The best escalation path is a product feature

AI products talk a lot about success paths. I am increasingly interested in what happens when the system is not confident enough, lacks the right tool, or encounters a case the workflow did not anticipate.

A graceful escalation path may be more valuable than another percentage point on an average benchmark. Ask a person. Use a stronger model. Narrow the task. Switch to read-only. Stop. Those branches determine whether the system fails usefully or simply becomes less predictable as the case becomes more important.

Escalation is often treated as an operational exception. I think it belongs in product design. The system should make the remaining safe options obvious rather than leave the user with a generic failure.