A deep, n8n-specific interactive guide — 13 chapters covering nodes, triggers, and JSON data flow, real CRM/invoicing/support-routing workflow builds, the AI Agent node with LLMs and RAG, and production hardening (error handling, scheduling, rate limiting, queue-mode scaling, and security).
Why automation matters and why n8n (fair-code philosophy, vs. Zapier/Make/Power Automate). Navigating the interface (canvas, node browser, execution history). Core concepts (nodes, JSON data flow, expressions). Building your first workflow (webhook → Sheets → email → Slack). CRM automation (lead capture, scoring, dedup). Invoice and payment automation (order-to-invoice, accounting sync, reminders). Customer support ticket routing with AI sentiment analysis. The n8n AI Agent node (LLMs, RAG, tool-calling). Error handling and retry logic (backoff, circuit breakers, dead letter queues). Scheduling, batching, and rate limiting. Performance, scaling, and security for production. Closing: monetization paths and modular design.
This reader goes deep specifically on n8n — the actual node types, JSON data model, expression syntax, and execution architecture (regular vs. queue mode) rather than a high-level survey of the no-code automation landscape. Every workflow shown (CRM lead scoring, order-to-invoice, AI-powered ticket routing) is built from the real chapter content of The n8n Automation Handbook, including exact node sequences and field mappings.
A plain LLM node call takes a prompt and returns a response — useful for extraction, classification, and summarization. The AI Agent node goes further: it runs a reasoning loop where the model chooses from a defined set of tools (database lookups, APIs, document search, email), executes one, observes the result, and decides the next action — without you scripting every step. Combined with RAG (retrieval-augmented generation against your own documents via a vector database), this lets n8n workflows answer questions and take actions grounded in your actual business data, not just the model's training knowledge.
The handbook treats error handling as first-class architecture: Error Trigger nodes for centralized failure handling, exponential backoff with jitter for retries, circuit breakers to stop hammering a failing service, and dead letter queues for items that need human review. On the scaling side: queue execution mode (Redis-backed) lets you add worker nodes horizontally instead of blocking the main instance, SplitInBatches and rate-limiting strategies (token bucket, fixed delay) respect external API limits, and credential encryption plus webhook authentication close the most common security gaps.
The AI Agent node runs an agentic loop rather than a single request/response. You give it a language model, a set of tools (functions or APIs it is allowed to call — for example lookup_customer, lookup_order, fetch_documentation, send_email), and a goal. The agent reasons about the goal, decides which tool to call first, executes it, observes the result, and decides whether to call another tool or return a final answer — iterating without you scripting each step in advance. A plain LLM node, by contrast, takes one prompt and returns one response; it cannot decide to look something up first. Use plain LLM calls for classification, extraction, and summarization; use the AI Agent node for multi-step tasks where the right sequence of actions depends on what earlier steps discover.
The pattern from The n8n Automation Handbook: a Typeform Trigger (or generic Webhook node for embedded forms) captures the submission in real time. A Set node standardizes the fields (firstName, lastName, email, phone, company). Before creating a CRM record, a Search node checks whether a contact with that email already exists (upsert pattern) to prevent duplicates. Chained IF nodes score the lead on company size, industry fit, budget, and timeline, tallying a total score with a Set node. Based on the score, the workflow routes to immediate Slack notification for high-value leads, a 24-hour follow-up queue for medium leads, or an automated nurture sequence for lower-scoring ones — while also creating/updating the CRM record via the HubSpot, Salesforce, or Pipedrive node.
An order event (Shopify "Fulfillment completed," WooCommerce "Order created," or a Stripe "Payment succeeded" webhook) triggers the workflow. Order details are extracted and mapped to an invoice created directly in QuickBooks, Xero, or FreshBooks via their API, with payment status set to Unpaid, Partially Paid, or Paid based on what was already collected at checkout. A Stripe Trigger node listening for payment_intent.succeeded, payment_intent.payment_failed, and charge.refunded events keeps the accounting system synced as money moves. A separate daily Cron-triggered workflow queries all unpaid invoices and sends escalating reminders at day 7, 14, 21, and 28, tracking a reminder count per invoice to avoid duplicate messages.
A Gmail Trigger node watches the support inbox. Incoming email text is normalized (lowercase, trimmed, length-capped), then passed to an OpenAI node with a structured prompt asking for JSON output: sentiment (positive/neutral/negative), intent (technical_support/billing/feature_request/complaint/feedback/other), urgency_keywords, and a one-sentence summary. IF/Switch nodes combine this AI analysis with rule-based logic to assign a category and a priority (Critical/High/Medium/Low, each with its own SLA), then a Set node determines the right assignee or team queue. The ticket is created in Zendesk, Freshdesk, or Intercom with the AI summary attached, and a separate hourly workflow monitors SLA compliance and escalates tickets at risk of breach.
Regular execution mode runs workflows directly on the main n8n instance — simple to set up, but a slow workflow blocks server resources, and if the server crashes, in-flight executions are lost. Queue mode uses Redis to enqueue triggered workflows for separate worker processes: the main instance never blocks, failed jobs are automatically re-queued, and you scale horizontally by adding more workers rather than upgrading a single server. Queue mode requires slightly more setup (Redis, worker processes, environment variables like EXECUTIONS_MODE=queue) and introduces a small enqueue/dequeue delay, but for any production system handling meaningful volume, the resilience and scalability it provides make it the recommended default over regular mode.