n8n, Make, Zapier, Copilot Studio or a custom agent: which to choose?

In short

• Make, Zapier, n8n and Copilot Studio publish their prices. To start, on annual payment: $9 per month at Make (price announced in April 2026), $19.99 at Zapier and €20 at n8n (checked on 13 October 2026). They suit you if someone at your company knows how to build and monitor them.
• A coded agent is justified when it connects several pieces of software, prepares actions that commit the company and must keep a log, approval levels and supervision.
• At Kendu Intelligence, we have chosen to build agents in code. And if a service provider offers you n8n, its licence limits what they can do: we detail it below.

You want software to do a repetitive task for you. You are told about Make, Zapier, n8n, Copilot Studio, or a “custom AI agent”. We sell the last one: so we also tell you when you do not need it.

An automation tool chains steps between your software: “when this happens, do that”. An AI agent adds an AI model that reads a text, understands a request and prepares a reply or an action.

Self-service tools and their public prices (October 2026)

“Self-service” means: you create an account and build it yourself. Prices checked on 13 October 2026 on the official pages, except that of Make Core: its pricing page does not show it as text, it comes from its blog (April 2026). Only Microsoft's French page states that its prices exclude tax; the others do not say whether VAT is included.

Each tool counts in its own way: one execution at n8n (a full run of a scenario, whatever the number of steps), one task per successful action at Zapier, generally one credit per action at Make. Compare on your real volume, not on the entry price.

What they do well: connect two tools, trigger a simple action

• They start fast. Free trial without a bank card for n8n's Starter and Pro plans, 14-day trial without a card for Zapier's Professional plan, free plan at Make.
• They have ready-made connectors. A connector is a bridge that lets two pieces of software exchange data.
• They show the scenario. A visual editor displays each step.

For example, an SME with 20 employees wants every enquiry received on its website to create a record in its CRM (the software that tracks its clients) and alert the salesperson. Two pieces of software, one rule, no delicate decision: a self-service tool is enough.

If you already work in Microsoft 365, look at Copilot Studio: according to Microsoft, users who have the Microsoft 365 Copilot licence can create internal agents at no extra cost.

Their limits for an SME: who builds, who monitors, who answers when it breaks down

The tool is self-service. The work around it is not.

• Who builds? You need to know each piece of software, its access rights, its special cases. In the European Union, among businesses with 10 people or more that had considered AI, lack of expertise is the main reason for not using it: 70.89% in 2025 (Eurostat).
• Who monitors? A password changes, a scenario fails on a Friday evening. An alert goes out. Someone still has to read it and know what to do.
• Who answers when it breaks down? The vendor answers for its tool, not for your process. At n8n, for example, the Business plan points to the forum; dedicated support is reserved for the Enterprise plan. If the person who built everything leaves, the scenarios may be left without instructions.
• Who approves? When an AI model writes a reminder, a mistake can be costly. You need to decide who reviews and keep a record, whatever the tool.

The n8n licence: what a service provider can do, and what it cannot

n8n can be installed free of charge on your servers (Community edition). But its licence, the “Sustainable Use License”, limits use to your “internal business purposes”, or to non-commercial or personal use. The official n8n FAQ explains what this means for a service provider.

Allowed, according to this FAQ:

• charging for the creation, installation and maintenance of scenarios, training and consulting;
• a managed service on the provider's own n8n installation, if clients do not access it directly;
• installing and maintaining n8n on the client's server, “provided that you are not also hosting for them”: if the provider does not also host it for them;
• triggering already-built scenarios through an API (an entry point for software) or a webhook (an automatic signal).

Forbidden:

• hosting n8n as a service where clients build their own scenarios;
• selling n8n as white label, that is, by removing the n8n brand;
• using features reserved for the Enterprise licence without that licence.

A case to have confirmed. To run n8n for several clients, the FAQ refers to license@n8n.io and announces a partner programme in preparation. If a service provider offers you this set-up, ask them for written confirmation from n8n.

When a coded agent is justified: log, human approval, frozen versions, tests

A “coded” agent is a programme written, tested and reviewed by developers, rather than drawn in a visual editor. It is justified when these needs add up:

• Several pieces of software to connect. Email, invoicing, bank, ERP (business management software): each link has its own rules.
• A log of every action. What the agent read, what it proposed, who approved, when, with which version. Months later, you can answer “who sent this reminder?”.
• Human approval by levels. Level 0: the agent proposes, your team does the work, we compare. Level 1: it prepares, a human approves each action. Level 2: it acts alone on safe cases, written in advance. Level 3: it acts alone, under monitoring. Some actions never go beyond level 1: a payment, sending a third party a document that commits the company, deleting data. And in the stop cases written in advance, the agent stops and hands over rather than guessing.
• Frozen versions and tests. n8n published versions 2.42.5, 2.42.6 and 2.42.7 on 8, 9 and 12 October 2026 (GitHub). Whatever the tool, the rule is the same: freeze a version that works and test each update before it reaches you. With us, each version of an agent will go through automatic tests on anonymised data, a review, then a short approval trial.

For example, an SME with 60 employees wants an agent to read the supplier invoices received by email, match them to orders in its ERP and prepare the accounting entry. Three pieces of software, documents to read, money at stake: this is the typical case for a coded agent.

Our choice, and why

At Kendu Intelligence, this is what we have decided.

• The agents will be written in code, with Claude (Anthropic's AI model) and Anthropic's agent development kit, then tested and versioned like software.
• Supervision and support will go through a tool in preparation. An agent that meets a repeated error is planned to report it itself: a ticket will open, we will be alerted, and you will see it in your client area.
• Deployment is your choice: your cloud in a European region, your servers, or a dedicated server with us. The AI model, for its part, may be called at a provider: see data and security.

Why? Because the human must keep control: the agent prepares, your team approves. This requires a log, levels and tests that we will write and check ourselves.

And if a self-service tool is enough for you, the audit will say so. It is on quote.

Going further

• Our standard agents: reminders, support, quotes, accounting documents.
• Our method: audit, agent rulebook approved by you, a trial where the agent proposes without acting, autonomy levels.
• Supervision: what is planned (dashboard, alerts, follow-up reports).
• Data and security: where your data goes, including what is sent to the AI model.
• Request an audit.