n8n AI Agent: What It Is and 5 Marketing Workflows I Run With It

n8n AI agent marketing workflows

n8n started life as a workflow tool: when this happens, do that. The n8n AI agent changes the picture. Instead of you mapping every step, you give a language model a goal and a set of tools, and it decides which tool to use and in what order.

I run a self-hosted n8n for my own marketing and for client automation projects. This post explains what an n8n AI agent is in plain English, how the AI Agent node fits together, and five marketing workflows where it earns its keep.

Want an AI agent working on your marketing?

I build n8n AI agents that research leads, update your CRM and draft content, with you approving anything that goes out. Happy to talk through your workflows.

Email me Know more about me

What is an n8n AI agent?

Standard n8n workflows operate on a fixed path. When a form arrives, n8n verifies the email, adds a lead to the CRM, and fires off an alert. Each execution follows this exact sequence of steps.

An n8n AI agent is a node that sits in a workflow with three things plugged into it:

  1. A chat model: the brain. Claude, GPT, Gemini, or a local model through Ollama.
  2. Memory (optional): so it remembers earlier messages in a conversation.
  3. Tools: the things it is allowed to do. Search the CRM, read a Google Sheet, call an API, run another workflow, query an MCP server.

You give it a goal (“find out what this company does and score how well it fits our ideal customer”), and the model works out which tools to call. The workflow around it still decides when the agent runs and what happens with the answer.

That last part matters. I think of the agent as a smart step inside a workflow, not a replacement for the workflow.

How the n8n AI Agent node works

Adding the AI Agent node reveals dedicated slots for the Chat Model, Memory, and Tool directly beneath it. These are the essentials.

  • System message: the agent’s job description. Be specific: who it works for, what a good answer looks like, what it must never do.
  • Tools: each tool needs a clear name and description, because that description is how the model decides when to use it. “Search SuiteCRM leads by email address” works far better than “CRM tool”.
  • Output parser: if the next node expects clean JSON, add a structured output parser so the agent’s reply matches a schema.
  • Max iterations: a ceiling on how many tool calls it can make. Keep it low while testing so a confused agent does not loop and burn tokens.

5 marketing workflows I run with an n8n AI agent

1. Lead research before outreach

A company name and website go in. The agent reads the homepage, checks the About page, pulls out what they sell, roughly how big they are and who the decision maker might be, and writes a two-line summary into the lead record. I review it before anything goes out.

2. Contact finder from a company list

For a list of companies, the agent scrapes each site, looks for named people and email patterns, and fills a sheet. A local model handles the reading, so I am not paying per page for something this repetitive.

3. Blog planning from a board

My blog ideas live on a Nextcloud Deck board. When I move a card into a “plan” column, a workflow sends it to an agent that writes a keyword table, an outline and the questions the post should answer, then writes it back to the card.

4. Inbox triage for website enquiries

New enquiries get classified (real lead, supplier pitch, spam, job applicant), and real leads get a short summary and a suggested reply. The routing itself is a plain workflow; the agent only does the judgement call. I covered the routing part in my n8n lead routing post.

5. An AI agent that talks to my other tools through MCP

n8n can both use and expose MCP servers. I have published several n8n workflows as MCP endpoints (CRM, project management, analytics, Tag Manager), which means an agent in Claude or n8n can work across all of them with one consistent connection. If MCP is new to you, start with my Google Tag Manager MCP post.

Which model should power your n8n AI agent?

  • For tool use and reliability: a strong hosted model (Claude, GPT, Gemini). They follow tool descriptions far better.
  • For cheap, repetitive reading: a local model through Ollama. Fine for extraction and classification; less reliable for multi-step tool use.
  • Mix them: I often use a local model for the bulk work and a hosted model only for the final judgement.

Mistakes to avoid

  • Giving the agent write access on day one. Start with read-only tools. Add writes one at a time, with a human approval step in between.
  • Vague tool descriptions. If the agent picks the wrong tool, the description is usually the problem.
  • No error path. Add an error workflow so a failed run tells you, instead of silently dropping a lead.
  • Letting it run unchecked. Look at the executions list for the first few weeks. You learn more from ten real runs than from any tutorial.

Is n8n the right AI agent builder for you?

If you already use n8n, the AI Agent node is the easiest way to add judgement to workflows you trust. If you are choosing a tool from scratch, n8n is a strong pick when you want to self-host, keep data on your own server, or mix local and hosted models. I compare it with Zapier in a separate post. If you want an agent like this built around your own CRM and inbox, see my AI agent development service, and CRM and ERP workflow automation for the workflows around it.

FAQ

What is an n8n AI agent? It is a node in n8n that combines a language model with memory and tools, so the model can decide which actions to take to reach a goal, inside a workflow you control.

Can an n8n AI agent use a local model? Yes. Connect the Ollama chat model node and point it at your Ollama server. Local models are good for reading and classifying; hosted models are better at complex tool use.

Is the n8n AI agent free? The node is included in n8n. If you self-host n8n you pay for your server and any hosted AI model you call; local models cost nothing per request.

Why is my n8n AI agent not using tools? Usually the tool description is unclear, the model is too small for tool calling, or the system message does not tell it when to use tools. Fix the descriptions first.

Leave a Comment

Your email address will not be published. Required fields are marked *