How can a workflow recover when a later step cannot undo an earlier action?

Map side effects before implementation

List steps that create external or durable changes, their order, and evidence that confirms each action. Decide whether compensation exists, such as sending a correction or creating follow-up work, and document where true reversal is impossible.

Separate preparation from commitment when possible. Validate inputs and permissions before the irreversible step, then make it a clear boundary with an idempotency key and a record of its result.

Recover deliberately after partial completion

If a later step fails, inspect run history to determine which actions already happened. Retry only safe steps; route non-repeatable outcomes to a reconciliation queue with context for deciding what happens next.

Make compensation visible and auditable, including who authorized it and what changed. Test failure points in a controlled environment so operators know whether the workflow resumes, compensates, or requires intervention.