# n8n for Business: Workflow Automation and AI Agents

> n8n is a workflow automation platform that connects business systems through nodes and runs AI agents through its AI Agent node. It fits when you automate internal processes and want to host the data yourself; it is the wrong choice for reselling as a service or without in-house operations skills.

Canonical URL: https://www.ruebstudio.de/en/guides/ai-tools-and-platforms/n8n-for-business  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-10-23

Many IT teams reach for n8n when they need to connect an ERP, a CRM, a shared mailbox and a ticketing tool without writing a custom integration for each pair. Since the AI Agent node became part of the core product, the same platform also runs AI agents. This guide explains how n8n is built, which licensing and hosting questions to settle before you adopt it, and where it stops being the right tool. Checked in October 2026; the latest stable release on [GitHub](https://github.com/n8n-io/n8n/releases) at the time of writing is n8n 2.44.6, published on 22 October 2026.

## What is n8n and what do companies use it for?

n8n lets you assemble processes as workflows on a visual canvas: a trigger starts the run, and nodes read, transform and write data in connected systems. Its [repository](https://github.com/n8n-io/n8n) describes it as a "fair-code platform to build and deploy AI agents and workflows". The focus is integration across many systems rather than a single application.

Common business uses share three traits: they repeat often, follow clear rules and touch several systems.

- **Record sync:** a web form creates a lead in the CRM, and the customer number from the ERP is written back.
- **Document handling:** a supplier's delivery note arrives as a PDF attachment, its line items are extracted and matched against the purchase order, and mismatches open a ticket for procurement.
- **Approvals and alerts:** a quote above a threshold needs sign-off; n8n asks the approver in Teams or by email and updates the status afterwards.
- **AI-assisted steps:** a language model classifies and summarises incoming support requests and routes them to the right queue.

What sets n8n apart from pure cloud automation services is the hosting model. You can run it on your own servers or in your own cloud account, which is often the deciding factor for organisations handling sensitive data.

## How does n8n work? The building blocks

A workflow is a trigger followed by a chain of nodes that pass data along as items. Triggers fire on schedules, webhooks or events in connected systems. When no ready-made node exists, the HTTP Request node calls any API, and custom logic goes into JavaScript or Python code steps.

The main building blocks:

- **Trigger nodes:** schedules, webhooks, incoming mail, changes in a database or SaaS app.
- **App nodes:** prebuilt connectors for widely used systems such as Microsoft 365, Google Workspace, Slack and databases. The [integration docs](https://docs.n8n.io/integrations/readme.md) speak of "hundreds of nodes", and community-built nodes can be installed on top.
- **Flow control:** branches, loops, merges and error paths.
- **Code:** the [README](https://github.com/n8n-io/n8n) lists JavaScript, Python and npm packages alongside the visual editor.
- **Credentials:** stored encrypted at instance level and referenced by nodes, so secrets do not live inside the workflow definition.

Workflows, credentials and execution history sit in a database. According to the [database docs](https://docs.n8n.io/hosting/configuration/supported-databases-settings/), SQLite is the default and PostgreSQL is the supported alternative.

### AI agents in n8n

n8n implements LangChain concepts as cluster nodes: a root node defines the main function and sub-nodes extend it. The AI Agent node is the root for agents. As the [LangChain overview](https://docs.n8n.io/advanced-ai/langchain/overview/) explains, you attach a chat model, memory, tools and, optionally, an output parser to it.

- **Chat model:** sub-nodes for Anthropic, OpenAI, Ollama and others. Ollama connects models you run on your own hardware.
- **Memory:** conversation history in Postgres, Redis or a simple window buffer. Memory attaches only to the AI Agent node, not to plain chains.
- **Tools:** actions the model may call, such as another n8n workflow or a code step.
- **Retrieval:** embeddings, retrievers, document loaders and text splitters for [retrieval-augmented generation (RAG)](https://www.ruebstudio.de/en/glossary/retrieval-augmented-generation-rag). Vector stores such as Qdrant or Pinecone are root nodes in their own right.

