Knowledge · agent
Controlling AI Agents: Approvals Instead of Flying Blind
How to stay in control of AI agents: the approval principle for external actions, clear instructions, minimum confidence, and traceability.
Control over AI agents doesn't come from trust, but from architecture: clear instructions, graduated automations – and the approval principle, under which actions with external effects run through a human's decision. A well-built agent works ahead boldly, but emails, calls, or master data changes only go out once you've reviewed the suggestion. That's the difference between delegating and flying blind.
The Basic Principle: AI Prepares, You Decide
The fear of agents "running out of control" usually rests on a false image: the agent that acts externally without being asked. Seriously built systems flip this around. The agent detects, researches, drafts, and proposes – but at defined points it stops and asks. You review the suggestion along with its context and approve or reject. This way you use the speed of AI without giving up responsibility.
In webRichtung agent, this point is called queries: a dedicated area that shows what's waiting for your okay. The assistant's basic rule there is: external effects – emails, calls, master data – run through your approval.
Building Block 1: Instructions That Name Conditions
Control begins before the first task. Good instructions tell the agent not only *what* it should do, but *under which conditions*: "Only create deadlines when the date, obligation, and source are clearly documented – merely report uncertain cases, don't create them." The clearer the conditions, the more predictable the behavior.
Building Block 2: Activate Automations Individually
Autonomy shouldn't be an all-or-nothing switch. Better: activate capabilities individually and equip them with adjustable settings. One example is the minimum confidence for automatic creation – if the agent isn't confident enough, it creates nothing but reports back instead. Guessing is not an option. This way you determine for each task how much autonomy is appropriate.
Building Block 3: Traceability
You need to be able to see what the agent did and why. This includes queries with justification and source (such as the document a detected deadline comes from) and statuses that document the progress – open, approved, rejected. Decisions thus remain verifiable, even after the fact.
What Control Does Not Mean
Two honest clarifications belong here. First: even a controlled agent can make mistakes – like a human employee. Approvals, good instructions, and occasional spot checks are therefore a working principle, not a one-time setup. Second: the approval principle is not a brake, but the prerequisite for speed. Because critical matters run through your decision, the agent can work ahead independently on the rest – without this safety net, you would have to set much tighter limits for it.
Conclusion
The right question is not "Can I trust the AI?" but "Is the system built so that I don't have to do it blindly?". Instructions, graduated automations, and approvals for external effects answer it. How these building blocks work together in practice is shown in the documentation on Automations & Approvals – and getting started works with Deploying an AI Agent in Your Company.
FAQ
How do I stay in control of an AI agent?
Through three building blocks: clear instructions on what the agent is allowed to do, graduated automations with adjustable settings like a minimum confidence – and approvals for actions with external effects.
What does the approval principle mean?
The agent only prepares actions with external effects – emails, calls, master data changes: you receive a query with context and decide before anything goes out.
What is a minimum confidence in automations?
An adjustable setting for automatic action: if the agent isn't confident enough, it doesn't create anything but reports the case for a decision – guessing is not an option.
Can an AI agent still make mistakes?
Yes, just like a human employee. That's why approvals, good instructions, and spot checks are part of the working principle – control is a process, not a one-time setting.
Does the approval principle slow down automation?
No – it enables it: because critical matters run through your approval, the agent can boldly work ahead on the rest. Without this safety net, you would have to set much tighter limits for it.