Every growing company eventually hits the same wall: leads arrive in one tool, invoices in another, support tickets in a third, and people copy data between them by hand. Automation platforms promise to fix that, and in 2026 all three market leaders also promise AI agents. The choice between n8n, Make and Zapier looks like a matter of taste. It is not. The platforms differ in how they bill you, where your data lives, how far AI steps can go and who can maintain the result a year later.
This guide compares the three on the criteria that decide real projects: billing units, AI capabilities, data protection, scalability and total cost of ownership. At the end you will find a selection process you can run in a week and the mistakes we see most often.
The short answer
If you only read one section, read this one:
- Zapier is the fastest way for a small team to connect popular SaaS apps without technical staff. It becomes expensive when workflows have many steps or run thousands of times.
- Make fits teams that think visually and need branching, loops and data transformation, run by an operations person rather than a developer.
- n8n is the default for engineering-led teams, AI-heavy workflows, high volumes and companies that need control over where data is processed. It rewards technical skill and punishes the lack of it.
The rest of this article explains why, with numbers.
What changed in 2026
Three shifts make this comparison different from the one you might have read two years ago.
n8n became the agency standard. Its valuation rose from $2.5 billion in October 2025 to $5.2 billion after a strategic investment by SAP in May 2026, and n8n is being embedded into SAP Joule Studio. The company reports 1.7 million monthly active developers and more than 1,400 enterprise customers (Tech.eu). Revenue is estimated at around $100 million ARR (Sacra).
Every platform now sells AI agents. Zapier added Agents and an MCP endpoint, Make advertises AI agents and agentic automation, and n8n ships agent and LLM nodes. At the same time, analysts warn that the label is used loosely: Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 and estimates that only around 130 of the thousands of vendors calling themselves agentic offer real agency (SD Times). Menlo Ventures found that only 16% of enterprise deployments are true agents; the rest are fixed chains of steps or routing around a single model call (Menlo Ventures).
MCP became the standard way to connect agents to systems. The Model Context Protocol was handed to the Agentic AI Foundation under the Linux Foundation on 9 December 2025, and there are now more than 10,000 public MCP servers (MCP blog). Automation platforms are becoming tool providers for agents, not just pipelines. We explain the protocol in MCP explained: connecting AI agents to your systems safely.
How each platform bills you
The billing unit is the most underrated difference between the three. It decides whether a workflow costs €20 or several hundred a month.
| Zapier | Make | n8n | |
|---|---|---|---|
| Billing unit | Task: each successful action step | Credit: most actions cost 1 credit | Execution: one complete workflow run |
| What does not count | Triggers, failed actions, Filter, Formatter, Paths, Delay and other built-in tools | Routers and error handlers | Number of steps and data volume |
| Free tier | 100 tasks/month, two-step Zaps, 15-minute polling | 1,000 credits/month | Community Edition, self-hosted |
| Entry paid plan | Pro from $19.99/month (annual) for 750 tasks; Team from $69/month for 2,000 tasks | Core, Pro and Teams, priced by credit volume in the plan configurator | Starter €20/month (annual) for 2,500 executions; Pro €50 for 10,000 |
| AI-specific billing | AI steps priced by model tier; Agents billed separately in activities (400 free, 1,500 on Pro from about $33.33/month) | AI features through Make’s AI provider may use more than 1 credit | Model costs paid directly to your LLM provider |
| Overage | Pay-per-task at 1.25× (annual) or 2.5× (monthly) the base rate, pause at 3× | Extra credit bundles of 1,000 or 10,000 | Business plan: €4,000 per 300,000 extra executions |
Sources: Zapier pricing, Make pricing, n8n pricing. Prices as published in October 2026; annual billing assumed where marked.
Worked example: one workflow, three bills
Take a typical AI workflow for inbound leads: a form submission triggers a lookup in the CRM, an LLM classifies the request, a second LLM call drafts a reply, the lead is enriched, a record is updated, a Slack message is sent and a follow-up task is created. That is roughly nine action steps. It runs 5,000 times a month.
- Zapier: about 9 tasks × 5,000 runs = 45,000 tasks. That lands on the 50,000-task Pro tier, $289 per month on annual billing, before AI step pricing.
- Make: about 9–10 credits per run, so 45,000–50,000 credits plus any extra credits consumed by AI modules. Check the current price for that volume in Make’s configurator.
- n8n: 5,000 executions, which fits the Pro plan at €50 per month, or a self-hosted instance where you pay only for the server and your time.
The difference grows with every step you add. Agentic workflows tend to have many steps and loops, which is exactly where per-step billing hurts. In all three cases you still pay the LLM provider; tokens are covered in LLM gateway: model routing, cost control and observability.
AI capabilities compared
All three can call an LLM, parse its output and branch on it. The differences are in depth and control.
Zapier has the most polished entry point: an AI step inside a Zap, a Copilot that builds Zaps from a description, Chatbots, Agents and an MCP endpoint that lets external AI assistants trigger Zapier actions. Enterprise customers can bring their own model via AWS Bedrock. The trade-off is limited control over prompts, retries, evaluation and logging, and a cap of 40 activities per agent run on paid plans.
