There's a gap between "I use ChatGPT at work" and "AI handles this automatically." Most people live on the chat side. They copy-paste outputs, manually trigger prompts, then move data between tools by hand. It works — but it's not automation. It's assisted manual labor.
n8n is the bridge. It's open-source workflow automation with 400+ integrations and native AI nodes, and it lets you turn your prompts into pipelines that run without you. More than 5,800 community templates exist for workflows ranging from content summarization to full lead qualification. You can self-host it for free or use their cloud.
This isn't a tour of n8n's features. It's two working pipelines, built step by step, with the actual prompts I use.
The mental model: prompts are function calls
If you've used a spreadsheet formula, you already understand the architecture. A formula takes inputs, does something, returns an output. Prompts work the same way — they take context (the input), process it through an LLM, and produce text (the output).
n8n pipelines are just chains of these function calls with data flowing between them. Output from step 1 becomes input to step 2. A condition node decides which path to take. An HTTP node sends the result somewhere.
The moment this clicked for me: I stopped thinking "what should I ask the AI?" and started thinking "what transformation do I need at this step?"
Pipeline 1: content monitoring and Slack digest
This pipeline monitors an RSS feed for new articles, classifies them for relevance, and posts a formatted summary to Slack — completely automatically.
Nodes you need:
- RSS Feed Trigger (or Webhook)
- AI node — summarize and classify
- IF node — filter by relevance
- AI node — format Slack message
- Slack node — post message
Step 1: RSS Feed Trigger
Set the feed URL and polling interval (I use 15 minutes for active sources, 1 hour for slower ones). Each new item triggers the workflow with these fields available: title, link, content, pubDate, creator.
Step 2: First AI node — classify and summarize
This is where most people underspecify. Don't just ask for a summary. Give the model a rubric.
You are a content analyst for a B2B SaaS company that sells project management software.
Incoming article:
Title: {{ $json.title }}
Content: {{ $json.content }}
Source: {{ $json.link }}
Tasks:
1. Relevance score (1-10): How relevant is this to our audience of project managers and team leads? Score above 7 means it's worth sharing.
2. Category: Choose one — [industry-news, competitor-update, thought-leadership, case-study, product-announcement, irrelevant]
3. Summary: 2-3 sentences, written for a busy project manager. Focus on the "so what."
4. Key takeaway: One sentence. What should a reader do or think differently after reading this?
Return as JSON:
{
"relevance_score": number,
"category": string,
"summary": string,
"key_takeaway": string
}
Setting the output format to JSON is critical here — it's what lets the next node parse and route cleanly. In n8n, enable "JSON" output mode on the AI node so the response is automatically parsed.
Step 3: IF node — filter by relevance
Condition: {{ $json.relevance_score }} >= 7
If true: continue to formatting. If false: stop. This is the power of AI classification — you don't need keyword rules or manual filters.
Step 4: Second AI node — format Slack message
Write a Slack message to share this article with the team. Keep it under 200 words.
Article details:
Title: {{ $json.title }}
Summary: {{ $json.summary }}
Key takeaway: {{ $json.key_takeaway }}
Category: {{ $json.category }}
Link: {{ $json.link }}
Format:
- Start with a relevant emoji based on category (📊 for case-study, 📰 for industry-news, ⚡ for competitor-update, etc.)
- Bold the title as a Slack link: *<link|title>*
- One paragraph with the summary and takeaway combined
- End with: "Worth 5 minutes →" followed by the link
Tone: Direct, no fluff. We're busy people sharing useful things.
Step 5: Slack node
Point it at your channel, use the message output from step 4. Done.
This workflow runs unattended. I set it up for 3 RSS feeds in about 45 minutes. It now surfaces 4-8 relevant articles per week to our Slack channel with zero manual work.
Pipeline 2: lead qualification and CRM entry
A form submission comes in. AI scores the lead, writes a personalized follow-up, logs it to your CRM, and routes high-value leads to a different email sequence. This one has real business impact.
