n8n AI Agent: what you can build yourself, and where n8n stops

In n8n you can build, without code, a single-purpose agent: one trigger, one model, one memory node and three or four tools, answering real messages by the end of the night. The platform stops at five limits: memory between sessions, error handling, live conversation, cost at scale, and maintenance. The standard licence does not let you host one instance and charge others for access.

Ask whether you can build an AI agent yourself and n8n comes up nine times out of ten. Fair enough - it is currently the shortest path from an idea to a working agent without writing code. Drag in the AI Agent node, attach a model and two tools, and the first version answers the same evening.

The trouble starts later. Not during the build, but at the three hundredth conversation, at the first failure at 11pm, and at the moment a client asks whether you can deploy the same thing for them.

This article covers both halves: what the AI Agent node genuinely does, what running it actually costs, and the point where n8n stops being the right home for the thing you are building. We build agents for a living and pick tools to fit the job, not the other way round - so treat this as a map rather than a pitch.

What does the n8n AI Agent node do?

It connects a language model to a set of tools and lets the model decide which tools to call to finish a task. That is the difference from a classic workflow, where every step is fixed in advance. The agent gets a goal and works out the order itself: check the calendar, write to the CRM, send a confirmation. You describe the destination, not the route.

You attach three kinds of component:

• A language model - Claude, GPT or a local model. The reasoning part.
• Tools - search, a knowledge base, an HTTP call, or any other n8n node. Each tool carries a description, and the agent decides when to use it based on that description.
• Memory - conversation history, so a customer does not have to repeat themselves in the next message.

Underneath, the node runs a loop. The model reads the goal, picks a tool, sees the result, and decides whether it now has enough to answer or needs another call. One customer question can trigger four or five passes around that loop before anything reaches the person waiting.

That loop is also where the real work sits, and it is not where beginners expect. The model only knows a tool exists because of the sentence you wrote describing it. Vague descriptions produce an agent that calls the wrong thing confidently. Precise ones produce an agent that behaves. Most of the tuning on a live n8n agent is rewriting those descriptions and the system prompt, not rewiring the canvas.

One capability gets far less attention than it deserves and matters most in a business deployment: n8n lets you require human approval before the agent runs a specific tool. Sending a message, changing a record, deleting data - you can gate all of it. Oversight is built into the node rather than bolted on afterwards, which is the single biggest reason we still reach for n8n when a process touches money or customer records.

What can you actually build in an evening?

A single-purpose agent: one trigger, one model, one memory node and three or four tools, wired on one canvas and answering real messages by the end of the night. That is not a demo. It removes work. What does not fit into an evening is the second half - the edge cases, the retries and the question of who owns it next month.

The anatomy is always the same. A trigger starts the run: an inbound email, a form submission, a chat widget on your website. The AI Agent node sits in the middle. The model hangs off it, memory hangs off it, and each tool is a node you already know how to configure - Gmail, Sheets, your CRM, an HTTP request to whatever has an API. If you have built any n8n workflow before, nothing here is new except the fact that the order is no longer yours to decide.

Three shapes come up most often in our inbox:

1. Inbox triage. The agent reads incoming mail, classifies it, labels it and routes it. Sales enquiries reach you with a summary attached.
2. Answers from a knowledge base. Load your price list, terms and service descriptions. The agent answers strictly from those documents and hands over anything outside that scope.
3. Lead enrichment. A form submission arrives, the agent checks public sources, adds industry and company size to the CRM and creates a task for whoever handles sales.

Each of those takes an evening if the process is already written down. If it is not, the evening goes on something else: writing down what happens between an enquiry arriving and the case closing, including who currently decides what. Skip that and you automate the mess.

Which process to point it at first is a separate question with a separate answer, and we work through it in our guide to AI automation for business. This article assumes you already know the target and want to know what the tool will and will not do once you aim it.

What does an n8n AI agent really cost?

The platform starts at 20 EUR a month for 2,500 executions, but the invoice has three lines: the n8n plan, model usage billed by whoever provides the model, and your own time keeping it alive. The third line is usually the largest and appears on no pricing page.

From the official pricing page, as of July 2026:

n8n prices in euros, so those are the figures that land on your invoice.

One mechanic matters before you pick a tier. An execution is one full run of a workflow, no matter how many steps it contains or how much data it moves. An agent that calls five tools in a single conversation burns one execution, not five. For conversational agents that is unusually good news, and it is why the entry plan stretches further than people expect - 2,500 executions is 2,500 conversations, not 2,500 tool calls.

