n8n AI Workflow Examples: 5 Practical Automations to Build in 2026

How n8n AI Workflows Are Structured

Every n8n AI workflow follows the same node architecture. Understanding this structure once makes all five workflow patterns immediately readable.

The AI Agent node is the orchestrator. It receives a prompt (from the trigger data), reasons about what tools to call, executes tool calls, and produces a final response. The Chat Model node provides the LLM backend. Memory nodes (optional) give the agent access to prior conversation turns.

For Ollama local models, replace the OpenAI Chat Model node with the Ollama Chat Model node and set your Ollama URL. Use `llama3.3:8b` or `qwen2.5:14b` for reliable tool-calling. Models below 7B parameters frequently produce malformed tool-call JSON.

Workflow 1: AI Email Responder

This workflow monitors a Gmail inbox, classifies incoming emails, and drafts a response using GPT-5. A human approval step before sending makes it safe to deploy without worrying about the agent sending incorrect replies.

Node Layout

Gmail Trigger → Set (extract subject + body) → AI Agent → IF (confidence >= 0.8) → Gmail (send draft) → Slack (notify)

Gmail Trigger Configuration

Set the trigger to poll every 5 minutes for unread emails in a specific label (e.g., "AI-Handle"):

{
  "resource": "message",
  "operation": "getAll",
  "filters": {
    "labelIds": ["Label_AI-Handle"],
    "q": "is:unread"
  },
  "pollTimes": {
    "item": [{ "mode": "everyX", "value": 5, "unit": "minutes" }]
  }
}

AI Agent System Prompt

In the AI Agent node's "System Message" field:

You are an email assistant for a software company. Your job is to draft professional, concise email replies.

Given the subject and body of an incoming email, produce:
1. A classification: support_request | sales_inquiry | partnership | spam | other
2. A confidence score between 0 and 1 for your classification
3. A draft reply (if not spam)

Rules:
- Keep replies under 150 words
- Never promise specific timelines
- Sign off as "The Support Team"
- Format your response as JSON: {"classification": "...", "confidence": 0.0, "draft": "..."}

Set Node, Extract Email Fields

{
  "subject": "={{ $json.subject }}",
  "body": "={{ $json.snippet }}",
  "from": "={{ $json.from }}",
  "threadId": "={{ $json.threadId }}"
}

Output and Approval Gate

Add an IF node after the AI Agent: if `confidence >= 0.8`, send the draft directly via Gmail. If `confidence < 0.8`, post to Slack for human review:

{
  "conditions": {
    "number": [{
      "value1": "={{ JSON.parse($json.output).confidence }}",
      "operation": "largerEqual",
      "value2": 0.8
    }]
  }
}

Workflow 2: AI Web Scraper and Summariser

This workflow takes a topic or URL as input, searches the web for recent information, reads the full page content, and produces a structured summary. It is the foundation for competitive monitoring, news digests, and research automation.

Node Layout

Webhook (POST with topic) → AI Agent [tools: SerpAPI + HTTP Request] → Set (format output) → Respond to Webhook

Webhook Input Format

Send a POST request to trigger the workflow:

curl -X POST https://your-n8n.com/webhook/research \
  -H "Content-Type: application/json" \
  -d '{"topic": "Anthropic Claude 4 release news", "max_sources": 3}'

AI Agent Configuration

Add two tools to the AI Agent node:

1. SerpAPI tool (built-in), set your SerpAPI key in n8n credentials 2. HTTP Request tool, set to allow GET requests (the agent uses this to read full page content from URLs in search results)

System message:

You are a research assistant. Given a topic, you will:
1. Search the web for the 3 most relevant and recent results
2. Read the full content of each result using the HTTP Request tool
3. Produce a structured summary with: key findings, source URLs, and a one-sentence verdict

Be specific. Include dates, figures, and named entities. Do not speculate beyond what the sources say.

Handling JavaScript-Heavy Pages

Some pages do not return content via a simple HTTP GET because they require JavaScript rendering. Add a filter in the agent's instructions:

If the HTTP Request tool returns less than 200 characters of readable text, skip that URL and move to the next search result.

This prevents the agent from citing JavaScript error pages as sources.

Workflow 3: AI Chatbot with Conversation Memory

This workflow builds a persistent chatbot that remembers previous messages within a session. It uses n8n's built-in Postgres Chat Memory node to store conversation history in a database, so memory survives workflow restarts and server reboots.

Node Layout

Chat Trigger → AI Agent [memory: Postgres Chat Memory] → Chat Response

Why Postgres Memory over Buffer Memory

Buffer Memory stores conversation history in RAM. When n8n restarts or the workflow is deactivated, history is lost. Postgres Chat Memory persists to a PostgreSQL database and survives restarts. For any customer-facing chatbot, use Postgres memory.