Nodes you need:
- Webhook Trigger (form submission)
- AI node — qualify and score lead
- IF node — route by score
- AI node — write follow-up email (two versions: hot lead / standard)
- Gmail or SendGrid node — send email
- HubSpot / Airtable node — create CRM record
Step 1: Webhook Trigger
Your form posts to the n8n webhook URL. The payload comes in with fields like: name, company, role, use_case, team_size, budget, email.
Step 2: AI node — qualify and score
You are a sales qualification assistant for a B2B software company.
Lead information:
Name: {{ $json.name }}
Company: {{ $json.company }}
Role: {{ $json.role }}
Use case they described: {{ $json.use_case }}
Team size: {{ $json.team_size }}
Budget range: {{ $json.budget }}
Score this lead on two dimensions (1-10 each):
1. Fit score: How well do they match our ideal customer profile? (ICP: team size 10-500, technical role or manager with budget authority, clear use case)
2. Intent score: How strong is their buying signal based on what they described?
Also provide:
- Priority tier: "hot" (both scores ≥ 7), "warm" (at least one ≥ 6), "nurture" (everything else)
- One sentence explaining the priority rating
- 3 bullet points: key qualifying signals you noticed
Return as JSON:
{
"fit_score": number,
"intent_score": number,
"priority_tier": string,
"reasoning": string,
"signals": [string, string, string]
}
Step 3: IF node
{{ $json.priority_tier }} === "hot" routes to the hot-lead email. Everything else gets the standard follow-up.
Step 4: AI node — personalized email
For hot leads:
Write a personalized follow-up email from our sales team.
Lead info:
Name: {{ $json.name }}
Company: {{ $json.company }}
Role: {{ $json.role }}
Use case: {{ $json.use_case }}
Key signals our AI noticed: {{ $json.signals }}
Write as if from a senior account executive named Sarah. The email should:
- Reference their specific use case in the opening line (not generic "thanks for signing up")
- Mention one specific way our product addresses what they described
- Propose a 20-minute call with a Calendly link placeholder: [CALENDLY_LINK]
- Be under 150 words
- Sound like a human wrote it, not a template
Subject line: Keep it specific to their situation, under 50 characters.
Return as JSON: { "subject": string, "body": string }
The resulting emails don't read like automation. They read like someone actually looked at the form.
Key n8n prompting concepts
Pass dynamic context with double curly braces. {{ $json.fieldName }} pulls data from the previous node. {{ $('NodeName').item.json.field }} references any earlier node. Always preview what data is actually available before writing your prompt.
Keep AI nodes single-purpose. One AI node that classifies, scores, and writes copy will get confused and produce inconsistent output. One task per node, clear output format per node.
Always specify output format. "Return as JSON with these fields" is not optional — it's what makes routing and downstream use possible. Freeform text outputs create parsing problems in every subsequent step.
Error handling matters. AI nodes can fail or return malformed output. Add an Error Trigger workflow that catches failures and pings you in Slack. In production, I add a simple regex check after every AI node: if the output doesn't contain the expected keys, route to an error handler instead of continuing.
When to use the API instead of n8n
n8n is the right choice when: you need integrations between multiple tools, you want a visual workflow you can hand off to non-engineers, or you're building on top of existing triggers (forms, emails, webhooks).
Reach for direct API calls when: you need low latency (n8n adds overhead), you're processing high volume (n8n cloud has rate limits), or your logic is complex enough that building it visually is slower than writing code.
For most business automation — the category these pipelines fall into — n8n is faster to build, easier to maintain, and more accessible to your whole team.
The gap between "I use AI" and "AI works for me" closes one pipeline at a time. Start with the content monitoring workflow. Run it for a week. Then ask what else runs on manual effort that shouldn't.
If you want to go deeper on agent architecture before building more complex workflows, the AI workflows vs agents lesson covers when to use which pattern. For the conceptual foundation on what makes something an agent, start with what is an AI agent.
