Step 1: Webhook Trigger

In n8n: New Workflow → Add first step → Webhook.

In the Webhook settings:

• HTTP Method: POST
• Path: kontakt-anfrage (your own path)
• Authentication: Header Auth -- generate a long random string as the secret
• Response Mode: Last Node (n8n responds once the workflow has completed)

n8n shows you the webhook URL, something like https://your-n8n-instance.com/webhook/kontakt-anfrage. You enter this URL in your contact form as the POST endpoint.

Test: in the n8n editor, click "Test workflow", then send a real inquiry through the form. If n8n receives the webhook, you'll see the data as JSON. If not, the path is wrong or the header-auth secret doesn't match.

Step 2: Classification with an LLM

After the Webhook node: + → AI → OpenAI → Message a Model. (If you're using Azure, there's a separate Azure OpenAI node with the same settings.)

Configuration:

You are a classification assistant for inbound
customer inquiries to a construction company.
Reply only with a JSON object of the form:
{"kategorie": "angebot" | "support" | "spam" | "sonstiges"}
No explanation, no prose.

Important: set Response Format to JSON Object. Otherwise the model will sometimes return Markdown-formatted answers and the workflow breaks.

Test with the inquiry from earlier. You should get back something like { "kategorie": "angebot" }.

Step 3: Routing via an IF Node

After the classification node: + → Flow → IF.

Condition: {{ $json.message.content.kategorie }} is not in spam, sonstiges.

The IF node splits the workflow into two paths. On the "false" path (spam/other), the workflow ends. Optional: a logging entry for spam, and moving "other" inquiries to a fallback folder.

On the "true" path, we continue with the next steps.

Step 4: CRM Lookup

This is where things get CRM-specific. We'll show it for a generic REST-API CRM (HubSpot, Pipedrive, and Zoho work similarly).

After the IF node (true path): + → HTTP Request.

In the next step we'll use the existence or absence of this contact.

Step 5: Extracting Structured Fields

Second LLM call. This time we extract the important information from the inquiry into a structured format.

+ → AI → OpenAI → Message a Model.

Extract the following fields from the customer inquiry as JSON:
{
  "leistungsart": "Which service is being requested?",
  "umfang": "Estimated scope in words, or open",
  "dringlichkeit": "hoch" | "normal" | "niedrig",
  "rueckfragen": ["Which information is missing for a
                   binding quote?"]
}
If information is missing: use null, do not guess.

Step 6: Drafting the Reply

Third LLM call. This is where the agent delivers the actual value: a reply draft that sounds like a human wrote it.

Two variants depending on whether the sender is an existing customer or not. Build it with two separate OpenAI nodes after a second IF node that checks whether the CRM lookup returned a hit.

Example prompt for existing customers:

You are a customer-service rep at [Company Name]. Write a
reply draft to an EXISTING CUSTOMER.
Tone: polite and direct, no marketing-speak.
Greet by first name (from CRM data).
Reference their inquiry concretely.
Ask the follow-up questions if any are listed in the JSON.
Close with a note on response time (24 hours).
Leave the signature open ("Your team at [Company Name]").

Variant for new customers: a slightly different greeting, a one-line introduction to the company in the first sentence, otherwise the same.

Step 7: Handing It Off as an Email Draft

Here's the critical point: the agent does not send the reply itself. It places it as a draft in the inbox of the responsible rep.

For Microsoft 365: + → Microsoft Outlook → Draft → Create. You'll need a one-time OAuth authentication against your Microsoft 365 account.

Optional: a Slack/Teams notification to the rep that a new draft is ready. + → Slack → Send Message.

What Goes Wrong Early On (and How to Fix It)

From experience -- the five most common issues during the first run:

JSON parse errors. LLMs love wrapping JSON in Markdown code blocks. Fix: Response Format → JSON Object in the OpenAI settings, plus an explicit "JSON only, no code blocks" in the prompt. Both -- one alone isn't enough.

Wrong webhook data structure. Depending on the form provider (Typeform vs. a custom HTML form vs. WordPress Contact Form 7), the JSON structure differs. In the n8n editor you can inspect the most recent webhook call and see exactly where each field lives. Only then write the references.

Hallucinated CRM data. When the CRM lookup turns up nothing (404), the LLM will sometimes still reach for "available data" and invent a first name. Fix: explicitly state in the system prompt "if no CRM record is available, use a neutral salutation" -- and check in code whether the CRM result actually exists.

Cost blowups. If your webhook gets flooded by a spam bot with 10,000 requests per hour, things get expensive fast. Fix: a rate-limit step before the first LLM call (n8n has no native one, but you can roll your own with Static Data) and a hard limit on the OpenAI side.

Email draft lands in the wrong mailbox. The default is the mailbox of the user n8n is authenticated as. If you want the draft in a shared mailbox, you have to configure the Outlook node accordingly.

When the Agent Is Ready for Production

Before you flip the switch, three tests:

Replay 20 historical inquiries. Take the last 20 inquiries from your inbox and run them through the workflow. Compare the generated drafts with what your reps actually wrote at the time. Match rate at 70 percent or above? Ready.

Test spam inquiries. Generate a handful of obvious spam texts ("Buy cheap viagra", "Investment opportunity") and check whether the workflow classifies correctly. If not: tighten the classification prompt.

Edge cases. A very long inquiry (with a 5,000-word code snippet inside), a very short one ("need help"), one in a foreign language. What happens? Do the steps fail cleanly or does the LLM produce garbage?

If all three tests come through cleanly, you can go live. But even then: the first 100 live inquiries get reviewed manually before the rep sends them out. That's the human-in-the-loop phase.

What You Can Extend Next

The agent is running. What comes next? Three sensible extensions:

Analyze attachments. If the inquiry includes a PDF (a spec sheet, a blueprint), the agent can read the document, pull out the key points, and weave them into the reply draft. With GPT-4o or Gemini 2.5, this works well.

Confidence score. Extend the classification prompt with a confidence value (0-100). On low confidence: special routing to a senior team member instead of an automatic draft.

Learning the tone. Collect the manual edits from the first 100 inquiries. Drop them into the prompt as few-shot examples. After three months, the agent sounds like your actual team.