AI & autonomy
Closed-loop execution
A workflow model that treats an external action as complete only after observing and reconciling authoritative outcome evidence.
Glossary
AI & autonomy
Closed-loop execution connects intent, authorization, external action, acknowledgement, and reconciliation. A system does not mark a message sent, payment settled, signature completed, or vendor assignment complete merely because it prepared or queued the request.
The loop needs a correlation key and explicit states such as prepared, authorized, sent, acknowledged, failed, unknown, reconciled, and reversed. Webhooks can help, but the design also needs polling or manual recovery when an event is delayed or missing.
Example: creating a payment request is an internal action. The loop closes only after the payment provider returns an authoritative status and the ledger and bank-facing record reconcile to that event.
Minimum evidence chain
Retain the triggering record, policy or approval, idempotency key, outbound request, provider identifier, acknowledgement time, status changes, exception handling, and resulting internal record.
Unknown is a real state
A timeout can leave the external provider successful while the local system saw no response. Retrying blindly can duplicate an action. Query by the original idempotency or correlation key and escalate unresolved outcomes.