The bad news is the jump between tiers. Going from 10,000 to 40,000 executions takes you from 50 to 667 EUR. There is nothing in between. A company that outgrows Pro mid-month pays thirteen times more for an order of magnitude it usually does not need, and the honest workaround is to watch your execution count and consolidate workflows before you hit the ceiling, not after.

Model usage sits outside all of that and scales with conversation length rather than conversation count. A long, tool-heavy exchange costs several times what a one-line answer costs, so the bill tracks how chatty your agent is. Model spend can pass the n8n subscription itself somewhere around 3,000 conversations a month, especially if every answer carries a large knowledge base. Work it out before you launch: conversations times average tokens per conversation times your model's published rate.

The same pricing page lists a Community edition, free to self-host with no execution limit. It has no SSO, no vendor support and fewer team features, and the server, backups, updates and 2am recovery are yours. It is not free. You pay in time instead of money, and time is the line item that people underprice.

For how those three lines compare with the other ways of deploying AI in a company, we break down the four deployment tiers in our AI costs guide.

Where does n8n stop?

At five limits: memory between sessions, error handling, live conversation, cost at scale, and maintenance. None of these is a flaw. They are the points where an automation platform stops being the right home for a product that serves your customers, and where the work turns from configuration into development.

1. Memory does not survive the session. The memory node holds the current conversation. n8n's own documentation describes Simple Memory as chat history for the current session, and a customer returning next week starts from scratch. You can work around it by attaching a database or vector store as a tool and writing the retrieval logic yourself, at which point you have a development project with a schema to maintain, not a configuration.

2. Error handling is on you. The model will answer badly. A tool will time out. An API will return something nobody planned for. A production-grade product retries, falls back to a safe answer and escalates to a person, and in n8n you build each of those behaviours yourself, branch by branch, for every path the agent can take. Across three scenarios that is an hour of extra wiring. Across thirty it is a system that itself needs managing, and the canvas stops being readable long before that.

3. There is no live conversation. n8n handles audio in batches: the OpenAI node offers Transcribe a Recording and Generate Audio, and the Twilio node can place a call and read a message out via text to speech. What is missing is the streaming layer. A phone agent has to listen and speak at once, let the caller interrupt mid-sentence and hold sub-second latency, while an n8n workflow runs request and response. Different category of tool, and we compared the options in AI voice agent platforms.

4. The economics stop working at scale. At a few hundred conversations a month, n8n is close to unbeatable on price. At tens of thousands, the Business plan plus model spend plus the hours somebody puts in reach the price of something purpose-built with no execution ceiling. The crossover is a calculation, not a principle, and for most companies it never arrives.

5. Maintenance never finishes. This is the one that surprises people, because it has nothing to do with the platform. Change your prices and the agent needs to know. Your model provider ships a new version and behaviour shifts under you without a single edit on your side. Swap CRM and the integration gets rebuilt. An agent is not a website you launch and leave. It is closer to a junior colleague who needs the occasional correction, and it degrades quietly rather than failing loudly.

Can you sell an agent you built in n8n?

Not by hosting one instance and charging others for access, and this catches out most people planning to build a business on the free edition. Worth reading the terms before the model, not after the first client.

n8n is not open source in the usual sense. It ships under the Sustainable Use License. Under the licence text you may use, copy and modify the software for your own internal business purposes or for non-commercial use, and you may distribute it to others only free of charge and for non-commercial purposes. Files marked as Enterprise require a separate licence.

In practice that draws a clean line through the three things people actually want to do:

• Building an agent for your own company - fine, including the free self-hosted edition. Internal business purposes is exactly what the licence describes.
• Configuring an agent on a client's own instance and their own account - ordinary implementation work. They run the software, you do the build, and you invoice for your time and expertise the same way you would for any other configuration job.
• Hosting n8n yourself and selling access to it as your product - outside the standard licence. That is a conversation with n8n about commercial terms, and they do have them.

The distinction is who runs the instance, not who wrote the workflow. Selling your own hours is not the problem. Reselling somebody else's software as a service is.

We are not a law firm and none of this replaces advice on your specific contract. But if your offer depends on running one instance for ten clients, read the licence yourself before you sign the first one, and ask n8n directly if the answer is not obvious. It is a short document and a cheap hour.

Build it yourself or buy it?

