n8n business automation: when it fits and how to run it well

Part 01

What n8n is, and what the license allows

n8n is a workflow automation platform. You build a workflow on a canvas: a trigger (a new email, a webhook, a schedule), then nodes that read, transform and write data across your tools. When a visual node is not enough, you drop in a Code node and write JavaScript or Python. Its GitHub README lists more than 1,500 integrations.

The license matters before you commit. n8n calls itself fair-code and distributes its source under the Sustainable Use License. It lets you "use or modify the software only for your own internal business purposes or for non-commercial or personal use." So a company can run n8n for its own processes at no license cost. Selling n8n itself to others, for example as a hosted product, falls outside it. Files marked .ee sit under a separate enterprise license.

You can run it two ways. n8n's own docs describe n8n Cloud (paid plans after a free trial) and self-hosting, where the free Community edition has "almost the complete feature set". Paid Business and Enterprise licenses add SSO, environments and projects. On pricing, n8n Cloud counts executions, and its pricing page defines one as "a single run of your entire workflow", however many steps it has.

Part 02

n8n vs Zapier vs Make: a fair comparison

All three tools connect apps and run logic between them. They differ on where they run, how they bill and how far you can go with code. The table uses the vendors' own documentation, checked on 10 October 2026. Prices change often, so we compare billing units rather than amounts.

Sources: n8n security page, n8n pricing, Zapier data privacy, Zapier task usage, Make organizations, Make pricing, Make Code.

The billing units change the maths. A Zapier flow with eight actions uses eight tasks per run. The same flow in n8n is one execution. For a few hundred runs a month the difference is small. For a fund that enriches every inbound deck, or an accounting team that processes every supplier invoice, it decides the bill.

Zapier is the right call when a marketing team needs ten small connections next week and nobody wants to run a server. Make is strong for visual scenarios with many branches. n8n wins when the data should stay on your infrastructure, when you need real code next to the visual steps, or when AI agents must pause for a human.

Part 03

Why self-hosting in the EU matters for GDPR

Every automation copies personal data: names in a CRM, IBANs on invoices, health details in a tenant's email to the property manager. Under the GDPR, the platform that runs the workflow is a processor, and moving that data outside the EU is a transfer governed by Chapter V of the regulation.

Zapier states that it "hosts data in AWS servers located in the United States" and covers EU data with Standard Contractual Clauses. That is legal, but it puts a transfer assessment on your desk. Make lets you pick an EU data center, and its help center warns that "you cannot change the location of the data center after you create the organization." n8n Cloud runs on Microsoft Azure with data "hosted in the European Union", according to its security page.

Self-hosting goes one step further. The workflows, the execution history and the encrypted credentials sit on a server you choose, in Frankfurt, Zürich or Madrid. Your DPO can see where everything is.

Part 04

What good n8n work looks like in production

A demo workflow runs once, on clean data, while someone watches. A production workflow runs at 3 a.m. on a PDF scanned sideways, while the CRM API times out. In September 2026 I built a workflow live at the n8n community meetup in Madrid, and most of the questions afterwards were about this part: what happens when something breaks.

1. An error workflow on every workflown8n lets you attach an error workflow that starts with the Error Trigger node and receives the failed execution, its URL and the last node that ran. One shared error workflow can serve all of them. It posts to the channel the owner actually reads.
2. Retries, then idempotencyTurn on Retry On Fail for nodes that call flaky APIs. Retries are only safe if a second run cannot create a second record, so every write uses a key: supplier VAT number plus invoice number, or the CRM's external ID. Check the key before writing.
3. Credentials in the credential storeNever paste an API key into a node parameter or a Code node. n8n encrypts stored credentials with an instance key; back that key up separately. Create credentials under company accounts, not a freelancer's personal login.
4. Workflows in GitOn Business and Enterprise plans, n8n has source control and environments built in. On the Community edition, export workflows as JSON with the CLI or API and commit them on a schedule. Either way, a change has a diff and a way back.
5. Monitoring that notices silenceError alerts catch failures. They miss the workflow that stopped triggering. Add a small scheduled check: if the invoice workflow has no successful run since yesterday morning, alert.
6. One Code node to fetch many sourcesFive HTTP Request nodes feeding one Merge node look tidy, but items multiply when one branch returns several rows, and an empty branch changes what the merge emits. One Code node that calls the five sources, catches each failure and returns one clean item is easier to read, test and retry.
7. Human approval where errors cost moneyUse a Wait node or a send-and-wait message for payments, deletions and anything that goes to a client. The workflow pauses until someone approves.
8. Log every AI decisionStore the input reference, the model, the prompt version, the output, the confidence and who approved. When a supplier asks why an invoice was rejected, you answer in a minute.

The canvas shows the happy path. Production work lives in the error branch.

Alpgency

Part 05

AI agents inside n8n, and when to move logic into code

Most of the new demand for automation now involves a language model. Eurostat counts 20.0% of EU enterprises with 10 or more employees using AI in 2025, up from 8.1% in 2023.

n8n's AI Agent node connects a chat model to one or more tools, and "the agent decides which tools to call to complete a task." A tool can be a CRM lookup, a database query, another workflow or an MCP server. You can also require human approval before a specific tool runs: the workflow pauses, a reviewer sees the tool and its parameters, then approves or denies.

That makes n8n a good home for a first agent. A property manager's agent reads a tenant email, looks up the lease and the open tickets, and drafts a reply. Sending the reply is the tool that needs approval. Reading is not.

Keep the agent narrow. Give it the three tools the task needs, not twenty. Put the deterministic rules (VAT format, totals, duplicate checks) in plain nodes before and after the model, where they always run the same way.

Move logic out of n8n into code when one of these is true:

• The logic needs unit tests, and a Code node has grown past one screen.
• Several workflows, or an app, need the same rules. Put them in one small service that n8n calls over HTTP.
• Volume or memory grows. n8n's docs list 320 MiB of RAM for a Starter Cloud instance and advise splitting large jobs into batches and sub-workflows.
• You need an evaluation set run on every prompt change, with results you can compare.

n8n then stays in charge of triggers, approvals and glue. The part that needs engineering discipline lives in a repository with tests.

Part 06

n8n business automation in practice: three examples and a checklist

1. Invoice intake to the ERP

An email trigger watches the accounting inbox. An AI step extracts supplier, VAT number, invoice number, lines and total from the PDF. Plain nodes check that the lines add up to the total, that the VAT number has a valid format, and that the supplier-plus-invoice key is new. Clean invoices become a draft bill in the ERP. Anything else goes to the accountant in Slack with the PDF and the reason. Every outcome lands in a log table.

2. Lead enrichment to the CRM

A web form posts to a webhook. One Code node queries the public company register (Zefix in Switzerland, the Handelsregister in Germany) and the company's website, and returns one item with legal name, size signals and sector. A scoring step rates the lead against your criteria and writes the reasoning next to the score. The CRM gets a contact, a company and a task for the owner. If the register is down, the lead still lands in the CRM, marked "not enriched", and a retry runs an hour later.

3. The Monday report

A schedule fires on Monday at 7:00. A Code node pulls last week's pipeline from the CRM, invoiced revenue from accounting and traffic from analytics. Plain checks compare each figure with the week before and with the source totals. A model writes three sentences of commentary from the checked numbers only. If a figure fails a check, the report still goes out, with the alert at the top instead of a wrong number.

FAQ

Frequently asked questions