Zapier, Make, n8n or Power Automate — and when none of them beat writing the integration.

Every comparison of these tools opens with a feature matrix, and every feature matrix says roughly the same thing: they all connect apps, they all have branching, they all have an AI node now. That matrix is why so many businesses pick the wrong one. What actually separates these platforms is the unit they bill you for — a step, an operation, a whole workflow run, or a seat — because that decision compounds every single time your automation fires. Below is one realistic workflow priced on all four, then where each one’s AI genuinely sits, and finally the honest answer to the question underneath all of this: whether you need an automation platform at all when an AI assistant can already read your inbox.

The same workflow, priced four ways

One automation with ten billable action steps, firing a thousand times a month — a lead lands, gets enriched, written to a CRM, tagged, notified, logged. Identical work on every platform. These are the list prices that job lands on.

What each one is actually for

Prices are August 2026 list prices on annual billing, in the currency each vendor publishes. App and integration counts are the vendors’ own figures.

The billing unit is the decision

Zapier and Make both count steps. n8n counts runs. Microsoft counts people. Everything else in this comparison is a detail that changes twice a year; the billing unit is structural, and it decides which platform gets cheaper as you succeed and which one punishes you for it.

Work it through once and it becomes obvious. A ten-step workflow firing a thousand times consumes ten thousand billable units on Zapier and Make, and one thousand on n8n. Add three steps because the process got smarter and Zapier and Make charge you 30% more for the same thousand outcomes, while n8n charges the same. Now run it ten thousand times instead: on n8n you move up one tier, on Zapier you are into six figures of tasks.

The counter-argument for Zapier is real and it is not about price: 9,000+ connected apps. If the tool you need to reach is in Zapier and nowhere else, the comparison ends there, and paying $129 to avoid building an integration yourself is a good trade. Just make that decision knowingly, because the default assumption — that these platforms are broadly interchangeable and you should pick on interface preference — is what produces a $400 monthly bill for a workflow that should have cost $9.

Which of them is genuinely built for AI

All four have shipped AI features, and they are not the same kind of thing. Zapier gives you Agents on every plan and AI steps that call a model without needing your own API key — useful, and firmly inside a linear Zap. Make has AI Agents in beta across all plans plus Maia, a conversational builder that assembles scenarios from a description. Microsoft sells the agent layer separately as Copilot Studio, at $200 for 25,000 credits a month.

n8n is the one built around it rather than beside it. Its AI Agent node is a LangChain integration: you attach a chat model and a set of tools, and the agent decides which tool to call and in what order, looping until the task is done. That is a materially different shape from “send text to a model, take the answer, continue” — and it is exactly the workload that per-step billing punishes, since a single agent run can call four tools and the model three times. On n8n that is one execution.

This is the practical reason n8n has taken so much ground with AI-heavy teams, and it is a pricing consequence more than an ideological one. If your automation is deterministic — form to CRM to Slack — the AI nodes are a rounding error and you should pick on catalogue and price. If the automation genuinely needs a model to decide something on each run, the per-execution model is the one that survives contact with the invoice.

The n8n licence footnote nobody reads

n8n is source-available, not open source. The 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 permits distribution only free of charge for non-commercial purposes. n8n calls this fair-code; it is not an OSI-approved open-source licence, and the distinction has teeth if you ever intend to resell an automation product built on it.

For the ordinary case — a business self-hosting n8n to run its own workflows — this is entirely fine, and that is precisely what the licence is written to allow. What it does not allow is offering n8n itself to your customers as a service. Worth knowing before it becomes a legal conversation rather than a technical one.

The other honest caveat about self-hosting: “free” means no licence fee. It does not mean free. Someone has to patch the server, watch the queue, restore from backup when a workflow silently stops firing, and notice that a credential expired. That is a real recurring cost, and for a small team it is frequently higher than the €20 a month the Cloud tier would have cost.

When Power Automate is the right answer

In two situations, and they are both narrow. First, when the business already runs on Microsoft 365 — the licensing is per user rather than per run, the connectors into SharePoint, Teams, Outlook and Dataverse are first-party, and the governance story is one your IT department already accepts.

Second, and more interestingly, when there is no API at all. Power Automate desktop flows are robotic process automation: they drive the user interface of applications directly, including “legacy applications, such as terminal emulators”, using UI elements, images or coordinates. When the system holding your data is a twenty-year-old desktop program with no integration surface, RPA is not a compromise, it is the only mechanism that exists. Attended desktop flows come with Premium at $15 per user per month; unattended robots start at $150 per bot per month.

Outside those two cases, the per-seat model works against you. Automation should get cheaper per outcome as it scales, and paying by headcount for work that removes the need for headcount is a strange shape to sign up for.

So why not just use AI for all of it?

This is the question worth taking seriously, because a modern assistant can already read a document, extract the fields, draft the reply and call your tools through MCP — the open standard that lets AI applications connect to external systems, described by its own documentation as “a USB-C port for AI applications”. Zapier exposes an MCP endpoint, Make ships an MCP server, n8n has MCP nodes. The connective tissue argument is over; assistants can reach your systems.

What an assistant does not have is everything around the intelligence. It has no trigger — something has to invoke it. It has no schedule, no retry when an API returns a 503, no idempotency when the same webhook arrives twice, no queue when four hundred orders land in an hour, no audit log showing what ran and what it did. An automation platform is mostly not the clever part; it is the boring, reliable part that makes the clever part run unattended at three in the morning without anyone watching.

