n8n vs a custom AI agent: when a workflow tool is enough

Use n8n (or Zapier or Make) when the job is a fixed sequence of steps between tools, even with an AI step in the middle. Build a custom agent in code when the work needs judgment across many cases, has to act in your own systems with strict rules and approvals, or runs at a volume where testing, cost and reliability matter. Many teams use both: the workflow tool for plumbing, the custom agent for the hard part.

What is n8n good at?

n8n is a workflow automation tool with hundreds of integrations, a visual editor, AI nodes and the option to host it yourself. It is excellent for moving data between tools, scheduled jobs, and workflows with one AI step, like "read the email, pull out the order number, update the CRM". Our team has led AI operations teams that ran n8n automation, and for many jobs it is the right answer.

n8n vs Zapier vs Make: which workflow tool?

Check each vendor's pricing page for current plans; all three change often.

When is a workflow tool enough?

• The steps are the same every time, and you can draw them on one page.
• The AI's job is to read, sort or extract, and fixed rules decide what happens next.
• Volume is moderate, and an occasional failed run can be retried by hand.
• The systems involved have ready-made connectors.

That covers most "we need an AI agent" requests we see. Our agent vs workflow guide has a three-question quiz to check.

When do you need a custom agent in code?

• Judgment on every case. The right next step depends on details, and the exceptions are the job.
• Your own systems. It has to act inside your ERP, product database or software with no connector, with its own permissions.
• Real rules and approvals. Limits by amount, customer and action, enforced in code and logged.
• Testing and versioning. You need to run the agent against hundreds of past cases before each change, and roll back.
• Volume and cost. At thousands of cases a day, per-execution pricing, latency and model costs need engineering.

Can you use n8n and a custom agent together?

Yes, and it's often the best setup. n8n handles triggers and plumbing: new email, new file, nightly sync. It hands the hard case to a custom agent through an API or webhook, and gets back a decision or a draft for approval. The agent can also expose its tools through the Model Context Protocol (MCP), so other assistants can use them.

What usually goes wrong with workflow-tool agents?

The workflow grows into a maze of branches as every exception gets its own path, and nobody can tell what it will do with a new case. Or an "AI agent" node gets a long prompt full of rules, and starts skipping some of them. Both are signs that the job needs judgment that a fixed flow can't hold, and it's time to move the hard part into a tested agent.

How do you decide for your workflow?

Write down the last 20 cases a person handled and what they did with each. If the actions fit in a short list of fixed paths, build it in a workflow tool. If every few cases needs a different call, you need an agent. We do this sorting in the first days of our First Agent Sprint.

Common questions

Is n8n secure enough for company data?

It can be, especially self-hosted, but security is in how you set it up: credentials, access and where the data goes.

Can n8n build real AI agents?

It has agent nodes that can call tools and models. They work well for contained jobs; complex agents with strict rules usually outgrow them.

Is a custom agent much more expensive?

It costs more to build and less to change safely at scale. See what a custom AI agent costs.