What is n8n? How it works, self-hosting and your first automations
n8n is an AI-ready automation platform you can self-host. Learn how it works, what it costs, what its licence allows and how to build your first workflows.
Read
AI automation works best wherever a repetitive process trips over messy text: emails, forms, invoices, reviews or meeting notes. Below are 15 practical examples across sales, finance, marketing, customer service and operations. For each one we cover what triggers it, what the AI model does and where a human still makes the call.
Regular automation follows rules: “if the subject line contains ‘invoice’, save the attachment to this folder”. AI automation adds a language model that understands content. It can tell an invoice from a quote, pull the VAT number and due date out of a PDF, or draft a reply in your company’s tone. Everything else (fetching data, writing to the CRM, sending notifications) still runs on plain rules, because they’re cheaper, faster and predictable.
Most businesses are still early on. Eurostat’s 2025 figures show that 20.0% of EU enterprises with at least 10 employees used AI, up from 13.5% a year earlier, with Denmark (42.0%) at the top and Poland (8.4%) near the bottom. The most common use was analysing written language (11.8% of businesses), which is exactly what most of the examples below rely on.
| # | Process | Trigger | What the AI does | Where a human decides |
|---|---|---|---|---|
| 1 | Lead qualification | Website form | Category and priority | Contacting the lead |
| 2 | Lead enrichment and first reply | New CRM contact | Company summary, draft email | Sending the email |
| 3 | Quote follow-ups | Daily schedule | Personalised reminder | Approving the text |
| 4 | Sales call notes to CRM | New recording or transcript | Summary and next steps | Checking the note |
| 5 | Expense documents by email | Email with attachment | Reading data from PDF or photo | Resolving mismatches |
| 6 | E-invoices: cost category and approval | New structured e-invoice | Suggested cost category | Approving the cost |
| 7 | Payment reminders | Overdue invoice | Polite, personalised message | Escalation |
| 8 | Content repurposing | New blog post | Social and newsletter drafts | Editing and publishing |
| 9 | Google review replies | New Business Profile review | Sentiment, draft reply | Publishing the reply |
| 10 | SEO report with commentary | Weekly schedule | Commentary on changes | Deciding what to do |
| 11 | Inbox triage | New email | Classification, draft reply | Sending the reply |
| 12 | Knowledge-base assistant | Chat or WhatsApp message | Answers from company docs | Taking over hard cases |
| 13 | Order status | Customer question | Checks the order via a tool | Complaints and exceptions |
| 14 | Tasks from meetings | Meeting transcript | Task list with owners and dates | Confirming assignments |
| 15 | Client or employee onboarding | Signed contract or form | Personalised plan and messages | Granting access |
Your website form posts to n8n via a webhook. A Text Classifier node tags the enquiry (quote, partnership, complaint, spam) and sets a priority based on the content, budget and timeline. The lead lands in the CRM with every field filled in, and a sales rep only gets a Slack or Telegram ping for hot enquiries. Spam stops clogging the database.
When a contact is added, the workflow fetches the website behind their email domain and the model sums up in three sentences what the company does and what it might need. That feeds a draft reply saved in Gmail or Outlook. The rep tweaks two sentences and hits send instead of starting from a blank page.
Every morning a schedule pulls quotes from the CRM that went out a few days ago and haven’t had a response. The model writes a short reminder that refers to what was actually quoted, not a generic “just checking in”. It goes for approval, and one click sends it from the rep’s own address.
A call recording or transcript goes into n8n. The model pulls out the client’s needs, objections, budget, agreements and next step, and the workflow adds the note to the deal and creates a dated task. The rep just reviews it. Before you start recording calls, tell participants and make sure you have a lawful basis for it.
Receipts, some supplier invoices, credit notes and contracts still arrive as attachments. A Gmail Trigger picks up emails with attachments, Extract From File pulls the text, and an Information Extractor node returns data in the schema you define: document number, supplier, VAT number, net and gross amounts, currency and due date. Rules then check that the line items add up to the total. Mismatches go to accounts, and everything else goes to a spreadsheet or your accounting system.
More countries now require structured e-invoices. In Poland, for example, invoices have been issued through the national KSeF system since 2026: from 1 February for the largest taxpayers and from 1 April for everyone else, with the smallest businesses allowed to stay outside it until the end of 2026 (official KSeF timeline). A structured e-invoice is an XML file with a fixed format, so you don’t need AI to read it. AI earns its keep one step later: the model suggests a cost category or cost centre from the line items and supplier history, and the workflow routes the invoice to the right manager for approval.
Every day the workflow checks for overdue invoices. The first reminder is polite and personalised (a long-standing client gets a different tone from a new one), and the second is firmer. After a set number of days, a human takes over. AI only writes the text, while rules decide when and to whom a reminder goes.
