That experience, more than any theoretical checklist, shows why this care matters.
Why it pays to isolate the entry point
I learned that the first step of any automation is where data enters the system, and that is where the silliest problems tend to show up. If the form is not properly built, everything downstream receives wrong or incomplete information. I ran into this directly.
In class, when trying to connect a WordPress form to Make, the first obstacle was not even technical in a complex sense: it was simply not being able to build a working form. The Elementor plugin I had installed required a Pro license just to create a form, and the Contact Form 7 plugin I tried next needed a paid add-on to fire a webhook directly. I had to switch approach and use Elementor's native integration with Make to move forward without paying for something I did not actually need.
Once the form finally connected to Make, the information arrived bundled into a single spreadsheet cell, because the initial mapping did not separate name, email, and message. Only after asking ChatGPT for help writing a splitting formula was I able to spread the information into readable columns.
If you are building something similar, it is worth checking this before showing the sheet to your team: feed it a test entry and confirm each field lands in the right column without mixing. And if the platform you are using has several versions of a form plugin, it is worth confirming upfront whether the integration you need is actually included in the free version, rather than discovering that only after you have already built the rest of the flow.
Why credentials and permissions deserve attention upfront
I learned that connecting tools like Gmail or Google Sheets to automation platforms usually requires enabling specific APIs in the Google Cloud Console. Without that, the automation simply will not move forward, even if the rest of the flow is well designed. I hit this wall myself.
This happened very concretely in class: when trying to connect Gmail and Google Drive to n8n, I got blocked repeatedly because the Gmail and Drive APIs were not enabled in the project, and because I had deleted some older credentials during another training. Fixing this required following Google's own documentation step by step, creating a new consent screen, enabling the needed libraries, and generating new client credentials.
I would recommend doing this permissions setup before any test with real data, because an error like this, visible in class as an "access blocked" or 403 error, can surface only after your team is already relying on the system, and at that point it is no longer an exercise, it is a live problem.
How to plan an app with Claude Code before you start coding
How to use AI in business beyond chatbot prompts
To design execution and handle exceptions, explore the business process automation guide.
Why row matching needs a clear key
I learned that when an automation updates an existing sheet, it needs to know exactly which row to change. If the matching rule is poorly defined, the system may never find the right row. This became clear when I tested the update step.
That is exactly what happened in class: when trying to get n8n to update the column with the AI-generated message, the "update row" step kept failing because the match was being made on the very column I wanted to write to, instead of using a stable identifier. Only after switching the matching key to the email field, with help from ChatGPT to understand the error, did the update start landing on the correct row.
Before handing the sheet to your team, test this with two or three similar entries and confirm each update lands exactly where it should. Always pick a field that is unique by definition as your key, such as an email address, and avoid fields that can repeat, like a name.
Why it is worth watching how AI interprets real cases
When an AI agent generates text to send to customers, the first result is rarely ready to go. In class, the first version of the generated message included context information that should not have appeared in the email body, with no formatting suited for reading.
Fixing this required rewriting the instructions given to the agent, explicitly asking for clean HTML with a defined structure: greeting, mirroring the request, how the company can help, one clear question, and a closing line along the lines of "we'll be in touch shortly." Only after that second attempt did the message come out in an acceptable format to send.
Before letting an agent like this respond to real customers, test it with varied messages and check whether the output stays consistently on-tone. It is also worth testing short or vague messages, just to see how the agent reacts with little information, because those edge cases are usually where the most visible errors show up.
Why an end-to-end test avoids surprises after turning on the automatic trigger
Only once every isolated step is working does it make sense to switch the whole flow to automatic. In class, once the form, the mapping, the credentials, and the message formatting were all sorted, I tested the full flow by sending fictional requests and confirming, in real time, that the message arrived in the inbox.
Even so, one detail was left unrefined: the automatic trigger on each new row was not entirely reliable, and one message arrived twice to whoever had submitted the form. It is a good example that even after a successful test, it is worth running the flow more than once before considering it ready to work unsupervised.
If you are validating a similar flow, I would leave the automatic trigger switched off during testing and run the flow manually, step by step, until confirming each step behaves predictably more than once. Only after that would I turn on the automatic trigger, ideally first with internal contacts before exposing the system to real leads.
Sextas Ímpares #125: How to automate tasks with AI assistance
This reflection comes from the class I gave live during Sextas Ímpares #125. In that session I built, live and with real stumbles, an automation connecting a WordPress form to Google Sheets, then to an AI agent in n8n, and finally to sending a personalized email through Gmail. It was all recorded, mistakes included, because that is what shows the real process of testing something before handing it to a team.