Hi, I am Romuald Członkowski. In this lesson I will show you how to build an AI agent in n8n by describing it, in plain words, to your AI assistant. The course grew out of a two-hour workshop at AI Summit Poland, and here you take it at your own pace.

A narrator reads the commentary for each slide and the lesson moves on by itself. When it is time for an exercise, it stops and waits for you. Pause is always within reach, and the K key hides and shows the commentary.

Before we start, one thing: you have just been through something. You filled in a form, clicked a link in an e-mail, and a moment later your environment was ready. Nobody did that by hand.

Everything in the blue frame was done by an automation: the sign-up queue, your n8n account, access to the AI model, the connector account, a copy of the starter workflow “Course start”, and the e-mail at the end.

That is a dozen or so workflows and one web page. My AI assistant built them while I said what I wanted and checked the result. It was not a single prompt: several days of work and plenty of corrections. In this course you will build the smallest piece of such a system, with the same method.

I run an agency called AiAdvisors. For this course something else matters more: I am the author of n8n-mcp, the connector through which assistants such as ChatGPT and Claude work inside n8n. You will use it in the exercises.

Most of us use AI like the left side: we open a chat, ask, and get a result. These tools are becoming a desktop for work. There is one drawback: someone has to sit there. Scheduled tasks help, but every run uses tokens, even when there is nothing to do.

In a business process a message arrives at three in the morning and has to be on the right desk by the time people come in. AI has to be built into a process that starts by itself. And not everything needs AI: whatever you can describe with a rule, a rule handles.

This is the thesis of the whole course. An assistant, that is Claude, ChatGPT or Codex, builds the workflow. It does that once, on your instruction. Inside the workflow another model does the work: it reads every message and decides, without you.

The word “workflow” will come up all the time, so first a definition: it is a process written down step by step so that a computer can run it by itself. Something starts it, then come the steps, and the result of one step goes to the next.

Until recently, automation stopped where somebody had to read a text and judge it. The AI step does exactly that.

n8n is a tool in which you put such workflows together from blocks. Each block is one action, a line shows where the data flows, and the button at the bottom runs the whole thing. There are hundreds of ready-made blocks, and after a run you can see every step and the data that passed through it.

Two things matter for a company: you can run n8n on your own server, so the data stays with you, and for internal use there are no licence fees. Someone has to maintain it, but that is not a burden. I do not represent n8n; I simply use it heavily myself.

Before the assistant starts building, look around n8n yourself. This is exercise zero, “Get to know n8n”, and it takes about ten minutes. You open the workflow “Course start” and look at the canvas: every rectangle is a node, and the lines show where data flows. The first node on the left is the trigger. Double-click a node to see its input, its parameters and its output. Then click “Execute workflow” and look at the “Executions” tab, which is the history of all runs.

The n8n interface is English only, which is why the exercises page has a glossary. Your login details are in the e-mail about your ready environment.

Your turn. course.aiadvisors.pl/exercises The lesson pauses here. Open the exercises page from the link in the commentary, do exercise zero, and come back when you are done. Then press “Resume”.

You know how the process works, and you describe it in words. The assistant turns that into a workflow. It does this through the n8n-mcp connector: with it the assistant sees your n8n, gets tools, and gets hints on what to fix when something goes wrong. Without the connector you are left copying and pasting.

Your n8n account, access to the AI model, the connector account and the starter workflow are already prepared. Connecting your assistant and installing the skills is your part, and you follow the guide on the start page.

An assistant that builds should have a strong model. I gave the same task to two models: one finished in two minutes, the other worked for over six, failed several times, and in the end announced success with an empty table.

A fast model quickly produces text, but not necessarily the right text. A stronger one is slower and usually gets it right the first time. Small models have their place too, just not here.

The plan is simple: one workflow, extended three times. First sorting messages, then urgent cases and a notification, and finally a draft reply. All the instructions are on the exercises page.

You will connect the course model to the workflow. There are two: “luna” is fast and good for everyday steps, and “sol” is stronger. We start every exercise with “luna”.

This is the heart of it. The assistant will build what you describe, so the whole skill is a good description. You do not need to program: it is enough to describe the business intent.

It does not have to be a single instruction. For bigger workflows you talk to the assistant and build step by step. In this course we practise the one-instruction version, because it teaches you to think about the whole process at once.

First a plan: the assistant shows how it understood the instruction, and you read it before it builds anything. Then a test: the assistant runs the workflow itself and shows the result. The second rule made a big difference for me.

The example is made up: a company called Green Office and ten messages from customers. Among them are enquiries, complaints, invoices and some spam. Four people, each handling something different.

The inbox is an address on the course server, not real mail, so nobody has to configure anything. In a company you would have Gmail or Outlook there.

This is what the result of the first exercise should look like. The workflow fetches the messages, the AI assigns a category, a plain rule picks the owner, and everything lands in the “queue” table, which is a list of cases to handle. The model in the AI step is “luna”.

Your assistant may arrange the blocks a little differently. That is normal. What counts is the result in the table.

