Why I no longer need n8n

For a while I automated my work by drawing workflows in n8n, and I was very proud of them. Today I ask an agent for almost all of that in a single sentence and I barely open the program. Here's what I had, what it cost me, when I switched and, without caricature, when workflows are still the best option.

You know the video: "automate your business with n8n, no coding". A clean desktop, some coloured boxes that connect themselves and, at the end, an email that lands in a spreadsheet as if by magic.

For anyone who doesn't know it, n8n is a tool for automating tasks by linking applications through workflows. On screen you see little boxes joined by arrows, and each box is a step: "an email arrives", "pull out this data", "save it here", "ping me on Telegram". You can use it in the cloud or install it on your own server.

I came to n8n at the end of 2024, when I wanted to stop copying a website's contact forms into my client spreadsheet by hand. It felt like a revelation: at last I could automate without writing code, and on my own server too, without depending on a monthly fee. I spent several weeks hooked, looking for things to automate like someone who's just bought a drill and wants to make holes in every wall.

A note of honesty before going on. The title says I no longer need it, and that's true, but I haven't deleted anything. It's still installed and still running one workflow, which I talk about below. I use n8n much less because agents have replaced it for almost everything, not because I've thrown it out of the window.

At my peak I had a dozen workflows. Three of them, described in plain language:

The first collected the website's contact forms, cleaned them up, added them to a spreadsheet with the date and sent me an alert on my phone. The second, every morning, checked the status of my servers and wrote me a summary on Telegram. The third took the invoices arriving by email, pulled out the amount and the supplier and saved them in a folder organised by month.

I had it installed on one of the internal network servers, inside a container, with tightly controlled access from outside to receive the forms.

I was immensely proud of the first one. I remember showing it to a friend the way you'd show off a new car: "look, I fill in the form and, ta-da, my phone buzzes". He gave me a look that said "and you spent three afternoons on that?". He had a point, but I was delighted.

The small print doesn't make it into the videos.

The workflows broke whenever something changed outside. The invoices one stopped working when one of the suppliers changed the format of its emails, and I didn't find out until three weeks later, when I went looking for an invoice that wasn't there.

Then there was JSON, which is the standard way programs pass each other structured data. While the workflow is simple, you don't see it. As soon as it gets complicated, expressions appear for pulling one piece of data out of another, and that's the end of "no coding".

And the updates, to n8n itself and to the server it lived on. One update changed how one of the boxes worked and I had to go through several workflows by hand.

I reckon it took me between three and five hours a month to keep it all running. And there was a whole weekend, a Saturday and half a Sunday, lost because the server restarted and n8n didn't come back up properly. The rush was because the server summary, the one that warned me of problems, was precisely one of the workflows that was down.

The change came sideways, unplanned. I wanted to rebuild the invoices workflow, the one that had broken. Instead of opening n8n, I opened Claude Code on the server and wrote something like this: "Every day, check the invoices inbox, pull the supplier, date and amount from each invoice, save the PDF in that month's folder and record the data in the expenses spreadsheet. If there's anything you don't understand, set it aside and let me know".

It suggested how to do it, I asked for a few changes, it scheduled it for every morning and I tested it with the previous month's invoices. All in all, just under an hour. The original workflow had cost me about five afternoons between building it and fine-tuning it.

The feeling, I won't deny it, was of having spent months drawing little boxes for nothing. Like learning to make fire with sticks the day before someone gives you a lighter.

But careful. I had to check what it had done for two weeks: on three invoices it misread the amount and on one it got the supplier wrong. It wasn't magic; it was much faster to set up and it still had to be watched. The difference is that when something went wrong, I didn't have to hunt for the faulty box: I explained the error to the agent and it fixed it, and I checked that the fix made sense.

This is how the two approaches look in my case, with figures that are my own estimates:

The agent wins by a mile on setup and on changes: if I want it to do something different, I tell it and that's that. It loses on three things. Repeating exactly the same thing a thousand times, where a workflow is more predictable. Pay-per-use costs, because every run consumes model time and that gets paid for. And silent errors: a broken workflow usually fails with a crash; an agent can get it wrong with complete confidence and without any alarm going off. I say more about that in The trouble with working with agents.

I don't want to draw a caricature. There are cases where I'd keep using workflows without hesitation.

When there's a high volume of identical repetition: the same operation hundreds or thousands of times a day, no thinking required. When the process always has to produce exactly the same result, with no interpretation. And in teams where nobody handles agents, but someone can look at a drawn workflow and understand it.

The workflow I still maintain is the contact forms one. It's simple, lots of them come in, it always has to do the same thing and it hasn't broken in a year. There's nothing to improve there.

If you still have to choose a workflow tool, the typical dilemma is n8n or Make. Make is easier to start with and you don't install anything, but it runs in the cloud and charges by usage. n8n gives more control because you can run it on your own server, at the price of paying for the server and the time to maintain it. Neither is better in the abstract.

And there are more alternatives to n8n: Zapier, which does the same in the cloud; Node-RED or Activepieces, which you can install yourself. And the alternative of a different kind, the one I use: asking an agent for the automation directly, in plain language. I'm not quoting prices because I haven't checked them.

There are things I miss. Seeing the process drawn out, with its arrows, and understanding it at a glance. Someone else being able to look at it and know what it does without asking me. Knowing exactly what's going to happen at each step.

When a client asked me how the automation of their invoices worked, with n8n I'd have shown them the screen. With the agent I had to explain it in words and show them the instructions, and it's not the same. It's not a disaster, but it's real. A drawn workflow is documentation in itself; instructions in plain language depend on the agent always interpreting them the same way, and I can't guarantee that.

If I were starting to automate tasks with AI today, I'd do it in three steps.

1. Pick one repetitive task you already do by hand. Just one.
2. Ask an assistant with access to your tools to do it first, and check the result every time until you trust it.
3. Only if it's repeated a huge number of times and has to come out identical every time, turn it into a workflow.

The opposite of how I did it, starting with the little boxes. And if the promise of automating everything in no time rings a bell, in Can you automate your life in ten minutes? I put it to the test with my own tasks.