Build when the process is internal, volume is low and a mistake costs you an hour. Buy when the agent talks to your customers and its failure costs you a customer. That line predicts outcomes better than comparing price lists, because the expensive part of getting this wrong is never the subscription.

The available data points the same way. MIT's NANDA report ("The GenAI Divide: State of AI in Business 2025") found that buying from a specialised vendor and adapting succeeds around 67% of the time, while building from scratch internally succeeds at about a third of that rate. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, mostly over escalating costs, unclear value and weak risk controls.

Neither number says "do not build in n8n". Both say the same duller thing: decide who owns the result and what success looks like before you open the canvas. Projects in that cancelled 40% rarely die because the technology failed. They die because nobody could say what the agent was supposed to be worth.

There is a third option that neither statistic covers, and it is often the cheapest. If the gap is that one person built the agent and nobody else understands it, the fix is not a bigger platform. It is a couple of days of AI training so more than one person can read the canvas and adjust a prompt. We have talked companies out of a build more than once on exactly that basis.

n8n usually earns its place in the stack, just not as the agent's brain. It works well as the integration layer connecting CRM, calendar and invoicing, while the decisions run in an AI agent built separately with its own memory, error handling and monitoring. Most deployments end up as some mix of the two rather than one or the other. If you want the wider three-way view of chatbot versus automation versus agent, that is in what you are actually buying.

Four questions that settle it

Answer these four honestly and you will know whether n8n is enough, without a proposal from anybody.

1. Who talks to the agent - your team or your customer? A colleague forgives a mistake and mentions it at lunch. A customer does not call back and does not tell you why.
2. How many conversations a month? Up to about a thousand, n8n wins on price and it is not close. Above that, count the model spend and your own hours before you compare anything.
3. What happens when it gets something wrong at 11pm? "Nothing, I will fix it in the morning" means build it yourself. "We lose the order" means you need escalation, retries and monitoring, and those are the parts n8n makes you assemble by hand.
4. Who maintains this in six months? Name the person. If you cannot, that is the finding. An agent without an owner degrades quietly: first it answers worse, then people stop using it, then someone switches it off.

Score it roughly. One or two answers on the serious side and n8n is still the right call. Three or four and you are past weekend-configuration territory. Even then the answer is not automatically to buy something. It is to decide who keeps this running and give them the time to do it.

Frequently asked questions

Is the n8n AI agent free?

The Community edition is free and self-hostable with no execution limit, so the software genuinely costs nothing. You still pay your model provider for every conversation, plus the server, backups, updates and your own maintenance time. Cloud plans start at 20 EUR a month for 2,500 executions and remove the hosting work in exchange.

What does the n8n AI Agent node do?

It connects a language model to tools and lets the model choose which tools to call, in what order, until the task is done. You can attach conversation memory, a knowledge base and any other n8n node as a tool. You can also require human approval before a sensitive action runs, which matters when the agent touches money or customer records.

Can I build an n8n agent without coding?

Yes. The first working agent needs no code at all, and most of the tuning afterwards is writing better tool descriptions rather than writing logic. Code shows up later, with unusual integrations, custom error handling and memory that survives between sessions - the things that separate an evening's demo from a product that handles customers unattended.

Will an n8n agent remember a previous conversation?

Not by default. The memory node covers the current session, and a customer coming back next week starts from scratch. Persistent history means attaching a database or vector store as a tool and writing the retrieval logic yourself. That is doable, but it turns a configuration job into a small development project with a schema to maintain.

Can I resell agents built in n8n?

Configuring an agent on a client's own instance and their own account is standard implementation work, and you invoice for your time as normal. Hosting your own n8n instance and selling access to it as your product falls outside the Sustainable Use License and needs commercial terms agreed directly with n8n. The test is who runs the instance, not who built the workflow.

Can n8n run a phone agent?

Not a live one. n8n will transcribe a recording and generate speech through the OpenAI node, and place a Twilio call that reads a prepared message out. What it lacks is the streaming layer a two-way call needs: listening and speaking at once, interruption mid-sentence and sub-second latency. That is a separate category of tool with a separate set of vendors, and we compare them in our voice platform guide. You can still use n8n behind such an agent as the integration layer.

When does n8n stop being enough?

When the agent talks directly to customers, handles more than a few thousand conversations a month, needs memory across sessions, or when its failure costs money the same night. At that point you are paying to build error handling, persistence and monitoring by hand, and that cost quietly overtakes the cost of something purpose-built that ships with them.