Copy the ready-made instruction from the exercises page and paste it to your assistant. For me the build took about five minutes. In this kind of work you spend a lot of time waiting for the model to finish.

Your turn. course.aiadvisors.pl/exercises The lesson pauses here. Open the exercises page from the link in the commentary, do the first exercise, and come back when you are done. Then press “Resume”.

The first run fetches only three messages as a trial. After I removed the limit, my run sent four to sales instead of three.

That is not an algorithm, it is a model, so results differ. You fix it with the description in the AI step, without rebuilding the workflow.

This message is disputed on purpose. In my run the model treated it as a complaint, and I had meant accounting. The lesson: you always check the AI result, and when it differs from what we wanted, we ask the assistant to improve the rule’s description.

The queue holds ten cases and they all look the same. A cracked desktop before the office opens cannot wait like a job enquiry.

You write what “urgent” means. I wrote: a case that has to be handled within two days. I also asked for a separate table for the cases where the AI is unsure, so that I could review them later.

Your turn. course.aiadvisors.pl/exercises The lesson pauses here. Open the exercises page from the link in the commentary, do the second exercise, and come back when you are done. Then press “Resume”.

Most people get three urgent cases. One person got four, even though they wrote the same rule as I did. A model is not deterministic, and you have to allow for that in your tests.

Here I saw a mistake happen. I asked for playful drafts, and instead of changing the workflow the assistant built a separate branch. I asked why it did that and told it to fix it. That is what working with AI really looks like: you have to look at what was built.

The draft for the tender quote said: “without a specification it is easy to mistake a desk for a spaceship”. That is why the draft waits in the queue and a human sends it. Automatic sending only makes sense after long testing with the team.

Your turn. course.aiadvisors.pl/exercises The lesson pauses here. Open the exercises page from the link in the commentary, do the third exercise, and come back when you are done. Then press “Resume”.

Now the exercises for the quick ones, and you can do each of them on its own. Changing to a stronger model is one sentence to the assistant or one field in n8n: from “luna” to “sol”. Then we swap the plain AI step for an agent with a knowledge base, which is eight company rules in a table.

The stronger “sol” works in the AI Agent node as long as the option “Use Responses API” is on. The drafts start quoting the rules from the base, for example eight percent off for 40 chairs. But for a quote needed “tomorrow by twelve” my agent promised 24 hours: the rule was followed, and common sense was missing.

Your turn. course.aiadvisors.pl/exercises The lesson pauses here. Open the exercises page from the link in the commentary, do whichever extra exercises interest you, and come back when you are done. Then press “Resume”.

The last exercise is your own process. The template is on the exercises page. The assistant alone is enough, you do not need n8n. The easiest way is to turn on the microphone and answer its questions out loud. The assistant builds nothing: it asks about whatever is missing and proposes a plan.

Your turn. course.aiadvisors.pl/exercises The lesson pauses here. Open the exercises page from the link in the commentary, do the exercise “Your own process”, and come back when you are done. Then press “Resume”.

I would start with an internal process. When a colleague from the next department gets a strange e-mail, the cost of a mistake is small. When a customer gets it, the cost is big.

Good first processes are documents (for example orders that someone retypes into a system today), incoming mail, and a simple connection between two systems. You can add an AI agent later.

If a first automation can be built in two hours, a small process can be rolled out in a week instead of spending a month defining a big project. Before going live you need tests on real data. For one of our clients we check the bot’s answers on thousands of examples before we switch the next type of case to automatic.

Who automates is changing too. At another client a warehouse manager, after a few weeks of joint consultations, built a label-scanning app for his team himself.

Your practice environment lasts fourteen days and is then deleted. To keep building you need your own n8n. There are two ways.

The first is n8n Cloud: the account is ready in a few minutes and you need no server. You need a paid Starter plan or higher, because the connector uses the API. The link on the slide is a partner link: n8n shares a commission with me, and you pay the same. The second way is your own server: no licence fees, and the data stays with you. Your assistant will walk you through the installation, because among its skills is one for setting up n8n on a server. Link: n8n.partnerlinks.io/n8n-mcp

Moving takes three steps. First download the workflow from the course environment as a file. Do it before your fourteen days end, because after that the environment disappears. Then in n8n-mcp.com add a new instance: you give its address and API key. From that moment the assistant works on your n8n.

Finally give the downloaded file to the assistant and ask it to recreate the workflow in the new instance. It will upload it, create the tables and ask for a model key. Credentials and tables do not travel with the file, so you provide your own model key.

One step is left: the survey. It is the last tab on the exercises page, “Course survey”. A few questions about what worked and what did not. Your answers make the next version of the course better.

Your turn. course.aiadvisors.pl/exercises The lesson pauses here. Open the exercises page from the link in the commentary, fill in the survey, and come back when you are done. Then press “Resume”.

Thank you for going through this lesson. If you have a process you want to automate, write to me at romuald@aiadvisors.pl. You can also have a look at n8n-mcp.com, the connector this course is built on.