Which AI model can I connect to n8n?

The familiar ones: Claude (Anthropic), the OpenAI models, and through separate nodes also European or local models like Mistral. You pick the model per workflow that fits the task and your privacy requirements.

Does my data then go to an American cloud?

Only if you choose that. You can use a European or local model, and you decide which fields you send to the model. Always start with the question: which data is allowed to leave here, and which isn't?

What if the AI gets something wrong?

You build around that. Put human checks on the steps that really matter, set hard rules where nothing is allowed to go wrong, and log every decision so you can read it back.

Guide · 1 min read

n8n with AI: agents and language models in your workflows

AI turns an n8n workflow from a pass-through into something that thinks along. How to put models like Claude and OpenAI to work, and keep it under control.

A plain n8n workflow moves data: from A to B, with a few rules in between. Add AI and the workflow also starts to interpret: summarising an email, gauging the tone of a message, classifying a document, drafting a reply. That's the difference between shuffling and thinking along.

The AI nodes in n8n

n8n has ready-made blocks for AI. The most important ones:

The language-model block: connect Claude or an OpenAI model and send it text with an instruction ("summarise this email in three bullets")
The AI agent node: a model that gets to choose for itself which tools it uses to reach a goal (more on that in building an AI agent)
A vector store: for when the AI has to answer based on your own documents (this is called RAG)

Three examples that pay off straight away

Triaging email. Incoming mail is read, sorted (quote request, complaint, invoice) and sent to the right person or folder, with a short summary attached.
Preparing quotes. From a request the AI pulls the key points, fills a template and sets up a draft, a human hits send.
Answering from your own knowledge. A customer's question is answered based on your manuals and documentation, not on whatever the model happens to think it knows.

Keeping it under control

This is what it's really about, and where a lot of hobby projects go wrong. AI in a work process needs guardrails:

Human checkpoints on the steps that touch money or reputation
Hard rules alongside the model, for what is simply not allowed
Logging of every decision, so you can see afterwards why the AI did what it did

That's how you build something your team dares to lean on, instead of a black box no one dares to look behind. It's the same approach we use when building AI workflows.

[Case, to be filled in: describe an AI workflow you built in n8n, with the human check that's in it and what it delivered.]

Getting started

If you've already done the n8n tutorial, you can add an AI step in fifteen minutes. Want to go deeper into the world of AI agents? That's the logical next step. And if you'd rather have it set up properly and safely straight away: we build it.

Frequently asked questions

Building an AI agent: how to approach it

An AI agent that actually gets things done isn't built by turning a model loose. It comes down to a sharp goal, the right tools, guardrails and a human in the loop.

Lees verder