Who owns the exception?

A useful question for an AI-assisted workflow is not only what it can complete. It is what happens when it cannot complete the work responsibly.

An ambiguous request. Conflicting sources. A missing approval. A situation the instructions did not anticipate.

Stopping is a workflow too

“Escalate to a human” sounds clear until we ask which human, with what information, and what happens while the work waits.

An escalation that sends an unexplained error to an unowned inbox is not much of a handoff.

What I want to explore

How much context should travel with an exception? When should a system offer options, and when should it simply stop? How does the reviewer distinguish uncertainty in the source material from a failure in the system’s own interpretation?

These questions belong in the design of the work, not just the instructions given to a model.

I am interested in examples of escalation paths that people can actually use, including the less impressive but important case where the right answer is to wait.