Lesson 4

Workflow Automation

15 min

Not every AI automation needs to be a Python service. For connecting SaaS tools, processing form submissions, or routing support tickets, a visual workflow tool like n8n is often faster to build and easier to maintain than custom code. This lesson builds a production-ready LLM workflow with n8n as the orchestrator.

n8n Architecture for AI Workflows

n8n executes workflows as directed graphs of nodes. Each node receives input JSON, processes it, and passes output JSON to the next node.

| Node type | Purpose | LLM use case | |---|---|---| | Webhook | Receive HTTP requests | Entry point for user submissions | | HTTP Request | Call external APIs | Call OpenAI / Anthropic API | | Code | Run JavaScript | Parse LLM output, custom logic | | IF | Conditional branching | Route based on LLM classification | | Wait | Pause for approval | Human-in-the-loop gate | | Send Email / Slack | Notifications | Alert on LLM output or errors |

Defining the Workflow

A document triage workflow:

Webhook (POST /triage)
    │
    ▼
HTTP Request → Anthropic API (classify document)
    │
    ▼
Code Node (parse classification JSON)
    │
    ▼
IF Node: category == "legal"?
    │yes                   │no
    ▼                      ▼
Wait (human approval)   Auto-process Node
    │
    ▼
Send Slack (notify reviewer)
    │
    ▼
HTTP Request → Internal API (route document)

The LLM Classification Node

javascript

// n8n Code Node: call Anthropic and parse response
const anthropicApiKey = $env.ANTHROPIC_API_KEY;
const document = $input.first().json.text;

const response = await $http.request({
  method: 'POST',
  url: 'https://api.anthropic.com/v1/messages',
  headers: {
    'x-api-key': anthropicApiKey,
    'anthropic-version': '2023-06-01',
    'content-type': 'application/json',
  },
  body: JSON.stringify({
    model: 'claude-haiku-4-5',
    max_tokens: 100,
    messages: [{
      role: 'user',
      content: `Classify this document. Return ONLY valid JSON: {"category": "legal|technical|hr|general", "confidence": 0.0-1.0, "reason": "one sentence"}

Document: ${document.substring(0, 1000)}`
    }]
  })
});

const text = response.content[0].text;
// Strip markdown code fences if present
const clean = text.replace(/```json\n?|\n?```/g, '').trim();
const parsed = JSON.parse(clean);

return [{ json: { ...parsed, original_text: document } }];

Human-in-the-Loop Approval Gate

javascript

// n8n Wait Node configuration
// Pause execution until an approval webhook is received

// Approval webhook URL: POST /webhook/approve/{executionId}
// Expected body: { "approved": true, "reviewer": "alice@company.com", "note": "..." }

// Code node after Wait:
const approval = $input.first().json;
if (!approval.approved) {
  // Route to rejection workflow
  return [{ json: { status: 'rejected', reason: approval.note } }];
}
return [{ json: { status: 'approved', reviewer: approval.reviewer } }];

Webhook Trigger with Validation

python

# Python-side: send a document for triage
import httpx

def submit_for_triage(document_text: str, document_id: str) -> dict:
    resp = httpx.post(
        "https://n8n.company.com/webhook/triage",
        json={
            "document_id": document_id,
            "text":        document_text,
            "submitted_at": "2024-03-27T10:00:00Z",
        },
        headers={"X-Webhook-Secret": WEBHOOK_SECRET},
        timeout=10.0,
    )
    resp.raise_for_status()
    return resp.json()

Error Handling in Workflows

| Failure point | n8n mitigation | |---|---| | LLM API timeout | Error workflow branch + retry | | Invalid JSON from LLM | Code node try/catch + fallback category | | Approval webhook not received | Wait node timeout → escalation | | Downstream API failure | Error node → Slack alert |

Always add an "Error Workflow" in n8n settings — it fires on any unhandled exception and prevents silent failures.

Summary

n8n orchestrates LLM workflows as visual node graphs, reducing boilerplate for SaaS integrations.
Use HTTP Request nodes to call LLM APIs and Code nodes to parse and transform the output.
Add human-in-the-loop gates via n8n's Wait node for sensitive document categories or high-stakes actions.
Always validate LLM JSON output in a Code node with try/catch and provide a fallback so a malformed response does not break the workflow.
Configure n8n error workflows for every production workflow to catch and alert on silent failures.

Multi-Step Agents Guardrails and Safety