Agents do the work, humans own the judgement
Jul 31st, 2026 by Jesper Kråkhede
With Project Perception and similar systems, AI agents will soon take real actions in our environments: isolating machines, changing configurations, closing attack paths. For management, the question is how to keep control.
The principle I believe in is simple. Agents do the work. Humans own the judgement.
That means a few things in practice. Define what an agent may do on its own. Low risk, reversible actions, like isolating a normal workstation, can be automatic. High impact actions, like changing Conditional Access, isolating a production server or disabling an admin account, need a human to approve.
Make sure every action can be traced. Who, or which agent, did what, when and why. If you can´t replay the decision, you can´t learn from it, and you can´t defend it.
Name an owner for each agent and each automated workflow. Someone who is accountable for what it does.
And review it regularly. Look at what the agents did, how often they were wrong and where the limits should move.
There is a compliance benefit here. Regulations like the Swedish cybersecurity act and the AI Act expect you to be able to show how decisions are made and who is responsible. An agentic system with full traceability can actually make that easier than the manual work it replaces, where the reasoning often lived only in someone´s head.
Autonomy without accountability isn´t innovation. It is a risk with a nice demo.