When a blog post goes live, the workflow grabs the text and the model drafts a LinkedIn post, a shorter Facebook version, three lines for graphics and a newsletter paragraph. The drafts land in a sheet or on Slack for editing. We don’t recommend auto-publishing without review, because one off-key post costs more than the half hour you saved.
The Google Business Profile Trigger fires on a new review. The model reads the sentiment and drafts a reply: a short thank-you that mentions something specific for positive reviews, and a calm response offering to talk for negative ones. Negative reviews go straight to the owner, because a quick reply builds trust with local customers.
On Monday morning the workflow pulls clicks, impressions and positions for your key pages from the Google Search Console API, compares them with the previous week and emails a table with three lines of commentary from the model. The commentary flags what dropped and what to look at, while the decisions stay with you. If you want a strategy behind the numbers, see our SEO and GEO service.
Every email to your general address goes through a classifier: order, complaint, invoice, quote request, partnership, spam. The workflow labels it, routes it to the right person and, for common questions, drafts a reply from your approved answers. A human still sends it, but starts from finished text.
Connect a website chat or WhatsApp Business to an AI Agent node that searches your company documents (terms, price lists, manuals) stored in a vector database. The assistant answers only from those sources, and when it doesn’t know or the customer is upset, it hands the conversation to a person. If you serve customers in the EU, Article 50 of the AI Act has required since 2 August 2026 that people are told they’re talking to an AI system, unless that’s obvious.
“Where’s my parcel?” makes up a big share of many shops’ inboxes. An agent with an order-lookup tool (n8n has nodes such as WooCommerce and Shopify) verifies the customer by order number and email, checks the status and tracking number, and replies. Returns, complaints and exceptions go to the team. We design e-commerce automations alongside the online shops we build.
A team meeting transcript goes into n8n, and the model extracts decisions and tasks with an owner and a due date. The workflow creates the tasks in Asana, ClickUp or Trello and posts a summary to the team channel. Tasks without a named owner aren’t auto-assigned. They go to whoever ran the meeting.
A signed contract or a kick-off form starts the sequence: a shared folder, a card in your project tool, tool accounts and a calendar of meetings. The model drafts a personalised plan for the first few weeks and welcome emails based on the contract scope or the role. Granting system access stays with a human, or with an approval step in Slack.
The best first candidate ticks four boxes:
Don’t start with an agent that’s meant to “handle all of customer service”. Anthropic’s guide to building agents recommends finding the simplest solution possible and only adding complexity when you need it. In practice, 13 of the 15 examples above are ordinary workflows with AI steps, and only two (12 and 13) need an agent that chooses its own tools. We explain the difference between those and full agents in AI agents for business: how they differ from chatbots.
There are three parts to the bill:
Got a process that eats hours of your team’s week? See how we deliver AI automation with n8n and AI agents, or get in touch. We’ll start with one process and show you the results on your own data. New to n8n? Read what n8n is and how it works first.
Start with a frequent, text-based process where mistakes are cheap, such as inbox triage, qualifying form leads or reading data from expense documents. At first, let the AI produce drafts and suggestions that a person approves. Switch on automatic sending only once the results are consistently good.
In these examples AI takes over the repetitive part of the job: retyping data, sorting and writing first drafts. Decisions, difficult customer conversations and accountability for the outcome stay with people. In practice, teams shift time from data entry to work that needs judgement.
Not to read them, because a structured e-invoice such as one from Poland's KSeF is an XML file with a fixed format. AI is useful one step later, for example to suggest a cost category or cost centre and route the invoice for approval. For documents outside e-invoicing systems, such as receipts or some foreign invoices, AI also helps extract the data.
You pay for the platform (no licence fee if you self-host, or n8n Cloud from €20 a month billed annually), for AI model tokens, and for build and maintenance. Model costs depend on how much text you process and which model you choose. Build costs depend on the number of systems and edge cases, so they are quoted individually.
Send the model only the data it needs, mask personal data (for example with n8n's Guardrails node) and sign data processing agreements with your model and platform providers. Check where data is processed and add the process to your records of processing. Self-hosting n8n also gives you full control over execution logs.
Insights

n8n is an AI-ready automation platform you can self-host. Learn how it works, what it costs, what its licence allows and how to build your first workflows.
Read
n8n, Make or Zapier? We compare pricing (checked November 2026), how each meters usage, AI agents and self-hosting to help you pick the right platform.
Read
Explainer video for business: when it pays off, how long it should be, how to write the script, how production works and whether to choose 2D, 3D or screencast.
ReadWe connect your tools — CRM, email, spreadsheets, store, accounting — into n8n workflows where AI reads, classifies and replies. You get hours back, not another app to babysit.
We use essential cookies and, with your consent, analytics (Google Analytics, Hotjar) and marketing cookies to understand how you use the site and improve it. Details in our privacy policy.