Routing inbound leads with N8N + Claude — the 6-node workflow I keep cloning

Most early-stage startups don't lose deals on price. They lose them on "sorry, we'll get back to you soon." Four hours later, the lead has already talked to two of your competitors.

Here's the workflow I built for a D2C skincare brand that took their median inbound response time from 4 hours to 11 minutes — and it's six N8N nodes.

The six nodes

1. Webhook — single endpoint hit by the website form, Instagram DM forwarder, and the Meta Lead Ads connector.
2. Set — normalise the payload into a flat shape: { name, contact, source, message }.
3. HTTP Request → Claude — classify the lead.
4. Switch — route by classification.
5. Slack — post into the right channel with the AI-suggested first reply pre-drafted, so the rep can hit Send in one click.
6. Google Sheets — append a row to the master CRM for later analysis.

The Claude call

The model does three things in one round-trip: classify intent (buyer, affiliate, noise), score budget hint (low, mid, high), and draft a first reply in the brand voice.

{
  "model": "claude-sonnet-4-6",
  "max_tokens": 400,
  "system": [
    {
      "type": "text",
      "text": "You triage inbound leads for a D2C skincare brand. ...",
      "cache_control": { "type": "ephemeral" }
    }
  ],
  "messages": [
    {
      "role": "user",
      "content": "{{ $json.message }}\n\nReturn strict JSON: {intent, budget, reply}"
    }
  ]
}

The cache_control marker is the unlock. The system prompt — brand voice, product catalog snippets, do-not-promise rules — is large and stable. Caching it cuts the per-lead Claude cost by ~85% once you cross a few thousand calls.

The Slack message

🟢 Buyer · mid budget
Name: Aanya · IG DM
"Hi, do you ship the vitamin C kit to Pune?"
 
Suggested reply:
Hi Aanya! Yes — Pune ships same-day from our Bombay warehouse.
Want me to drop the order link here?
 
[👍 Send]   [✏️ Edit]   [🗄️ Archive]

The rep glances at it, taps Send, and moves on. The Slack message includes buttons hooked back to the same N8N workflow — that's another two nodes if you want to fully close the loop.

What I'd watch

• The classification prompt drifts. Pin it. Re-tune it monthly with the 20 messages that confused Claude — not by re-reading the prompt cold.
• Don't auto-send the AI reply. Always require a human tap, at least for the first six months. The cost of one weird message landing in a customer's IG inbox is much higher than the cost of one extra tap.
• Log everything to the Sheet. Eventually you'll want to know which source ships the best conversion, and only the audit log can tell you.

That's the whole thing. Six nodes, one Claude call, one Slack channel — and your "we'll get back to you soon" problem is gone.