n8n is enough when the work is a clear sequence of steps between apps (read an email, pull out a few fields, update the CRM, ask someone to approve), with one or two AI steps, moderate volume and someone technical to look after it. Custom code is worth it when the AI has to reason across many steps, every change needs tests and a review, volume is high, or the agent acts on money and customers. Many good setups use both: n8n for the plumbing, code for the hard part.
What does n8n do well (and where do Make and Zapier fit)?
n8n, Make and Zapier are workflow tools: you draw a chain of triggers and steps in a visual editor, and the tool runs it. They are very good at most of the plumbing of automation: connecting apps, moving data, schedules, retries and notifications.
- Zapier is the quickest way to connect two popular apps. Almost anyone can build and maintain a simple flow.
- Make gives a non-developer a visual canvas with branching, loops and a large set of ready connectors. It runs in Make’s cloud.
- n8n suits technical teams. It runs in n8n’s cloud or self-hosted in your own environment, it lets you drop into JavaScript or Python inside a step, and its AI nodes are strong.
The self-hosted option is n8n’s biggest practical advantage. The Community edition is, in n8n’s words, “the free edition”, the version that runs without a license key (n8n Docs). Your data stays on your servers, and the cost stays flat as volume grows, apart from hosting.
If your job is “when a form comes in, check it, write it to three systems and tell the team”, a workflow tool is the right answer. Writing custom code for that is slower, costs more, and leaves you with something only a developer can change.
Can n8n build AI agents?
Yes, and this is where n8n has moved furthest. Its agent features cover the building blocks you would otherwise write yourself:
- Agents with tools and memory. n8n’s docs describe agent tools as “workflows, custom code, built-in n8n integrations, and MCP servers”, plus session and episodic memory (n8n Docs).
- Human approval for tools. The workflow pauses until a person approves or denies a tool call, and the review can happen in Slack, Microsoft Teams, Gmail, Outlook, WhatsApp and other channels (n8n Docs).
- Evaluations. You can compare an AI workflow’s output against known test cases, from quick checks during development to metric-based scoring over time (n8n Docs).
- MCP both ways. An agent can call MCP servers, and the MCP Server Trigger node makes n8n workflows available as tools to MCP clients such as Claude (n8n Docs). Our guide to MCP explains what that means.
So “n8n can’t do AI agents” is no longer true. The real question is how much reasoning, testing and control your agent needs, which is what the rest of this guide is about.
What does n8n cost, and what does its license allow?
Pricing. n8n Cloud is priced by workflow executions, and an execution is one run of a whole workflow however many steps it has. On n8n’s pricing page in November 2026, Starter was listed at €20 a month billed annually for 2.5K executions, Pro at €50 for 10K, Business at €667 for 40K (self-hosted only), and Enterprise on request (n8n pricing).
What the free edition leaves out. The self-hosted Community edition does not include environments, external secrets, log streaming, SSO, sharing of workflows and credentials, or version control using Git (n8n Docs). On the pricing page, Git version control and multiple environments are listed from the Business plan (the comparison table shows them for Enterprise) and metric-based evaluations from Pro (n8n pricing). Those are exactly the features you miss once an automation becomes important, so price them in from the start.
The license. n8n is not open source in the strict sense. Its Sustainable Use License lets you use or modify the software “only for your own internal business purposes or for non-commercial or personal use”, and distribute it only free of charge for non-commercial purposes (n8n Docs). The license FAQ is practical: consulting and automation services are permitted, and you can charge for building and maintaining workflows, but you cannot host n8n as a service that lets your customers build their own workflows (n8n License FAQ). If you plan to put n8n inside a product you sell, read that FAQ before you design anything.
Where do workflow tools hit their limits?
None of these is a hard wall. They are points where the effort to keep a visual workflow reliable starts to exceed the effort of writing code.
- Complex branching. Ten nested IF nodes, loops inside loops and shared state between branches are hard to read on a canvas. Code with functions and types is easier to reason about at that size.
- Testing and evals. n8n’s evaluations help, but code lets you run hundreds of test cases on every change in CI and block a release that gets worse.
- Version control and review. Without Git in your edition, there is no reviewed diff: a change goes live when someone publishes it. Code changes are diffed, reviewed and can be rolled back line by line.
- Guardrails and approvals with nuance. “Approve every payment” is easy in n8n. “Approve payments over a threshold, from new suppliers, unless the PO matches” is easier to get right, and to test, in code.
- High volume. n8n can scale: its docs say queue mode “provides the best scalability” by running across multiple worker processes (n8n Docs). But at high, spiky volume you also want batching, back-pressure and cost per run under control, which is easier to tune in your own service.
