n8n 表单接单 AI 分流工作流：14 节点实现校验、LLM 分流与告警

A production-shaped n8n workflow that replaces "someone watches the contact form inbox" with an automated pipeline: validate → LLM triage → route → log → notify → acknowledge, with a failure branch that retries and shouts.

14 nodes, imported and run on n8n 2.35.7 (self-hosted, community edition).

What it does

1. Webhook receives the form POST on /webhook/intake-form.
2. Validate & Normalize (Code node) checks the required fields, trims and lowercases the email, and returns either {valid: true, ...} or {valid: false, missing: [...]}.
3. IF — Required Fields OK splits the two cases. Invalid input still produces a notification instead of a silent drop.
4. OpenAI — Triage & Summary asks gpt-4o-mini for JSON only: urgency, category, a one-sentence summary for the sales team, and a suggestedReply.
5. Switch — Route by Urgency sends urgent and normal down separate paths.
6. Google Sheets — Append Lead, then Slack — Notify Team, then Gmail — Auto Acknowledgement sends the drafted reply to the sender.
7. Error branch (bottom row): Error Trigger → Wait 30s → retry the original call → Slack #alerts, so a failed run is loud rather than invisible.

Node reference

Design decisions worth pointing at

• Validation is a Code node, not a chain of IF nodes. One place to read, one output contract, and the invalid path carries which fields were missing so the alert is actionable instead of "something failed".
• The LLM is asked for JSON only, and the Switch parses it defensively — $json.message.content ? JSON.parse(...).urgency : $json.urgency — so a model that returns plain text degrades to the fallback branch instead of throwing.
• The error branch is a first-class path, not an afterthought. Retry with 30 s backoff, then alert with lastNodeExecuted and the error message. Silent failure is the failure mode that actually costs money in automation work.
• Credentials live in n8n's credential store, never in the workflow JSON. The exported file in this repo contains no secrets — the placeholder values (YOUR_SHEET_ID, https://api.example.com/retry) mark where your own values go.
• The workflow file is version-controllable. It is a plain JSON export, so it can be reviewed in a pull request, diffed between versions, and restored after a mistake.

Import

# CLI
n8n import:workflow --input=workflow/form-intake-ai-triage.json

# UI: Workflows -> ... -> Import from File -> select the JSON

The import is inactive by default ("active": false). Activate it only after the credentials below are attached and the sheet ID is replaced.

Then test the webhook without wiring up a real form:

curl -X POST http://localhost:5678/webhook/intake-form \
  -H "Content-Type: application/json" \
  -d '{"name":"Dana Reed","email":"Dana@Example.com","company":"Northline",
       "message":"Our invoice exports take two days every month and we keep missing cut-off."}'

Expected: a row appended to Leads, a message in #leads, an acknowledgement email to the sender, and HTTP 200 back from the webhook.

Also call it with {"name":"","email":"","message":""} — that is the path that exercises the invalid-input alert, and it is the one people forget to test.

Credentials checklist

See docs/credentials-checklist.md for the handover version of this list.

Minimum scopes and the two Google-side gotchas (sharing the file with the service account / OAuth user, and enabling the Sheets API) are in the checklist.

What changes for a real deployment

The JSON here is a working template, not a drop-in. On a paid job the parts that get rewritten are:

• the validation rules (which fields are actually required, format checks, spam and duplicate-submission handling),
• the routing rules (urgency is rarely just two buckets; SLA timers usually appear),
• the lead destination (Sheets is fine for a small team, Airtable/CRM/HubSpot once volume grows past a few hundred a month),
• idempotency — if the form can be submitted twice, a dedupe check on email + timestamp key belongs before the Sheets append.

Limitations

• Retry is single-shot with a fixed 30 s backoff. For a rate-limited API, exponential backoff with jitter is the correct version — the shape is the same, the Wait node's delay just becomes computed.
• The HTTP - Retry Original Call node retries a call, not automatically the failed node — n8n has no built-in "re-run from the failed node" expression. For real use, the more robust pattern is Retry On Fail on the fragile nodes plus this branch as the final alert.
• No queueing. A burst of submissions runs concurrently; if the downstream API has a concurrency limit, add Split In Batches or Limit.
• Gmail's automated reply comes from a human-looking mailbox. Check the client's expectations before enabling it — some teams prefer a reply-to address instead.

License

MIT — see LICENSE. The workflow is a template; replace the placeholder endpoints and IDs with your own before running it.