From the Classroom to the Client: Living with the EU AI Act in 2026

Sunday Reflection — 2026-10-11

A few years ago, someone asked me what would finally bring order to the chaos of “AI in production.” My answer then was, “Nothing short of German paperwork.” Now the EU AI Act has arrived, and every consulting job this year echoes that prediction. We’ve swapped wild-west experimentation for risk assessments, documentation, and enough checkboxes to fill a motor pool. It isn’t all downside, but there’s no going back.

The most honest moment I’ve had with a client this month was watching a project manager realise that her favourite AI feature would need either a total rewrite or weeks of justification for the regulator. It’s all well and good in the planning stage; then someone opens the Act and realises you can’t deploy half your backlog without another round of forms, tests, and (my favourite) an “explainability statement.” In Germany, the answer is “more process.” Meetings multiply. Brits sigh and look for a workaround. On a live project last quarter, we spent three days building a tool that scanned prompts for “unapproved data categories.” It worked, but one intern could have explained why users jumped ship to tools without the guardrails. At least this is better than the old days: mines only go off if you miss something important. The EU is betting software can be made that safe, too.

mermaid flowchart LR A[Project Idea] --> B[Compliance Check] B --> |Passes| C[Development] B --> |Fails| D[Redesign or Discard] C --> E[Launch] D --> B

This year, my students in Berlin start every AI project with a list of “prohibited use cases” and a section on user rights. Five years ago, their ideas were wild—one wanted to build facial recognition for bike locks. Now, most play it safe. Good for some, frustrating for others—one team came up with a great healthcare chatbot that we couldn’t even discuss without an extra week just to check if it ticked a ‘high-risk’ box. I see the logic behind these rules. Especially in banking or health, oversight is overdue. But I can’t shake the feeling that teaching students to think more about process than possibility leaves the next generation a bit more timid than the last.

Regulation can stop the worst mistakes but tends to dampen the spark as well. It’s the same in software, the army, or the classroom: you need guardrails, but someone has to decide when the rules protect and when they just get in the way.