Two features matter most for production use. [Human-in-the-loop for tools](https://docs.n8n.io/build/integrate-ai/ai-examples/human-in-the-loop-for-tools.md) pauses the workflow before an agent runs a risky tool, for example sending an email or deleting a record, until a reviewer approves or denies it in the n8n chat, Slack, Microsoft Teams, Outlook or another supported channel. [Evaluations](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/understand-why-to-test.md) run a dataset of test cases through a workflow so you can compare outputs before switching prompts or models.

### n8n and MCP

n8n speaks the [Model Context Protocol (MCP)](https://www.ruebstudio.de/en/glossary/model-context-protocol-mcp) in both directions. As a client, agents in n8n can call tools exposed by external MCP servers. As a server, the [MCP Server Trigger](https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-langchain.mcptrigger.md) publishes attached tools and workflows at a URL over Server-Sent Events or streamable HTTP; stdio is not supported. Bearer or header authentication is optional, so switch it on deliberately.

## Deployment options: cloud, self-hosted and editions

n8n comes as a managed cloud service and as self-hosted software. The [vendor's overview](https://docs.n8n.io/choose-n8n/) lists Starter, Pro and Enterprise cloud plans, and Community, Registered Community, Business and Enterprise editions for self-hosting. The choice affects where data lives, which features you get and how much operational work lands on your team.

**n8n Cloud.** The vendor runs the instance, and by its own description no technical expertise is required. Setup and maintenance are handled by the vendor. In exchange, your data and credentials sit with a third party, so check the hosting location and the data processing agreement. According to the database docs, the Starter and Pro cloud plans run on SQLite; PostgreSQL is reserved for the Enterprise scaling plan.

**Self-hosted.** You install n8n with Docker or npm on your own infrastructure. The vendor is explicit that infrastructure and maintenance are "your responsibility". In return, data, credentials and execution logs stay inside your network, and you can pair n8n with models you host yourself; [hardware sizing for local LLMs](https://www.ruebstudio.de/en/guides/local-ai/hardware-local-llm-gpu-server-sizing) explains what that takes. This is the route for projects where data must not leave the organisation, a topic covered in more depth under [sovereign AI](https://www.ruebstudio.de/en/sovereign-ai).

**Community or licensed edition.** The free Community edition covers workflows, nodes and AI agents. According to the [feature list](https://docs.n8n.io/hosting/community-edition-features/), it lacks among other things:

- SSO via SAML and LDAP
- projects and sharing of workflows and credentials
- Git-based version control and separate environments
- external secrets (connecting a secrets manager)
- log streaming
- multi-main mode

The Registered Community edition, unlocked by registering an email address, adds folders, "debug in editor" and custom execution data. Once several teams build workflows with separate permissions, or you need a promotion path from test to production, a licensed edition is hard to avoid.

Deeper comparisons of hosting models and of n8n against Make and Zapier will appear in this guides section; this overview page lists them automatically once they are published.

## Licensing: what the Sustainable Use License allows

n8n is not open source by the OSI definition. It ships under the Sustainable Use License, which n8n calls fair-code. The [licence text](https://github.com/n8n-io/n8n/blob/master/LICENSE.md) permits use and modification "only for your own internal business purposes or for non-commercial or personal use". For most internal automation that is fine.

Three situations need a closer look:

1. **Charging others for it:** you may only distribute or provide the software to others free of charge and for non-commercial purposes. Shipping n8n inside a product you sell is not covered by this licence.
2. **Running it for third parties:** whether an instance serving group companies, customers or partners counts as "internal business purposes" is a matter of interpretation. Get clarity in writing before you build on it.
3. **Enterprise code:** files with ".ee" in their file or directory name are not covered by the Sustainable Use License and require a valid n8n Enterprise License. Without that licence you may not use this code.

This section is orientation, not legal advice. Have the licence reviewed by counsel for any commercial scenario.

