Required fields
Does the output contain every field needed by the next system?
RUNBOOK
AI output needs checks before it becomes action.
Inside n8n, evaluation can be simple: required fields, source support, confidence, policy rules, risk level, and human review before sensitive actions.
Source support
Can the answer point back to the email, PDF, CRM note, or knowledge source?
Risk level
Does the output affect money, customers, legal terms, safety, or a system of record?
Auto-continue
Safe, complete, low-risk outputs can continue to the next workflow step.
Human review
Uncertain, incomplete, or sensitive outputs should wait for a person.
Fallback
Broken inputs, API errors, and missing context should create a clear exception task.
PRINCIPLE
Evaluation is part of the workflow, not an afterthought.
Good AI automation is visible, testable, and reviewable.
Scope an AI workflow