n8n, Make or Zapier vs a Custom AI Agent: When to Switch

Key takeaways

• One five-step lead workflow costs 6–7 Zapier tasks, 5+ Make credits or 1 n8n execution per run: compare your workflow's monthly bill, not plan prices.
• At 1,000 leads a month that workflow fits Make Core at $12, n8n Starter at €20 or Zapier Professional at $129 (October 2026, billed annually).
• Self-hosted n8n is free for internal use; you need a commercial license once outside users build or configure workflows, including through an AI agent.
• Switch to a custom agent when two or more signals appear: long-term memory, many tools, real human handoff, data residency, per-run cost at volume, tests and versions.
• Hybrid is often the answer: n8n keeps schedules and integrations, and only the step that needs judgment moves into a service with its own tests.

n8n, Make and Zapier are enough while your automation is a fixed chain of steps with a model call or two inside. Move to a custom AI agent when it must remember context across conversations, choose among many tools, hand cases to people by design or keep data on your servers, or when per-step fees outgrow hosting: Zapier Professional costs $489 a month for 100,000 tasks, billed annually.

By AI agent we mean software where a language model chooses the next step and calls tools, such as your CRM, instead of following a fixed path. We build both kinds: a content factory on 40+ n8n nodes and AI call notes in our own code for Encore, a Dubai agency.

Inside: October 2026 prices, one workflow counted in each platform's billing unit, n8n's license rules, six signals to switch and a decision checklist.

n8n vs Make vs Zapier vs a custom agent. List prices from the official pricing pages of Zapier, Make and n8n as of October 2026, billed annually. n8n shows prices in euros or dollars depending on where you open the page; in dollars, Starter is $20, Pro $50 and Business $800.

n8n vs Make vs Zapier: why do price comparisons mislead?

Because each platform counts a different thing. Zapier counts actions, Make counts module operations including the trigger, and n8n counts whole runs, so one workflow can be 6 units on one platform and 1 on another. Compare what your workflow costs per month, not what a plan costs.

The definitions, from each vendor:

• Zapier task. "A task is any successful action that runs in Zapier," and "Zap triggers never use tasks," per Zapier's help center (updated August 21, 2026), and the pricing page agrees: "Triggers don't count toward your task limit." Zapier's task usage rates, though, list "Trigger and action steps" at "1 task per step", so budget one extra task per run in case the trigger is counted. Built-in tools such as Filter, Paths, Formatter and Looping use 0 tasks. An AI by Zapier step costs 1, 3 or 5 tasks depending on the model tier, per Zapier's task usage rates.
• Make credit. "Every time a module in a scenario performs an action, it consumes credits," and reading data from an app or webhook counts. Most actions use 1 credit; some features on Make's AI Provider use more. Router and error-handler modules are free.
• n8n execution. "An execution is a single run of your entire workflow." The pricing page adds: "It doesn't matter how many steps are in the workflow or how much data it processes."

One lead-qualification run, counted three ways

Take a common workflow. A web form sends a lead, the workflow finds the contact in the CRM, an AI step scores the lead and drafts a reply, the CRM record is updated and the sales chat gets a message. Five steps, one run per lead.

One run of the workflow above, counted by each vendor's published rules, and the plan it needs at 1,000 leads a month. Prices as of October 2026, billed annually. The step breakdown is our assumption, not a vendor figure.

On this shape Make is the cheapest. Change the shape and the order moves. Stretch the same flow until each run uses 20 tasks and Zapier needs 20,000 tasks a month, $189 on Professional, while n8n still counts 1,000 executions. Run a workflow on every chat message instead of every lead, and n8n's per-run count climbs, while on Zapier the actions a Filter step skips use no tasks.

None of these lines is the whole model bill. On Zapier you can bring your own model at 1 task per AI step and pay the model provider directly, and on any platform a model you connect with your own key is billed per token. Our AI agent cost guide works that token formula on official price lists.

Can you self-host n8n for commercial use?

Yes, as long as n8n stays your internal tool. The self-hosted Community Edition is free under n8n's Sustainable Use License, and you need a commercial license when people outside your company build or configure workflows themselves, or when n8n itself is what you sell.

"You may use or modify the software only for your own internal business purposes or for non-commercial or personal use." — n8n, Community license

The n8n license FAQ says this license "is free and applies to the self-hosted version of n8n only" and draws the line in plain terms:

• Allowed: running your own business on it, as long as only your people create or modify workflows; building automations for clients on your instance "as long as your clients do not have the possibility to create or edit them"; using n8n behind the scenes in your own product.
• Not allowed: hosting n8n as a service where clients build workflows; white-labeling it; letting external users build or configure workflows "whether via a custom UI, our API, MCP, or an AI agent acting on the user's behalf."

