01Start with the process, not the tool
The usual question is «n8n or Make?». It is a fair question, but it comes too early. First you need to name the process: what triggers it, what data it needs, which systems it touches, what result is expected and who is responsible if something goes wrong. Once that is written down, the tool almost picks itself. Without it, any tool ends up automating the mess.
In practice, the first process should be small, repetitive and easy to check: sorting the emails in one inbox, summarising a conversation, creating a CRM record from a form. How to choose it is covered in the guide to automating orders, bookings and CRM with AI. If the process needs to decide steps, rather than follow a rule, read the difference between an AI agent and a chatbot first.
02What can be connected in practice
AI is useful when it works with the systems where the data already lives. In a small business, that usually means a mix of these:
- CRM: create or update contacts, log a conversation, prepare the next step on a deal.
- Email: sort incoming messages, extract order or request details, draft a reply.
- Calendar: suggest slots, create a booking, prepare a reminder.
- Online store: check an order status or stock, flag an issue.
- Invoicing: leave a draft ready for a person to review and issue.
- WhatsApp or the website chat: collect an enquiry and route it to the right system, telling people they are talking to an AI.
The point is not to connect everything at once. Each connection adds permissions, data in transit and something to maintain. That is why the first phase is written with concrete names: this inbox, this CRM, this calendar.
03When n8n or Make make sense
n8n and Make are automation platforms with connectors for many tools and AI steps that can sit inside a flow. They make sense when the process reads well as a sequence of steps, volume is reasonable, exceptions are few and known, and someone on the team can open the flow and understand it. They are also useful to validate an idea quickly, before investing in custom development.
Their strength is visibility: the flow can be seen, reviewed and paused. Their limit shows when the logic grows into an application hidden across dozens of nodes, when very fine permissions are required, or when volume and exceptions call for control the platform does not offer comfortably.
04When you need something custom
A custom integration —a dedicated service that talks to each tool’s API— makes sense when the process is critical, the logic is complex, the data is sensitive or the flow has to be part of a product. It takes more work up front and clear maintenance, in exchange for full control over permissions, logs and error handling.
| Criterion | n8n or Make | Custom integration |
|---|---|---|
| Volume | Low or medium, with predictable peaks | High or variable |
| Permissions | Whatever each connector offers | Defined in detail, per action |
| Exceptions | Few and known | Many, or with their own rules |
| Maintenance | Visible to the team, no coding | Needs development and documentation |
| Speed to validate | High | Lower at first |
The table does not name a winner. It is common to start with a flow in n8n or Make to validate the process and move to custom code only where it is needed. The decision is made in the diagnosis, with the process on the table.
05Permissions, data and what is never left unreviewed
Each connection should have the minimum permissions for its task: a flow that only reads orders does not need to change them. Credentials live in the platform’s or the server’s secret store, never in the text of a step. And it is decided up front which data reaches the AI model, which must stay in the source system, and where each run is logged.
In the work I propose, some actions never leave a flow without a person approving them: a customer message that is not an agreed standard notice, a change to an order, a charge, a discount or a firm CRM entry. The flow prepares the draft and leaves it in an approval queue. I develop that pattern in the guide to internal AI agents with human approval.
06How it is tested before going live
A flow is not tried out on real customers. It is first tested with sample data and edge cases: an email with missing details, a duplicate order, an unexpected model response. Then it runs in parallel, with no effects, comparing what it would do with what a person did. Only then is it switched on, with error alerts and a simple way to pause it.
After launch, I review which cases the flow resolves and which go to a person. That shows whether to widen the scope, fix a rule or leave it as it is. The full method of diagnosis, project and support is described in the AI integration service.