Anthropic’s own engineering guidance draws the same line and is unusually blunt about it: workflows are systems where models and tools follow “predefined code paths”, agents are systems where models “dynamically direct their own processes and tool usage”, and “agentic systems often trade latency and cost for better task performance”. The recommendation is to find “the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.” Coming from a company that sells the models, that is worth more than any vendor comparison chart.

The practical rule: use a deterministic workflow when the rules can be written down, and reach for a model only at the step where genuine judgement is required — classifying an enquiry, extracting fields from an unpredictable document, drafting a first reply. A model wired into every step of a process that had clear rules is slower, more expensive and less reliable than the if-statement it replaced.

When none of them is the right answer

There is a third path these comparisons never mention, because none of the vendors benefit from it: writing the integration. Once a workflow is high-volume, business-critical and stable, a small piece of code calling two APIs directly costs nothing per run, has no seat licence, cannot be repriced by a vendor, and does not break when a platform deprecates a connector.

The honest threshold is somewhere around “this runs constantly and will still be running in three years”. Below it, a platform wins easily — you are paying for a thousand integrations you did not have to build and a UI that lets a non-developer see what happened. Above it, you are renting a permanent tax on something you could own. That is the point at which API & Integrations work costs less over two years than the subscription does.

The mistake to avoid is the reverse: building a custom integration for a process nobody has validated yet. Prove the workflow on a platform first, where changing it costs an afternoon, and rewrite it once the shape has stopped moving.

How to choose in ten minutes

Count your steps and your monthly runs, then multiply. If steps times runs is under a few thousand, the price difference between all of them is noise and you should pick on which app catalogue covers your tools — usually Zapier. If it is in the tens of thousands, the per-execution model is worth real money and n8n starts winning on arithmetic alone.

Check the connector before anything else. The cheapest platform that cannot reach your CRM is infinitely expensive. Zapier lists 9,000+ apps, Make around 3,000, n8n roughly two thousand integrations plus an HTTP Request node and custom code nodes for anything it does not cover natively.

Ask whether a model has to decide something on each run. If yes, per-execution billing and native agent nodes matter, and n8n is the default. If no, ignore the AI marketing entirely — it is not what you are buying.

Ask where the data has to live. If customer records cannot leave your infrastructure, self-hosted n8n is the only one of the four that answers that cleanly.

How we set this up in practice

For a single, well-defined workflow — form submissions into Slack, a CRM record into an email sequence — Zapier / Make Automation is €299 and takes two business days: scoping, trigger mapping, error handling, tested end-to-end in your live environment, plus a walkthrough so it is not a black box. The platform subscription is yours, not ours.

For something with more moving parts, Workflow Automation is €699 and starts with a one-hour process-mapping workshop, then builds in n8n or Make across two or three of your existing tools, with error alerting, documentation and two weeks of monitoring after go-live — because real, messy data breaks automations that pass every demo.

We choose the platform after the mapping session, not before, and we will happily tell you the workflow does not need one. If the process runs twice a month, the automation costs more in subscription and maintenance than the manual work does, and we would rather say that than sell you a retainer. Our B2B process automation guide covers which processes are worth automating in the first place.

Sources

Every price, billing definition and feature claim above was taken from the vendors’ own pages in August 2026. Prices change — check them before committing.

• Zapier — Pricing ↗

  The Professional task ladder ($19.99 for 750 tasks up to $3,389 for 2 million, annual billing) and Zapier’s own definition of a task: counted when a unit of work completes successfully, with failed actions, triggers, filters, paths and Formatter excluded. Also states 9,000+ connected apps, Zapier Agents on all plans and MCP availability.

• Make — Pricing ↗

  Free 1,000 credits, Core $9, Pro $16 and Teams $29 per month at the 10,000-credit tier, and the operation definition: “Each module action in your scenario, like adding a Google Sheet row or fetching Gmail account data, counts as one credit.” Lists 3,000+ apps, 350+ AI apps, AI Agents in beta and an MCP server.

• n8n — Pricing ↗

  Starter €20, Pro €50 and Business €667 per month on annual billing, with 2,500 / 10,000 / 40,000 workflow executions and unlimited steps, users and workflows. States plainly that n8n charges “for full executions, not for each step” — one run of the whole workflow, whatever it contains.

• n8n Docs — AI Agent node ↗

  The LangChain-based agent node: attach a chat model plus one or more tools and the agent decides which tool to call to complete a task, with Tools, Conversational, ReAct, SQL and Plan-and-Execute variants.

• n8n — Sustainable Use License ↗

  The licence text itself: use or modification is permitted “only for your own internal business purposes or for non-commercial or personal use”, and distribution only free of charge for non-commercial purposes. Source-available, not OSI open source.

• Microsoft — Power Automate pricing ↗

  Power Automate Premium at $15 per user per month paid yearly (cloud flows plus attended desktop RPA and Dataverse entitlements), Process at $150 per bot per month, Hosted Process at $215, and Copilot Studio at $200 for 25,000 Copilot credits per month.

• Microsoft Learn — Introduction to desktop flows ↗

  Defines desktop flows as robotic process automation that drives application interfaces directly — including “legacy applications, such as terminal emulators” — using UI elements, images or coordinates. This is the capability the other three platforms do not have.

• Anthropic — Building effective agents ↗

  Distinguishes workflows (models and tools following “predefined code paths”) from agents (models that “dynamically direct their own processes and tool usage”), notes that “agentic systems often trade latency and cost for better task performance”, and recommends “finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.”

• Model Context Protocol — Introduction ↗

  The open standard for connecting AI applications to external systems — “like a USB-C port for AI applications” — now supported across Claude, ChatGPT, VS Code and Cursor, and exposed by Zapier, Make and n8n alike.