The line between autonomy and escalation

Every AI system needs an escalation point: the exact condition under which the agent stops acting on its own and calls in a human. Most companies don't define this. They trust the system's good judgment and only discover the problem when something breaks.

The result is two equally bad extremes. On one side, the agent that escalates everything paralyzes the automation and turns the workflow into manual work with a useless layer of technology in the middle. On the other, the agent that decides what it shouldn't: silent risk, where no one knows the error happened until it turns into a real consequence.

In the projects I structure, the first thing I document is not the happy path. It's the escalation map. Three questions I answer before releasing any agent:

𝟭. What is the irreversible action in this flow?
Sending an email, processing a payment, deleting a record, publishing content - any action without ctrl+Z needs a human checkpoint, no exceptions.

𝟮. What is the minimum acceptable confidence level?
The agent acts when confidence is above X. Below that, it escalates. That number is defined with the team: it's not a system default, it's a business decision.

𝟯. Is the context familiar or ambiguous?
The agent acts within the mapped patterns. A new situation, an outlier, an unknown variable - always escalate. Autonomy only within what has been documented.

Unlimited autonomy becomes delegation without accountability. A company that doesn't define where the agent stops ends up discovering that limit in the worst possible way.

Save this post if you're building or operating AI agents - this framework will come in handy when the project scales.

Tell me in the comments: what is the biggest escalation risk in the flow you're automating today?