Database Setup

Create a dedicated table in your PostgreSQL database:

CREATE TABLE IF NOT EXISTS n8n_chat_histories (
  id SERIAL PRIMARY KEY,
  session_id VARCHAR(255) NOT NULL,
  message JSONB NOT NULL,
  created_at TIMESTAMP DEFAULT NOW()
);

CREATE INDEX ON n8n_chat_histories(session_id);

n8n's Postgres Chat Memory node creates this table automatically if you enable "Create Table If Not Exists" in the node settings. The SQL above is for manual creation or inspection.

Postgres Chat Memory Node Settings

{
  "sessionIdType": "customKey",
  "sessionKey": "={{ $json.sessionId }}",
  "contextWindowLength": 10
}

Set `contextWindowLength` to 10 (the last 10 message pairs). Larger windows increase LLM token usage on every request. For GPT-5.2 at 10 message pairs, a typical conversation uses 2,000-5,000 tokens per turn, costing $0.01-0.05.

Chat Trigger Configuration

Enable the n8n Chat Trigger node with "Allow File Uploads" off (for text-only chatbots). The session ID is generated automatically per browser session, pass it to the memory node's sessionKey.

Workflow 4: Document Analysis and Data Extraction

This workflow accepts a PDF or document file, extracts its text, and uses an LLM to answer questions about it or extract structured data. Common use cases: invoice processing, contract review, research paper summarisation, and compliance document checking.

Node Layout

Webhook (file upload) → Extract from File → Split into Chunks → AI Agent → Respond to Webhook

Extract from File Node

{
  "operation": "pdf",
  "options": {
    "joinPages": true
  }
}

This extracts all text from the PDF into a single string. For large documents (50+ pages), enable "Join Pages" and expect the output to be 50,000-200,000 characters.

Chunking for Large Documents

GPT-5.2 has a 128,000 token context window. For documents under 80,000 words, you can pass the full text to the agent. For larger documents, add a "Recursive Character Text Splitter" node before the agent:

{
  "chunkSize": 2000,
  "chunkOverlap": 200,
  "separator": "\n\n"
}

Then loop through chunks with a n8n "Loop Over Items" node and aggregate results.

Extraction System Prompt (Invoice Example)

Extract the following fields from this invoice document and return them as JSON:
- invoice_number (string)
- invoice_date (ISO 8601 date)
- vendor_name (string)
- total_amount (number, in the invoice currency)
- currency (3-letter ISO code)
- line_items (array of {description, quantity, unit_price, total})

If a field is not found, set its value to null. Return only the JSON object, no explanation.

Output Validation

Add a Code node after the AI Agent to validate the extracted JSON:

const output = JSON.parse($input.first().json.output);
const required = ['invoice_number', 'invoice_date', 'vendor_name', 'total_amount'];
const missing = required.filter(f => output[f] === null);

return [{
  json: {
    ...output,
    extraction_complete: missing.length === 0,
    missing_fields: missing
  }
}];

Workflow 5: Social Media Content Agent

This workflow takes a blog post URL or topic, researches it, and generates platform-specific content for LinkedIn, X (Twitter), and Reddit. A schedule trigger runs it daily from a content queue stored in a Google Sheet.

Node Layout

Schedule Trigger → Google Sheets (get next queued item) → AI Agent [tools: HTTP Request] → Set (format per platform) → 3x HTTP Request (post to each platform API)

Google Sheets Queue Structure

Create a sheet with columns: `status`, `topic`, `source_url`, `linkedin_posted`, `twitter_posted`, `reddit_posted`, `created_at`.

The workflow reads the first row where `status = "queued"`, processes it, then updates the row to `status = "done"`.

AI Agent System Prompt

You are a social media content writer. Given a topic and optional source URL:

1. If a URL is provided, read its content using the HTTP Request tool
2. Generate three versions of content:

LINKEDIN: Professional tone, 150-200 words, no hashtags in body, 3 hashtags at end, include a question to drive comments
TWITTER: 240 characters maximum, conversational tone, include 2 hashtags inline
REDDIT: Authentic tone, no marketing language, frame as a question or discussion prompt, 100-150 words

Return as JSON: {"linkedin": "...", "twitter": "...", "reddit": "..."}

Posting to Platform APIs

For LinkedIn (OAuth 2.0 required, set up LinkedIn credentials in n8n):

{
  "method": "POST",
  "url": "https://api.linkedin.com/v2/ugcPosts",
  "headers": {
    "Authorization": "Bearer {{ $credentials.linkedInOAuth2Api.accessToken }}",
    "Content-Type": "application/json"
  },
  "body": {
    "author": "urn:li:person:YOUR_PERSON_ID",
    "lifecycleState": "PUBLISHED",
    "specificContent": {
      "com.linkedin.ugc.ShareContent": {
        "shareCommentary": { "text": "={{ $json.linkedin }}" },
        "shareMediaCategory": "NONE"
      }
    },
    "visibility": { "com.linkedin.ugc.MemberNetworkVisibility": "PUBLIC" }
  }
}

Troubleshooting

Frequently Asked Questions

const result = JSON.parse($input.first().json.output);
return [{ json: result }];