The FAQ sums up the test: "The SUL test is not about UI visibility — it is about who controls the workflow logic and whose commercial interests are being served." The AI agent clause is the one product teams miss. If your customers describe an automation in chat and an agent assembles it on n8n, the customer sets the workflow logic, and the free license no longer covers it. If customers only use agents you built and receive their outputs, it does. Paid features such as Git version control on the Business plan cost money either way.

When is it time to move from n8n to a custom AI agent?

When the agent stops being a step in a workflow and becomes a system with its own state, tools and people around it. These are the six signals we check. One alone rarely justifies a rebuild; two or more usually justify at least a hybrid.

1. State and memory across conversations. Agent memory in a workflow tool keeps a thread in context. When the agent must recall a customer's history for months, and you need to search, correct or delete that memory, it belongs in a database you control.
2. Many tools. Each tool is a description the model reads on every call and one more way to fail. Past a handful, you want per-tool permissions, logs and tests, which are easier to keep in code.
3. Human handoff by design. An alert is not a handoff. A real one has a queue, an owner, a timeout and an agent that resumes once a person answers. n8n lists human approval for tool calls on its pricing page; an inbox with service levels is product work.
4. Data that must stay in your infrastructure. Zapier and Make run in their own clouds. Self-hosted n8n or your own service keeps workflow data on your servers, though the model provider still receives the prompts unless you host the model too.
5. Per-run cost at volume. When the monthly platform fee exceeds a server plus a developer's upkeep hours, the math has flipped. Zapier Professional at 100,000 tasks is $489 a month; at 6–7 tasks per run, that is about 14,300–16,700 runs.
6. Testing and versioning. Agents still fail often enough to need a test set. Stanford HAI's AI Index 2026 reports that "AI agents made a leap from 12% to ~66% task success on OSWorld," a benchmark of real computer tasks, "though they still fail roughly 1 in 3 attempts on structured benchmarks." On n8n Cloud, workflow history keeps 1 day on Starter and 5 days on Pro, while Git version control and multiple environments need Business at €667 a month. In code, prompts, tools and test cases share one repository, and every change runs the tests before release.

If none of these apply, stay where you are. A no-code workflow you can read on one screen is cheaper to own than an agent nobody on your side can change.

Is n8n production-ready for AI agents?

Yes, as long as the path is fixed and the judgment steps are few. Two of our production systems run on it.

The content factory for a media publisher under NDA turns new videos into finished articles with 0 written by hand. The whole line is n8n: a schedule checks the tracked channels, Apify pulls the full transcript and stats, and a clean record lands in a source table. Then two Gemini agents, built as n8n AI agents, work in sequence. The first reads the transcript and returns the angle, title and key points as structured data; the second writes the article in the house tone. A cover is generated, and nothing reaches the CMS without a review step. It is 40+ nodes, and 100% of articles are built on full transcripts.

AI Content FactoryClient under NDA

Case study · Under NDARaw video in, published articles out0hand-written articles

Judgment lives in two steps, so only two steps are agents. The rest is plain workflow, where n8n is strongest.

The Instagram agent for a marketing client under NDA pushes n8n further. One n8n agent on Google Gemini runs 100+ accounts from a Telegram chat. It has memory that keeps follow-ups in context and seven tools: a Think tool for planning, four Apify tools for Instagram data, a reel breakdown and a second Gemini agent for web search. n8n's Error Trigger catches every failed run, and a person gets a Telegram alert with the step, the account and the reason: 0 silent failures. Sensitive DMs go to a person by design. Until memory has to outlive a thread or handoffs need a queue, the platform is the cheaper owner.

When does a custom build pay off?

When AI is part of a product your team or your customers use every day, or when the "agent" is really a model trained on your data. In both cases a workflow tool would be the wrong owner of the core logic.

For Encore, a Dubai PR and events agency, we moved the team off Asana into a CRM of its own and built AI call notes next to it. A call comes in from the browser, a file upload, Google Meet or Zoom. Deepgram Nova-3 transcribes it with speakers and timestamps, Claude writes the summary and decisions, and tasks with an owner, due date and priority go to the project board in one click while the assignee gets a Telegram message.

An AI assistant inside the CRM opens and closes tasks. From the first CRM commit to Asana data in production took 9 days, and within 3 weeks 70 client calls were transcribed and searchable. Here the AI writes into the product's own data model, so it lives in the product's code.

