n8n Self-Hosted AI Agents: An Inbox That Triages Itself

Most automation platforms charge you for every step a workflow takes. n8n charges for the run. Its pricing page puts it plainly: an execution is a single run of the whole workflow, however many steps it has and however much data it handles.

That one decision shapes where n8n fits best. Workflows with many steps cost the same as workflows with few. And AI agents, which take many steps by nature, suddenly become cheap to run.

Add the fact that n8n can be installed on your own servers, and its sweet spot comes into focus: AI agents that work on sensitive data, run by a team with enough technical confidence to host them.

The build: a support inbox that triages itself

A software company receives around 3,000 support emails a month. A person reads each one, decides whether it is a bug, a billing question, a how-to or an urgent outage, looks up the customer, and routes it. Urgent emails sometimes wait an hour because they look like everything else.

The agent that takes over the first pass works like this:

A new email arrives in the shared support mailbox.
The agent reads it and decides what it is about.
It looks up the customer in the CRM, through a tool it is given: plan, account owner, open tickets.
It searches the knowledge base for articles that answer the question.
It creates a ticket with a category, a priority and a one-line summary.
For how-to questions, it drafts a reply using the matching article and holds it for a person to approve.
For anything that looks like an outage, it posts to the on-call channel immediately.

Steps 3 to 7 are where the agent earns its keep. n8n’s AI Agent node is given a chat model and a set of tools, and decides for itself which tools to call and in what order. A billing question needs the CRM lookup and the ticket. A how-to needs the knowledge base and a draft. The agent does not follow a fixed path, which is exactly why it is useful for a varied inbox.

Why self-hosting matters here

A support inbox is full of personal data: names, email addresses, account details, sometimes payment problems or medical context. Sending all of it to a third-party automation service is a decision that legal and security teams often, quite reasonably, want to avoid.

The free Community Edition of n8n can run on your own server, inside your own network. The email, the CRM lookup and the ticket never leave your infrastructure. The only data that goes outside is what you choose to send to a language model, and you choose that model. It can be a hosted service your organisation has already approved, such as Azure OpenAI, or a model running on your own hardware, in which case nothing leaves at all.

For healthcare, legal and financial services, that is often the difference between an agent being approved and being shelved.

What it costs to run

Each triaged email is one execution, whether the agent took three steps or eleven.

On n8n’s own cloud, the Starter plan covers 2,500 executions a month for €20, billed annually, and Pro covers 10,000 for €50. At 3,000 emails a month, Pro covers it with room to grow.

Self-hosted, the Community Edition has no execution limit and no licence fee. What it does have is upkeep: a server to run it on, backups, updates, monitoring and someone who notices when it stops. A small virtual machine costs little. The person who looks after it costs more, and that is the real comparison to make.

The model calls are a separate cost, and usually the larger one. An agent that reads each email, decides what to do and drafts a reply makes several model calls per message. We go into that arithmetic, and how to keep it in check, in where agentic AI gets expensive.

Guardrails that make it safe to switch on

An agent with access to your CRM and ticketing system needs boundaries that live in the workflow, not in the instructions you give the model.

Give it narrow tools. The CRM tool should look up one customer by email, not run arbitrary queries. The ticket tool should create tickets, not close them.

Keep a person on anything outbound. Drafting a reply is cheap and safe. Sending one is not. In the first months, every draft waits for approval.

Log what it decided and why. Each ticket should record the category the agent chose and the evidence it used. When it gets one wrong, that record is how you fix the instructions.

Run it in parallel first. For the first two weeks, let the agent triage alongside the person who does it today, and compare. The match rate is your evidence for letting it run alone.

n8n now speaks the Model Context Protocol, which is becoming the common way for AI agents to use tools. It works in two directions.

The MCP Client Tool node lets an n8n agent use tools that live elsewhere, such as an internal service your developers already expose. The MCP Server Trigger does the reverse: it publishes your n8n workflows as tools that other AI assistants can call.

In practice, that means the “look up a customer” workflow you built once can be used by the support agent, by an assistant your sales team uses, and by a developer’s coding tool. We cover building those tools on .NET in agentic workflow automation with MCP.

The licence point most teams miss

n8n is “fair-code”, not open source in the strict sense. Its Sustainable Use License lets you use and modify it for your own internal business purposes, which covers the support inbox above and almost every in-house use.

What it does not allow is distributing it or offering it to others as part of a paid service. If your plan is to host n8n and sell access to customers, or build a commercial product on top of it, read the licence first and talk to n8n about an enterprise agreement.

When n8n is the wrong choice

n8n rewards technical confidence. If nobody on the team is comfortable with servers, credentials and the occasional bit of JavaScript, the hosted cloud plans help, but a more guided tool may suit you better.

Some paid features matter as a team grows. Single sign-on, separate development and production environments, and version control through Git sit in the higher tiers, with the self-hosted Business plan at €667 a month billed annually. Budget for them if several people will build workflows.

And if the job is mostly customer messaging for a local business, an all-in-one platform such as GoHighLevel does more out of the box. Our comparison of all three sets them side by side.

The short version

n8n is at its best for workflows with many steps, AI agents especially, on data you would rather keep in-house, run by a team that is happy to own a server. Price it by the run, not the step. Keep the agent’s tools narrow and a person on everything outbound. And if you plan to sell it on, read the licence first.

Prices from n8n’s pricing page, November 2026, billed annually. Licence terms from the Sustainable Use License in n8n’s repository. The support-inbox example is illustrative.

We build AI agents on n8n and on .NET, and can deploy them on your own infrastructure when the data calls for it. See our AI agents and copilots service, or get in touch.