At a London panel I moderated with Simplito’s Kasia and Mirek, we explored the growing challenge of building AI that is sovereign, private and properly controlled.

For this episode of Futurise, I continued that conversation with Błażej, CTO of Simplito, going deeper into one of the biggest questions facing enterprise AI:

How much autonomy should we actually give an AI agent?

As agents gain access to enterprise data, systems and processes, privacy and security become fundamentally different problems. It is no longer enough to control the model or protect the data — organisations need to control what the agent can access, what it can do and what happens when it acts.

Błażej takes a deliberately challenging position: he doesn't believe fully autonomous agents have a place inside highly secure enterprise environments.

That opens up a fascinating discussion about the AI harness, privacy-by-design, sovereign infrastructure, permissions, governance and risk management — and whether the future of enterprise agentic AI is actually about controlled agency rather than autonomy.

For European organisations, the EU AI Act adds another layer of complexity, requiring careful consideration of risk, human oversight, monitoring and accountability for relevant high-risk systems.

But there is an interesting twist: the very controls that can make European AI adoption more considered are attracting interest from US organisations too. Rather than seeing European privacy and governance requirements simply as constraints, some US organisations are looking to the European approach as a way of building AI that is safer, more controlled and more trustworthy.

A conversation about privacy, sovereignty and where we draw the line between AI that can assist, AI that can act, and AI that should never be allowed to act alone.