How to Build AI Intent Routing in n8n

February 2, 2025 · 9 min read

The #1 question on the n8n community forum about AI agents: "How do I get my AI agent to follow system prompts and route to the right action?" (106 views and climbing).

The answer isn't better prompting. It's intent routing — classifying what the user wants first, then sending them down the right path. Here's how to build it in n8n.

Why Intent Routing Beats "One Big Prompt"

Most people try to build AI chatbots by cramming everything into one system prompt: "You are a helpful assistant that handles support, sales, feedback, and general questions."

This fails because:

• The AI gets confused — with too many responsibilities, it picks the wrong action
• You can't customize responses — support needs a ticket, sales needs a CRM update, feedback needs a database entry
• It's expensive — you send the entire context to the AI every time, even for simple routing
• Debugging is impossible — when something goes wrong, you don't know which "mode" the AI was in

Intent routing fixes all of this. One small, fast AI call classifies the intent. Then deterministic n8n logic routes to specialized handlers. Each handler has its own focused prompt and actions.

The Architecture

Five nodes. That's it. The key insight: the classifier is tiny. It only needs to output one word (SUPPORT, SALES, FEEDBACK, or GENERAL). This means:

• You can use a cheap, fast model (Claude Haiku or GPT-4o-mini)
• Set max_tokens: 10 — classification takes milliseconds
• The expensive model only runs in the handler that needs it

Step 1: The Classifier Prompt

This is the most important part. Your classifier prompt must be:

• Exhaustive — every possible intent is listed
• Mutually exclusive — no overlap between categories
• Constrained — output is ONLY the label, nothing else

The constraint "respond with ONLY the label" is critical. Without it, the AI will add explanations, which breaks your Switch node.

Step 2: The Code Node (Safety Net)

AI models don't always follow instructions perfectly. Add a Code node after the classifier to validate and normalize the output:

If the AI returns anything unexpected, it falls through to GENERAL. Fail safe, not fail open.

Step 3: The Switch Node

n8n's Switch node is perfect for intent routing. Set up one output per intent, with a fallback for anything unmatched:

• Output 0: intent === "SUPPORT" → Support Handler
• Output 1: intent === "SALES" → Sales Handler
• Output 2: intent === "FEEDBACK" → Feedback Handler
• Fallback: → General Handler

Step 4: Specialized Handlers

This is where intent routing really shines. Each handler can:

• Support: Create a Jira/Linear ticket, search knowledge base, send to Slack #support
• Sales: Update CRM, send to sales team, trigger nurture sequence
• Feedback: Save to database, tag sentiment, notify product team
• General: Use a conversational AI with a focused "helpful assistant" prompt

Each handler gets its own system prompt, its own tools, its own logic. No confusion. No prompt bloat. Clean separation of concerns.

Advanced: Adding Confidence Scores

Want to handle ambiguous messages? Modify the classifier to return a confidence level:

Then in your Code node, route LOW confidence messages to a human review queue instead of auto-handling them.

The JSON Body Gotcha

If you're calling the AI API via HTTP Request node, remember: set Specify Body to "String", not "JSON". n8n validates JSON before evaluating expressions, so {{ $json.message }} fails as invalid JSON. This trips up almost everyone.

When to Use This Pattern

• ✅ Customer support chatbots with multiple categories
• ✅ Slack/Discord bots that handle different commands
• ✅ Email triage (route to right department)
• ✅ Any system where user input could mean multiple different actions
• ❌ Single-purpose tools (just use one prompt)
• ❌ Free-form conversation (use a proper AI agent instead)