n8n vs custom code: a task-shape decision

n8n or custom code: the task decides, not the tool

Use n8n when the task is repetitive and deterministic; write code when it is not. That is the whole verdict, and it is a condition rather than a number: no items-per-day threshold or budget line survives contact with real workflows, but the shape of the task does. For repetitive, deterministic work - the same steps, in the same order, on inputs whose form you can predict - n8n or Zapier are perfectly good, and this studio says so against its own interest: it builds custom AI systems for a living and still routinely works with both tools itself. Connector tools fail when the problem or the workflow is non-deterministic - when a step requires judgment about what arrived rather than plumbing to pass it along. And there is a second, slower condition that catches up with organisations that automate seriously: as an org becomes more AI-native, the cost of paying for connector tools overtakes the cost of building in-house. This page walks the three legitimate options and the line between them, applying the same judgment that runs through the rest of this studio’s AI writing.

The three options, compared on what actually differs

The buyer at this decision is not choosing between two things but three: a connector tool, a plain script in software already owned, or a real system built in code. Most published comparisons collapse the middle option, which is a mistake - it wins a whole category of tasks outright.

The last row is not an afterthought. A real system changes hands with documentation, training and tests, and how its output gets evaluated - and what happens when its success recruits more scope - are each deep enough subjects to have their own pages in this territory.

When to choose each

Each option wins a real category of work and fails in a characteristic way. The two worked cases below are the studio’s own examples, from the same operational territory, on purpose: the line between them is the line this page is about.

When a connector tool wins

n8n or Zapier wins when the workflow is repetitive and deterministic and spans several tools: a form submission that becomes a CRM row, a payment that triggers an invoice, a file drop that fans out into notifications. The steps are fixed, the inputs are predictable, and the tool’s whole value is that nobody writes or maintains integration code. This studio reaches for these tools itself when the task has that shape - paying for custom development there would be buying expertise the problem does not need. The failure mode is the first non-deterministic step: a node that must judge what a document means, decide which of several unclear cases applies, or handle inputs whose form keeps changing. The canvas assumes predictable outputs wired to predictable inputs, and stretching it over judgment work produces the familiar dead workflow - branch piled on branch, still wrong on the cases that matter. There is also a slower failure that is financial rather than technical: connector pricing scales with usage, so an organisation that keeps getting more AI-native keeps paying more for plumbing it could by then own - the point at which that cost overtakes the cost of building in-house is the point at which the subscription has become the wrong purchase.

When a plain script wins

The most under-recommended option in this comparison costs nothing at all. A well-maintained daily PnL sheet that needs a daily summary sent to leadership is Google Sheets plus GScript, for free - nothing else needed. No connector subscription, no build engagement, no new tool for anyone to learn: the data is already structured, the logic is fully deterministic, and the script lives inside software the organisation already runs. A surprising share of automation asks have exactly this shape, and the honest answer to them is an afternoon of scripting, not a purchase. The failure mode is scope: scripts are written by whoever needed them and maintained by nobody, so a script that quietly grows into the thing a workflow depends on becomes an unowned dependency - and the moment the task stops being deterministic, a script fails the same way a connector tool does, only with less visibility when it breaks.

When custom code - a real system - wins

Code wins when the problem is non-deterministic: when the work is judgment, not plumbing. Reconciling invoices, purchase orders and receivables against cash-in-bank and working capital is the studio’s own example - the inputs arrive in inconsistent forms, the matching requires judgment, the exceptions are the actual work, and what the organisation needs is what the founder describes as a ten-times-more-powerful micro-CFO on call. No connector canvas and no sheet script produces that, because the value is not in moving data between tools but in reasoning over it. A system with that job needs to be built, tested against the cases it must not get wrong, documented, and handed to a trained team - which is why it is an engagement rather than a subscription. The failure mode runs in the other direction: building a real system for a task a connector tool already handles is paying deep expertise for shallow work, and it is the first situation in the refusal list below - the studio turns that work down rather than take the fee.

n8n vs Zapier is the wrong question

Asked which of n8n or Zapier is better, this library gives the answer it gives about every tool pairing: it does not rank vendors, because the studio builds on these tools and a vendor ranking from a builder is marketing wearing a review’s clothes. What matters for the buying decision is that both are connector tools, and they sit on the same side of every line this page has drawn - right for repetitive, deterministic workflows, wrong for non-deterministic problems, and subject to the same subscription arithmetic as an organisation’s automation grows. Choosing between them is a configuration-level decision about integrations, hosting and pricing that an afternoon of reading settles; choosing whether a connector tool is the right category is the decision that moves money. The same reframing answers the AI workflow question. n8n can orchestrate AI workflow automations, and the published examples - summarise this, classify that, draft a reply - work exactly when the AI step behaves deterministically enough to wire into a canvas. The test to apply to any such example is the one this page opened with: if every step’s output is predictable enough that the next step never needs to exercise judgment, the connector tool is the right home for it. The moment that stops being true, you are no longer configuring a workflow - you are operating an AI system, and it deserves to be built, tested and handed over like one.

The part most pages leave out

When not to choose Foxnut Studios

Situations where another option is the better call, and where we say so in the first conversation rather than the fourth.

• The client wants a superficial AI solution - something that could be bought off the shelf or configured in an afternoon, and does not require the studio's expertise. An off-the-shelf tool, or a generalist implementer. If the problem does not need deep expertise, paying for deep expertise is the wrong purchase - the studio says so and points at the simpler option.
• The client's senior leadership does not have conviction in making AI work for the organisation. Nobody, yet. Without conviction and commitment from the top, these projects tend to fail despite the best implementation - the honest move is to decline until leadership is committed, not to build something that will be abandoned.
• Some of the organisation's data is offline or simply undocumented. The organisation itself, doing the documentation work first. Data that is unstructured and scattered across many tools and databases is workable - the studio works with that routinely. Data that exists only in someone's head or on paper is not, and no consultant can fix that from outside.

Foxnut Studios works on briefs like this one from Bengaluru and Paris. If you want the shape of that before you talk to anyone, here is what an AI engagement covers and what you keep.