Why Low-Code Internal Tools Rot: The Requirement Change n8n Can’t Draw

The first version of the internal tool is a pretty canvas. A webhook, a sheet, a Slack message. n8n or Zapier or Make draws it in an afternoon. The requirement that kills it arrives on a Thursday: “also if they already paid, don’t create the folder, update the row, and if the tax ID is new, wait for a human, unless it’s the same company under a different name.” That sentence has loops, state, and a join. The canvas can fake it with IF nodes until the graph looks like a subway map nobody will touch. Then a credential expires and the rot is official.

I have shipped Zapier-heavy ops stacks and the Python that replaced them when the canvas stopped being a drawing of the job. This is not “n8n is bad.” n8n is excellent at the first sentence. It is a poor place to hide the fourth revision of a business rule. The rot is what happens when you keep drawing instead of admitting the rule is now a program.

What rot looks like in a real workflow

Nodes named IF3 copy 2. A Google Sheet that is a database and a UI and a lock. A wait node that is someone’s lunch break. An error path that emails you and also retries and also writes a row that the next run will treat as new. Two editors, two versions, no diff. The production workflow is the one that has not been opened in three months because opening it is how you break the thing that still mostly works.

n8n will let you put a Code node in the middle. That node is the confession. Once half the logic is JavaScript in a browser textarea, you have a program with worse packaging. Check it into git? Maybe. Test it? Not really. Review a change? Screenshot the canvas.

Make and Zapier rot the same way with prettier paths. The brand is not the disease. The disease is a requirement that became a state machine while the tool stayed a mural.

The change the canvas cannot draw

Idempotency: “run this twice, get one folder.” Canvas people add a lookup node. Then the lookup is eventually consistent. Then you get two folders on a retry. A program would use a unique key and a transaction. A canvas uses hope and a sheet tab named LOCK.

A join across two systems that do not share an ID. The first version keyed on email. The change is “email is not unique.” n8n can filter. It cannot give you a migration you will trust. You will write a one-off and leave it in a second workflow named fix dupes — DO NOT DELETE.

A human in the loop that is not a binary approve. “Fix the VAT number.” That is a form with validation and a timeout and a second actor. You can Rube Goldberg it. You will hate it in a month. This is when I open a small app or a Python service with a queue.

Compliance: “show me who changed the rule.” The canvas has a version history if you paid and if you remember to look. A git repo has a blame. Internal tools that touch money grow a need for blame. I have watched finance ask for that and the Zapier folder go quiet.

Ownership rot, not just logic rot

The first author of the workflow leaves. The canvas has no README that matches the nodes. The new person adds a branch instead of asking why the old branch exists. Six months later the company has two ways to create a customer and a Slack channel that still gets the old format. Git would have shown the archaeology. The canvas shows a maze.

Credentials rot on a different clock. A Google OAuth screen, a dead Slack token, a Stripe restricted key that someone rotated in the dashboard and not in n8n. The workflow is “fine” until a Monday. Low-code tools hide credentials in a vault that is not your password manager’s vault. When the person who clicked Connect is gone, you inherit a ghost login.

Naming rot: Production, Prod2, Copy of Production (2). I have deleted the wrong one. I now export and I treat unpublished copies as hostile. n8n’s UI will not save you from a human who is afraid to delete.

When n8n should stay

One person, a when/then pipe, no join that would embarrass a junior DBA. I have argued that n8n beats a weekend of custom CRUD for that shape. The rot starts when you promote the pipe to the company’s source of truth. Keep the pipe. Do not keep adding IF nodes as a culture.

Stay on the canvas if the requirement change is “also notify this channel.” Leave when the change is “also understand this entity over time.”

What I replace it with, and what I do not

A Python service with a queue and a real unique constraint. Apps Script when the data already lives in Google and the logic is still small. A thin CRUD when people need a screen all day. I do not replace a working three-node n8n with a platform rewrite because I am bored. I replace it when a change request takes longer to draw than to type, or when a retry would cost money.

n8n can call that service. The hybrid is healthy: canvas for glue, program for the rule. The rot is pretending the rule is glue.

How rot hides in green executions

The dashboard says 200 successful runs. Twenty of them created duplicate customers. Success meant HTTP 200, not “the business rule held.” Low-code tools are optimistic about success. Tests are pessimistic. Rot is optimism that aged.

I add a check that is not the workflow: a daily query for duplicate keys, a sum that must match Stripe, a folder count. When the check fails and the canvas is green, the canvas is no longer the tool. It is the suspect.

A week I will not pretend was the tool’s fault

We had a Zapier canvas that created Drive folders for new deals. Sales asked for “if the deal is a renewal, use the old folder and append a date, unless legal already opened a matter, in which case wait.” I drew it. It worked on the happy path. A renewal with two emails created two folders and a legal matter that pointed at neither. The canvas had no test for that triple. A fifty-line Python function with a unique deal_id and a status column did. The rewrite took less time than the fourth IF branch would have. I was angry at myself, not at Zapier. Zapier did the job I had originally named. I had renamed the job in a meeting and refused to rename the implementation.

n8n would have given me the same week with a different logo. The requirement change was the event. The drawing was the delay.

Teaching the next person

If I cannot pair-draw the workflow in fifteen minutes, it is already rotting. New hires should not need a tour of hidden sticky notes. A small module with a README is kinder than a generous canvas. I used to think the canvas was the documentation. It is a picture of last quarter’s understanding.

I now write the rule in a comment or a spec first. If the spec is a paragraph, canvas. If the spec is a list of unlesses, code. That filter has saved more Saturdays than a new n8n node.

A rule I use on Fridays

If I cannot explain the latest requirement without a nested “unless,” I stop drawing. I write the unless in code or I push back on the requirement. Drawing the unless is how internal tools become folklore.

Low-code rot is not a moral failure. It is a category error that felt like speed. n8n cannot draw a requirement that is now a program. You can keep asking it to. The graph will accept the nodes. The Thursday after a holiday, nobody will know which node is the tax ID. I have been that nobody. I write Python sooner now, and I leave the canvas for the sentences that still fit on one line. If your internal tool is a mural of unlesses, the next requirement will not fit in a node. It will fit in a function, or it will wait in a ticket until someone brave deletes half the graph. Be the person who deletes. The pretty picture was last quarter. The rule is this Thursday, and Thursday wants a diff.