## Prerequisites: data, systems and organisation

Before n8n goes live you need three things: systems with reachable interfaces, a person or team that runs the instance, and rules for who may build and approve workflows. Miss one and you end up with undocumented automations nobody can maintain.

**Systems and interfaces.** For every process, confirm that each system offers an API, webhooks or at least database access. Older ERP systems without an API can often only be reached through file exports or middleware, which works but breaks more easily.

**Operations.** Self-hosting needs a server or container host, database backups, TLS, an update routine and monitoring. SQLite is enough for a single instance. When load grows, [queue mode](https://docs.n8n.io/hosting/scaling/queue-mode/) spreads executions across worker processes; that setup requires Redis as the message broker and PostgreSQL as the database, and SQLite is not supported there.

**Privacy.** A self-hosted instance sends telemetry to the vendor by default, including execution counts and an instance pulse every six hours. The [telemetry docs](https://docs.n8n.io/deploy/host-n8n/configure-n8n/security/control-telemetry.md) explain how to switch it off with `N8N_DIAGNOSTICS_ENABLED=false`, and version checks with `N8N_VERSION_NOTIFICATIONS_ENABLED=false`. Configured this way and combined with locally hosted models, n8n can be operated in line with the GDPR; whether a specific workflow is depends on the data, the purpose and your contracts.

**Organisation.** Decide who builds workflows, who approves them, and who gets paged when one fails at 3 a.m. Without that, an automation platform turns into shadow IT; the guide to an [AI governance framework](https://www.ruebstudio.de/en/guides/ai-strategy/ai-governance-framework) covers the organisational side.

## Rolling it out step by step

The safe path runs from one narrowly scoped process through a trial instance to a governed platform. Do not start with an agent. Start with a deterministic workflow and add AI steps only where fixed rules fall short.

1. **Pick one process.** High volume, clear rules, a result you can check, such as sorting incoming purchase orders from a shared mailbox.
2. **Choose a hosting model.** Cloud for a quick trial without sensitive data; self-hosted as soon as personal or confidential data flows through it.
3. **Start a trial instance.** The container from the README is enough for a first test:

```bash
docker volume create n8n_data
docker run -it --rm --name n8n -p 5678:5678 \
  -e N8N_DIAGNOSTICS_ENABLED=false \
  -v n8n_data:/home/node/.n8n docker.n8n.io/n8nio/n8n
```

The editor then runs at `http://localhost:5678`. Production adds PostgreSQL, TLS and backups.

4. **Build the workflow without AI first.** Trigger, data retrieval, logic, output. Build the error path at the same time.
5. **Add an AI step where it helps.** For example an AI Agent node that classifies free text. Protect any tool that writes or sends with human-in-the-loop approval.
6. **Test with real cases.** Set up evaluations with realistic and awkward examples before going live.
7. **Write down how it is run.** Credentials, monitoring, updates and ownership. A production checklist will follow in this guides section.

In our own operations, linking every workflow to an error workflow that posts the workflow name and execution ID on failure has proved worthwhile. It makes silent failures visible.

## Limits and risks

n8n is strong at integrations and at processes you can draw as a graph. It is weaker when agents must hold complex state across many steps, when you need rigorous software testing and code review, or when your team has no capacity to run yet another platform.

- **Licence boundary:** commercial redistribution and running it for third parties are not covered by default (see above).
- **Maintainability:** visual workflows with dozens of nodes and embedded code are hard to review. Without Git-based version control, which the Community edition lacks, you also lose a proper change history.
- **Agent depth:** the AI Agent node handles tool calls, memory and retrieval. For agents with many intermediate states, branching plans and recovery after errors, code frameworks usually give you more control; the [comparison of AI agent frameworks](https://www.ruebstudio.de/en/guides/agent-frameworks/ai-agent-frameworks-comparison) covers the options. The difference from n8n: there you write the agent as tested, reviewed code instead of wiring it together on a canvas.
- **Security:** an n8n instance holds credentials for many core systems. A compromised account or an exposed webhook has a wide blast radius, and an MCP endpoint without authentication is a real exposure.
- **Community nodes:** these are written by third parties. Check who maintains them before you rely on them in production.
- **The cost of ease:** because workflows are quick to build, many get built. Without an inventory, a year later nobody knows which automation carries which process.

n8n is a poor fit if you want to sell the automation as a product, if nobody in-house will own operations, or if the heart of your project is an autonomous agent with its own state logic.

