Updated 17 October 2026. An automated workflow can complete many steps and still leave one awkward decision unresolved. That is where the organisation discovers whether autonomy came with an operating model.
Design the exception queue
Name the conditions that require a person: missing evidence, conflicting instructions, unavailable permissions or a consequence outside the approved scope. Each condition needs a destination and enough context for the next person to act.
An illustrative agent preparing a supplier payment might stop when the invoice and purchase order disagree. The useful handoff includes the mismatch, the relevant records and the action not yet taken. “Something went wrong” sends the next person back to the beginning.
Give the receiving team a way to resolve, reject or return the task. Record whether the agent may resume and what must be rechecked before it does. A stale approval should not silently survive a changed amount or recipient.
Watch the queue as a product signal. Repeated exceptions may reveal a missing capability, unclear instructions or a process that should remain manual. The answer is not automatically to remove the review step so the dashboard looks smoother.
| Question | Evidence or action |
|---|---|
| Trigger | Specific unresolved condition |
| Handoff | Context and action status |
| Resolution | Owner decision and safe resumption |
The practical next step
Run one exception from beginning to end. If nobody can say who takes it next, the workflow is not fully designed, however autonomous the happy path looks.
Keep the resulting decision, its owner and unresolved questions in a short decision record. For claims about agent behaviour, connect the decision to evaluation evidence rather than relying on a successful demonstration alone.
Further primary reading: NIST AI Risk Management Framework. The review map and product recommendations above are my analysis, not a certification or a legal assessment of a particular deployment.
