At TechBBQ, Europe's AI conversations kept returning to one question: who is actually in control

TechCrunch summed up the TechBBQ conference: whatever topic an AI panel announced — investment, regulation, products — the conversation invariably converged on control. Who answers when an autonomous system errs: the model vendor, the company that deployed it, or the specific employee who pressed «run»? The European habit of settling such questions through regulation only sharpened the debate.

Behind the philosophy is a practical shift: AI has stopped merely advising and started acting — sending, paying, changing data. An AI agent that executes process steps itself is exactly as much more convenient than a chatbot as it is more dangerous: the cost of error turns from a wrong answer into a wrong action. So the «who is in control» question is moving from conference halls into company regulations.

The answer practitioners converge on is mundane: an agent needs explicit authority boundaries, actions above a set threshold must be approved by a human, and every step must leave a trace from which you can reconstruct who decided what. We hold to the same principle in 1Cifer: agents' serious actions go through routed approval, and the full decision history is kept. When deploying agents, build these three elements in from day one — cheaper than adding them after an incident.