## How do I get started?

Start with one process, one trial instance and one named owner. After four to six weeks you will know whether the interfaces, data quality and operations hold up, and whether an AI step adds real value or a plain rules-based workflow is enough.

Three questions settle most decisions:

1. **Must the data stay in-house?** If so, self-host with telemetry disabled and, where needed, local models.
2. **How many people will build?** Once several teams need separate permissions, a licensed edition becomes the practical choice.
3. **Is it a process or an agent?** Processes with targeted AI steps are where n8n shines. If an autonomously planning agent is the core, compare n8n with code-first frameworks.

To see which business processes suit agents, the overview of [AI agents for companies](https://www.ruebstudio.de/en/ai-agents) covers areas such as [quotation](https://www.ruebstudio.de/en/ai-agents/quotation) and [technical service](https://www.ruebstudio.de/en/ai-agents/technical-service). More background is available in the [guides section](https://www.ruebstudio.de/en/guides).

## Frequently asked questions

### Is n8n open source?

Not by the OSI definition. The source code is public, but it ships under the Sustainable Use License, which n8n calls fair-code and which limits use to internal business, non-commercial and personal purposes.

### Can a company self-host n8n for free?

For your own internal business processes the licence allows it. SSO, Git-based version control, environments and log streaming are missing from the Community edition and need a paid licence.

### Which language models does n8n support?

The docs list chat model sub-nodes for providers including Anthropic, OpenAI, Google and Ollama. Ollama lets you plug in models running on your own hardware.

### Can n8n expose workflows to MCP clients?

Yes. The MCP Server Trigger node publishes connected tools and workflows at a URL over Server-Sent Events or streamable HTTP, with optional bearer or header authentication.

### When do I need PostgreSQL instead of SQLite?

SQLite is the default and works for a single instance. Once you scale out with queue mode and several workers, PostgreSQL and Redis are required.

## Sources

- [n8n: Sustainable Use License (LICENSE.md)](https://github.com/n8n-io/n8n/blob/master/LICENSE.md)
- [n8n: GitHub repository and README](https://github.com/n8n-io/n8n)
- [n8n: Releases on GitHub](https://github.com/n8n-io/n8n/releases)
- [n8n Docs: Choose how to use n8n](https://docs.n8n.io/choose-n8n/)
- [n8n Docs: Community Edition features](https://docs.n8n.io/hosting/community-edition-features/)
- [n8n Docs: Supported databases and settings](https://docs.n8n.io/hosting/configuration/supported-databases-settings/)
- [n8n Docs: Queue mode](https://docs.n8n.io/hosting/scaling/queue-mode/)
- [n8n Docs: LangChain in n8n](https://docs.n8n.io/advanced-ai/langchain/overview/)
- [n8n Docs: MCP Server Trigger node](https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-langchain.mcptrigger.md)
- [n8n Docs: Human-in-the-loop for tools](https://docs.n8n.io/build/integrate-ai/ai-examples/human-in-the-loop-for-tools.md)
- [n8n Docs: Understand why to test (Evaluations)](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/understand-why-to-test.md)
- [n8n Docs: Advanced AI](https://docs.n8n.io/advanced-ai/)
- [n8n Docs: Nodes and integrations](https://docs.n8n.io/integrations/readme.md)
- [n8n Docs: Control telemetry](https://docs.n8n.io/deploy/host-n8n/configure-n8n/security/control-telemetry.md)
