The part of an AI workflow that usually needs the most design is not the model call. It is the review surface after the model call. For document workflows, I like this structure: 1. Capture the document 2. Extract structured fields 3. Run a second check on sufficiency or risk 4. Merge the extraction and assessment 5. Write a review record 6. Notify a person only when action is needed The mistake is letting the AI output disappear into the next step too quickly. If the workflow says: "Evidence is sufficient" or: "This contract has medium risk" or: "This invoice needs approval" then the reviewer should be able to see what the automation used to reach that state. The review record should usually include: - source file - extracted fields - missing fields - risk level - rule or threshold used - next action - reviewer status That makes the workflow easier to debug and easier to trust. The pattern is especially useful in n8n because the workflow can stay simple: - parser node - assessment node - merge - code formatting - Sheets/Airtable log - Slack or email notification A safer production structure is not "AI decides everything." It is "AI prepares the review, and the workflow makes the handoff visible." Where do you usually put the review surface in your builds: the database, a Slack message, a dashboard, or inside the workflow execution log?

I know a few people in this community do cold email and I was wondering what reply rates are you guys seeing and have you built your own automations to automate cold email and how is it working out for you?