Field notes
AI draft with human approval
HITL helps, but only if the approve step is really before the send.
What Crash Labs saw
Gate FAIL · risk 100/100 · 2 outbound action(s) · coverage 100%.
"Very simple Human in the loop system email with AI e IMAP" should not go live yet: 4 critical issue(s) (score 100/100). Biggest risk: Nobody gets alerted when this workflow fails.
Why this breaks in production
critical
Ran 2 actions twice on a retry
2 email(s) ran again on the second delivery. In production that means duplicate emails, charges, or messages for the same event.
Fix → Before you send email or charge a card, store the order/message id. If that id was already handled, skip. Put a short queue in front of bursts so one webhook storm cannot spam.
critical
Sent the same email 20 times in a burst
18 of those actions happened after the first run. Example from the ledger: 50 copies of the same email went to info@n3witalia.com (subject: "[Approval Required] Order question").
Fix → Before you send email or charge a card, store the order/message id. If that id was already handled, skip. Put a short queue in front of bursts so one webhook storm cannot spam.
critical
Under heavy retry, saw 50× the same outbound action
50 copies of the same email went to info@n3witalia.com (subject: "[Approval Required] Order question").
Fix → Before you send email or charge a card, store the order/message id. If that id was already handled, skip. Put a short queue in front of bursts so one webhook storm cannot spam.
critical
Burst kept succeeding and burned budget quietly
In production you may not notice until the API bill spikes or the account is empty.
Fix → Add spend/quota alerts and a daily cap. Do not rely on the provider to fail loudly.
Run this pattern on your workflow
Upload your n8n JSON at Crash Labs. Same gate, blast-radius ledger, and shareable report for clients.