Software that adjusts a pharmaceutical production line while it runs has to convince operators, quality teams and inspectors that every change is safe. Felix Müller has spent years working on that problem.

Felix Müller is the CEO and co-founder of Plus10, a German company that spun out of Fraunhofer research in 2019. Plus10 builds software that learns how a machine behaves from live pressure and temperature curves, spots a process drifting before it produces scrap, and shifts it back. It works in pharmaceutical and medical production, where tools must meet good manufacturing practice rules.

Felix explains why every update to a learning model needs a written explanation of what changed and how much more confidence it adds, and why an update with no benefit isn’t allowed. He describes the parallel models that check the live one. He also describes human review as a temporary step. Operators spend 6 to 8 weeks judging the system’s recommendations, then it runs on its own, because people asked to approve something 20 times an hour stop thinking critically about it.

We also talk about why pilots fail in regulated industries, why his most useful testers are the sceptics on the shop floor, why factories now automate because they can’t staff nights and weekends, how agents fit when guardrails must keep every recommendation inside the rules, and whether the European Artificial Intelligence Act helps European companies. Felix has advice for students too: he still spends 2 to 3 hours a week learning, as a chief executive.

We closed by asking whether we should hold artificial intelligence to higher standards than humans. Felix says an independent body would have to set and watch those standards, and he doesn’t yet know who that would be.

Hosted by Luis Lancos and Antonio Vieira Santos. License to Act is sponsored by DataWhisper, which does not appear in the conversation unless a guest raises it.

Support the show