Case study · EncoreWebsite, CRM and AI call notes for a Dubai agency3products shipped by one team

Recommendation systems sit even further from no-code. For ZOOP, a creator-first social app, we built the recommendation system, not the app: events stream through Kinesis and Lambda into a DynamoDB feature store and an S3 event lake, a two-tower model and ANN search pick candidates, a ranker orders them, and SageMaker serves the scores and retrains on a schedule. 10,000+ monthly users get a personal feed, and ZOOP owns its algorithm instead of renting a third-party feed API.

For a top-15 global grocer under NDA, we built three ML systems on 100,000+ SKUs from scratch: personal picks, shopper intent and recipe to cart. No workflow platform trains a two-tower model or fits a recipe's ingredients to pack sizes.

How does a hybrid setup split the work?

Keep orchestration in n8n and give agents only the steps that need judgment. n8n handles schedules, triggers, retries and integrations; agents classify, choose or draft. When one agent outgrows the platform, it moves into its own service and n8n calls it over HTTP.

Moving one agent out takes four steps:

1. Fix the contract. Define the agent's input and output as structured data, the way the content factory's angle agent returns angle, title and key points.
2. Collect the tests. Gather 30–50 real cases with the expected result before you rewrite anything. They become the eval set, a fixed list of cases every change must pass.
3. Rebuild behind the same contract. The new service gets its own logs, cost tracking and the memory or tools that pushed it out.
4. Swap and compare. Replace the n8n agent node with an HTTP request to the service and run both on the same cases until the results match.

Stay on no-code, go hybrid or build custom: a decision checklist

Answer these per workflow, not per company. Most companies end up with all three.

Stay on Zapier, Make or n8n if:

• the path is fixed and AI is one or two steps in it;
• your monthly units fit a plan tier with room to grow;
• the data may live in the vendor's cloud, or you self-host n8n;
• a person reviews the output before it reaches a customer.

Go hybrid if:

• one step needs memory, many tools or a real handoff, and the rest is plain integration;
• your team already maintains the integrations in n8n.

Build custom if:

• the agent is part of your product and your customers use it every day;
• you need to set or audit its memory, permissions and logs yourself;
• data must stay in your infrastructure end to end;
• platform fees at your volume exceed hosting plus upkeep;
• the core is a model trained on your data, as with recommendations.

Before you talk to a studio, write the workflow, monthly volumes, systems and data rules into our one-page software brief template. Then put our 12 questions for an AI automation agency to each studio and compare the answers. If the agent will talk to customers on WhatsApp, our guide to WhatsApp AI agents for UAE businesses covers Meta's rules and message fees on top of this choice. Our own AI automation services follow the same rule: n8n where it fits, code where it has to be.

Questions we get asked

Is n8n free?

The self-hosted Community Edition is free under the Sustainable Use License for internal use. n8n Cloud is paid: Starter costs €20 a month for 2,500 executions and Pro €50 for 10,000, billed annually as of October 2026. Business, at €667 a month for 40,000 executions, is self-hosted only and adds SSO, multiple environments and Git version control.

Which is cheaper at scale, n8n or Make?

It depends on the shape of the workflow. Make bills credits per module action, so cost grows with the number of steps; n8n bills one execution per run, whatever the step count. Short workflows that run very often favour Make; long workflows with many steps favour n8n. Count units per run and runs per month, then compare the tiers.

Can my clients edit workflows we build for them on our n8n?

Not under the free Community license. n8n's license FAQ allows building automations for clients on your instance only as long as clients cannot create or edit them. If clients need to build or configure workflows themselves, through a UI, the API, MCP or an AI agent acting for them, you need a commercial agreement with n8n.

What should we prepare before moving off Zapier or Make?

List every Zap or scenario with its trigger, steps, monthly runs and units used, plus the accounts and API keys it depends on. Mark which steps call a model and which need a person. Collect 30–50 real inputs with the expected result: they become the test set that proves the new build behaves the same before you switch it on.

Do we have to move everything at once?

No. Move one workflow or one agent step at a time. Keep the old version running, send the same inputs to both and compare the outputs before switching traffic. Most companies keep simple integrations on Zapier, Make or n8n for years and move only the parts that need memory, many tools, strict data rules or their own tests.

Sources

1. Zapier pricing · Zapier, 2026
2. Task usage rates · Zapier, 2026
3. Tasks in Zapier · Zapier Help Center, 2026
4. Make pricing · Make, 2026
5. n8n pricing · n8n, 2026
6. Sustainable Use License · n8n Docs, 2026
7. License FAQ · n8n Docs, 2026
8. The 2026 AI Index Report · Stanford HAI, 2026

Share this article

About the author

Artem Zelenskii

Head of AI

Leads AI and ML: 5+ years in production. Apple Developer Academy graduate.

Meet the leadership