Make puts AI modules and agents into its visual canvas. That is a strength for operations teams who want to see every branch. For complex agent logic, scenarios can become hard to read and test, and AI modules may consume more credits than standard actions.
n8n offers LLM and agent nodes, memory, tool calling, vector store integrations and a code node in JavaScript or Python. You can run your own models, call any HTTP API and expose workflows as tools. That is why agencies building production agents default to it. It also means more decisions are yours: how to version prompts, how to test, how to restrict what the agent may do.
A useful rule: the more an AI step can change data or talk to customers, the more you need versioned prompts, test sets and logs. That favours the platform your team can engineer properly, not the one with the nicest demo.
Data protection, hosting and EU compliance
For European companies, this section often decides the matter. In Germany, 66% of companies already using AI name data protection requirements as a barrier, 56% legal uncertainty and 54% unpredictable costs (Bitkom).
- n8n states that its hosted plans store data within the EU on servers in Frankfurt, Germany. Self-hosted, data stays wherever you run it, including your own EU servers. Note that the self-hosted edition is source-available under n8n’s own fair-code license, not a classic open-source licence, and that self-hosted Business and Enterprise plans need a license key that checks in daily.
- Make and Zapier are cloud services. Their pricing pages do not detail data residency, so check the DPA, the subprocessor list and the hosting region with each vendor before you route personal data through them.
- For all three, every LLM provider you call becomes a subprocessor. Document it.
If your automation talks to customers, for example a chat or email assistant, Article 50 of the EU AI Act requires disclosure that people are interacting with AI since 2 August 2026 (AI Act Explorer). Our guide to GDPR for AI chatbots and assistants covers DPAs, logs and data residency in detail.
Who should pick what
| Your situation | Best fit | Why |
|---|---|---|
| Small team, no developer, popular SaaS tools, under a few thousand runs a month | Zapier | Fastest setup, largest app catalogue, built-in Copilot |
| Operations team, complex branching, data transformation, visual thinkers | Make | Visual scenarios, routers and error handling do not cost credits |
| AI-heavy workflows, many steps, thousands of runs | n8n | One execution per run, full control over LLM calls |
| EU personal data, need to choose hosting location | n8n (self-hosted or EU cloud) | Data location under your control |
| Product team embedding automation into its own SaaS | n8n or custom code | Check the n8n license before reselling; custom code for core features |
| Customer-facing agent with refunds, bookings or payments | Custom agent plus any platform as glue | Needs evals, permissions and monitoring beyond any no-code tool |
When none of them is the answer
Automation platforms are excellent glue. They are weaker as the core of a business-critical AI product. Warning signs:
- the workflow makes decisions with money, contracts or health data;
- you need regression tests for prompts and model versions;
- latency matters, for example in voice;
- the logic has grown into dozens of nested branches nobody can review.
In these cases the usual pattern is a small custom service for the agent itself, with the platform handling triggers, notifications and simple integrations. Market prices reflect that: a single automation on n8n or Make with an LLM typically costs $3,000–15,000 to build, while a custom AI agent starts around $10,000–50,000 (TheCrunch, Lets-Viz).
A practical selection process in five steps
- List your top five workflows with steps per run, runs per month and the data each touches. This alone usually narrows the choice.
- Calculate the bill on each platform using its billing unit, then add LLM costs. Use the worked example above as a template.
- Classify data. Anything with personal or financial data needs a DPA, a subprocessor list and a hosting decision before the first test.
- Decide who maintains it. A platform your team cannot debug at 2 a.m. is the wrong platform. If nobody can, budget for a retainer: the most common 2026 model is around $5,000 to build plus about $1,000 a month to run (TheCrunch).
- Build one workflow end to end on the shortlisted platform, with logging and an error path, before migrating anything else.
Typical mistakes
- Choosing by app count. A catalogue of thousands of connectors does not help if the one system that matters, your ERP or industry software, needs a custom API call anyway.
- Ignoring per-step billing. Teams prototype on a free tier, then discover that the production version with error handling and AI steps costs ten times more.
- Putting prompts inside nodes without versioning. Prompts change behaviour as much as code does. Keep them in version control and test them.
- Letting the agent act with admin rights. Give each workflow the narrowest credentials it needs. Prompt injection is still unsolved: even OpenAI says it may never be fully eliminated (analysis).
- Self-hosting without an owner. Unpatched, unmonitored automation servers are a security problem waiting to happen.
- Building everything in one platform forever. Keep business logic portable. Our article on owning your AI and avoiding vendor lock-in shows how to plan an exit.
Conclusion
There is no universal winner. Zapier buys speed for small teams, Make buys visual control for operations, and n8n buys cost efficiency, data control and engineering depth. For most AI-heavy workflows in Europe, n8n is the sensible default. Zapier stays the fastest start when volume is low and data is not sensitive. Whatever you choose, the success of an AI automation depends less on the platform than on clear metrics, narrow permissions, tests and someone who owns it after launch.
If you want a second opinion on your workflows, we design and run automations on all three platforms and build custom agents where they are needed, through our AI agents and automation and systems integration teams. Book a free 15-minute automation audit and we will tell you which platform fits your volume, data and team, even if the answer is the one you already use.