n8n vs Zapier vs Make for AI Workflows

Most comparisons of these three tools list features. The thing that actually decides it for AI workflows is how each one counts usage, because AI workflows tend to have several steps and run often. So rather than a feature grid, this post takes one realistic AI workflow, maps it onto each tool, and works out what each pricing model charges for it. Then it covers the question most comparisons skip: when to stop using any of them and write code.

Prices below are from each vendor's official pricing page as of November 2026. They change, so check before you buy.

The example workflow

The workflow is a common first AI automation: triage incoming support emails. It has one trigger and three actions.

• Trigger: a new email arrives in the support inbox.
• AI step: a model classifies the email (billing, bug, sales, other) and drafts a reply.
• Action: create a ticket in the helpdesk with the category and draft.
• Action: post a summary to the team's chat channel.

Assume 2,000 emails a month. That is one run per email, so 2,000 runs.

How does each tool charge for the same workflow?

Each tool counts something different, and that difference matters more than the headline plan price.

Zapier counts tasks. Its pricing page says a task is a successful action step, including AI by Zapier and Code by Zapier steps, while triggers, filters, paths and failed actions do not count. AI steps are priced by model tier, so an AI step may cost more than one task. The calculation, assuming each AI step counts as one task:

• 3 action steps x 2,000 runs = 6,000 tasks a month (more if the AI step costs more than one)
• The Pro plan tiers listed are 2,000 tasks at $49/month and 10,000 tasks at $129/month on annual billing ($73.50 and $193.50 billed monthly)
• So this workflow needs the 10,000-task tier: $129/month annually, or $193.50 monthly

Make counts credits. Its pricing page says "most actions consume 1 credit", and a run uses one credit per module that runs. Make's credits documentation adds that AI modules using Make's own AI provider are charged by tokens and model. As one example from that page, Claude Sonnet 4.5 gives 301 input tokens and 60 output tokens per credit. If you connect your own AI provider account on a paid plan, you pay Make only for operations and pay the provider for tokens. With your own provider key:

• About 4 modules x 2,000 runs = about 8,000 credits a month, plus model tokens billed by your provider
• Through Make's AI provider instead, a prompt of 1,500 input tokens and a 300-token reply on Sonnet 4.5 would add 1,500 / 301 + 300 / 60 = about 5 + 5 = 10 more credits per run, or about 20,000 more a month
• Make prices plans by monthly credit volume; check the current price for your tier on its pricing page

n8n counts executions. Its pricing page says "An execution is a single run of your entire workflow", regardless of how many steps it has. The model is called with your own provider credentials, so tokens are billed by the provider.

• 1 execution x 2,000 runs = 2,000 executions a month
• The cloud Starter plan includes 2,500 executions at €20/month billed annually; Pro includes 10,000 at €50/month
• So this workflow fits Starter: €20/month annually, plus model tokens
• The self-hosted community edition has no licence fee; you pay for the server and your own time

The pattern is clear. Per-step pricing (Zapier) and per-module pricing (Make) scale with the number of steps, and AI workflows tend to gain steps over time: a validation step, a lookup, a fallback. Per-run pricing (n8n) does not.

What about AI features, ease of use and hosting?

Price is not the only factor. Who maintains the workflow and where your data can live often decide it.

On self-hosting, read the licence before you plan anything beyond internal use. n8n's Sustainable Use License says you may use or modify the software "only for your own internal business purposes or for non-commercial or personal use". Running it for your own operations is fine. Building a product that resells n8n workflows to your customers needs a closer look, and usually a commercial agreement.

Whichever tool you choose, set it up in your own account from the start. I covered why in who owns your automations when the freelancer leaves.

When should you move an AI workflow to custom code?

Move to code when the workflow stops being an internal shortcut and starts behaving like software. These are the signals I look for on client work.

• It is part of your product. If customers trigger it or see its output inside your app, it needs the same code review, version control and deployment as the rest of the product.
• You need real tests. Visual tools make it hard to run a set of saved examples through every change. For an AI step whose quality matters, an evaluation set run in code is far easier to maintain.
• The logic has outgrown the canvas. Deep branching, loops over large lists, retries with backoff and complex error handling are clearer as a few hundred lines of TypeScript or Python than as fifty boxes.
• Usage pricing dominates. At high volume, per-step or per-run charges can exceed the cost of a small service on your own infrastructure. Do the arithmetic above with your real volumes.
• Latency matters. If a person is waiting on the result, you want control over streaming, timeouts and caching.

You do not have to move everything at once. A common pattern is to keep the workflow tool for triggers and simple integrations and call a small code service for the AI step. That service can have tests and logging while the business still sees and edits the surrounding workflow.

If you are not sure which workflows deserve this effort at all, start with an AI automation audit.

Work with me

I build AI workflows in n8n, Zapier or Make when that is the right fit, and move them to code when they outgrow it. See workflow automation and AI agent development, or browse my work and book a scoping call.