is already visible in practice: at individual software teams of major MedTech manufacturers, industry reports indicate that a substantial share of new code is now created with the help of generative AI, while the technical review remains entirely in the hands of engineers.

This exact interplay is at the center of the next sepp.med Afterwork Exchange. Tom Jachmann, who is responsible for the software platform of modern computed tomography (CT) systems at Siemens Healthineers, uses concrete practical examples from legacy modernization to show where AI is already boosting speed and quality today, and where human responsibility remains irreplaceable.

Join the live discussion on which roles are changing in your engineering team right now, and secure your spot at the sepp.med Afterwork Exchange.

Sources

• Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity (arXiv:2507.09089). METR.
• Harvey, N., & DeBellis, D. (2025, September 23). Announcing the 2025 DORA Report: State of AI-assisted software development. Google Cloud.
• Wessling, J. (2025). 2025 GenAI code security report. Veracode.
• Harding, B. (2025). AI Copilot code quality: 2025 research (211 million changed lines). GitClear.
• NSF. (2026). IEC 62304, draft of the 2nd edition: Optimized safety classes, extended scope and AI.