Three Nodes to Wire an AI Intake Triage in n8n
We wrote about the spec-form triage for the transmission shop in Kokomo: 12 to 20 custom build requests per week, each requiring 20 to 35 minutes of skilled-tech time to read the spec sheet, match it to the product catalog, and draft a quote recommendation. The AI triage cuts that to a two-minute review.
We've described what it does. Here's how it actually wires together in n8n.
The Premise
The spec form already submits. On every WooCommerce or Gravity Forms submission, the site fires a webhook. That's the trigger. We don't need to change the form, the hosting, or anything the customer touches. We need to catch the POST, process it with Claude, and put the output somewhere a tech can act on it.
Three nodes.
Node 1: Webhook
An n8n Webhook node listens for the POST from the form plugin. When the form submits, the payload arrives as JSON: all 40 fields, plus a file URL for the dyno sheet the customer uploaded.
No transformation needed yet. The webhook captures it and passes the raw payload to node two.
Node 2: HTTP Request to Claude
An HTTP Request node calls https://api.anthropic.com/v1/messages directly. The body is a JSON payload with a system prompt and a user message assembled from the form fields.
The system prompt is where the domain knowledge lives. For this shop, it tells Claude what it's reading: a high-performance automotive spec form for converter and transmission builds. It specifies the output format: converter stall range, transmission recommendation, any flags from the dyno data or field combinations that don't add up. It tells Claude to return valid JSON and nothing else.
The user message is the form payload, formatted as a readable block. Something like:
Engine: 454 BBC
Cam: solid roller, 260/270 @.050
Stall target: 3,200-3,600
Application: bracket drag
Dyno peak: 620 hp @ 6,400 rpm
[...remaining fields]
Dyno file: [URL]
Claude reads the fields, matches them to known converter stall ranges and transmission specs, and returns a JSON object: recommended stall range, recommended transmission, any notes or flags.
This is not a guess. The spec form was designed by people who know what data a good quote requires. Claude pattern-matches structured data to known recommendations. The edge cases, unusual field combinations, dyno anomalies, are what goes into the notes field and ends up in front of the tech.
Node 3: Route to Review
The third node parses the JSON response and routes it somewhere a tech can act on it. Options: email to the responsible tech, a Slack message to the quotes channel, or a write to an Airtable record that populates a review queue.
The review artifact shows the AI draft alongside the original submission. The tech reads both, corrects anything off, and approves. The quote goes out.
Total tech time per submission: two minutes instead of 25.
Why This Works
The spec form is pre-structured intake data. Claude doesn't need to extract meaning from a vague description. It needs to read 40 labeled fields and return a match against a product range it can be trained on via the system prompt. That's a reliable task.
The pattern holds for any structured intake that routes to a skilled person. HVAC equipment selection, custom fabrication quotes, specialty repair forms: if the intake form is already structured, the webhook fires already, and the output is a recommendation a human reviews before it goes out, three nodes is the build.
The first month you'll tune the system prompt. By month two, the draft quality is high enough that most reviews are approvals. The tech's value moves from reading forms to catching the edge cases the prompt hasn't seen yet.
That's the right use of a skilled tech's time.