Define the business output
Start with the thing the team needs: a ticket, a quote, a CRM task, a report, a draft reply, or a reviewed record.
RUNBOOK
Design the n8n workflow before adding AI.
Most failed AI automations fail before the model is involved. The real issue is unclear inputs, weak validation, missing review, and no fallback path.
Map the inputs
List the forms, emails, PDFs, spreadsheets, APIs, and system events that can trigger the workflow.
Build the deterministic path
Use rules, lookups, transformations, required fields, retries, and logs before asking AI to reason.
Add AI only where useful
Use AI for extraction, classification, drafting, summarizing, or comparison. Do not use it to hide a broken process.
Route exceptions
Bad data, low confidence, missing fields, duplicate risk, and risky actions should go to review instead of failing silently.
Hand it over
Document how the workflow runs, what can fail, how to replay a case, and who reviews exceptions.
RULE
If the workflow is unclear, AI will make it unclear faster.
The safest order is process, data, rules, review, then AI.
Map your first workflow