- Multi-step reasoning. An agent that plans, calls tools, checks its own work and retries differently is possible in an agent node, but debugging why it chose step four is easier with your own traces and prompts in code.
- Observability. Run history in the editor is good for one workflow. Across dozens of workflows and agents you want structured logs, traces and alerts in the monitoring tools you already use, and log streaming is an Enterprise feature on n8n’s plans (n8n pricing).
What does custom code add, and what does it cost?
Custom code means a small service you own, written in Python or TypeScript, running in your cloud, calling the models and APIs it needs.
What it adds:
- tests and eval suites that run on every change, with a release blocked if quality drops;
- Git history, code review and rollback for every prompt, rule and integration;
- guardrails written as code: validation of every model output, thresholds, policy checks and approval rules of any complexity;
- logs and traces in your own monitoring stack, so you can see why an agent did what it did;
- any model, framework or system, including legacy databases and internal services no connector reaches.
What it costs: a developer to build it and someone technical to maintain it. Ops staff can no longer change a step themselves, small changes take a pull request, and a first release takes weeks instead of days. With us, a custom AI agent built around your systems starts from $8,000 at a fixed price, which is more than a workflow build. That is only worth paying when the limits above are real for you, not because code sounds more serious.
n8n or custom code: which should you choose?
| Your situation | Best fit | Why |
|---|---|---|
| Moving data between popular apps, a few rules | Zapier or Make | Fastest to build, anyone can maintain it |
| Data must stay in your cloud, technical ops person available | n8n, self-hosted | Free Community edition, flat cost as volume grows |
| One or two AI steps (classify, extract, draft) inside a clear process | n8n or Make | AI nodes are good enough; n8n has built-in approval steps |
| Agent that plans across many tools and steps | Custom code | Reasoning is easier to test, trace and debug |
| Every change must be tested and reviewed | Custom code, or n8n on a plan with Git | Release discipline matters more than editor speed |
| High, spiky volume where cost per run matters | Custom code | Batching, caching and model choice under your control |
| Acts on money, contracts or customers at scale | Custom code with approval gates | Nuanced guardrails and audit trails |
| Workflows your customers build themselves inside your product | Custom code, or an n8n commercial agreement | The Sustainable Use License does not cover it |
How do we build it: n8n, Make or code?
We do not have a favorite tool. Our Ops Agent costs $2,000 setup + $299/month and goes live in about two weeks, built on n8n, Make or custom code, whichever your team can maintain. After mapping the process with the people who do the job, we recommend one tool per workflow in writing and explain the trade-off.
For AI-heavy workflows, a hybrid is often the best fit:
- The workflow tool does the plumbing: triggers, schedules, connectors, retries, approval messages and run history. Your ops team can see and adjust it.
- Code does the hard step: reading messy documents, reasoning over several records, validating totals. It runs as a small tested service that the workflow calls, with its own eval set and Git history.
- Controls stay the same either way: anything irreversible waits for a named person, each workflow runs under its own least-privilege account, low-confidence results go to a person, and every run is logged.
Document-heavy jobs such as invoice intake are a good example. Our Document Agent ($3,000 setup + $399/month) extracts the fields and validates them in code (line totals, purchase-order match, supplier), while the surrounding workflow posts records and asks for approvals. You own every workflow: it lives in your account or your self-hosted n8n, documented so your team or another provider can maintain it. Our AI automation service covers the same approach for a set of workflows, plus rescuing automations that fail silently.
How do you migrate when you outgrow n8n?
You rarely need a rewrite. Move the part that hurts, keep the rest.
- List what you have. Export each workflow, note its owner, its triggers and which systems it touches. Unknown workflows are a bigger risk than any tool choice.
- Find the painful step. It is usually one node: the AI step that misreads documents, the branch nobody understands, or the run that times out at peak.
- Write the tests first. Turn past runs into test cases: the input and the output a person agreed was right.
- Move that step into code and call it from the existing workflow over HTTP or as an MCP tool. The trigger, approvals and notifications stay where they are.
- Run both side by side on live cases until the new step matches or beats the old one, then switch.
- Repeat only if needed. Many workflows never need to leave n8n, and that is fine.
What should you do next?
- Watch the hybrid working: our live workflow automation demo turns an order email into a priced invoice, a sheet row and a team update, and asks before anything goes to a customer.
- Compare packaged and custom prices on the pricing page, or price your own process in two minutes with the project estimator.
- Weighing up the build cost in general? Read how much an AI agent costs.
- Or book a free 30-minute call: bring one workflow, and we will tell you plainly whether n8n is enough for it.