Automating routine work

Connect apps and add AI steps to workflows you repeat every week. About 1 h, advanced level.

You'll be able to

• Map a weekly task as a trigger, a series of steps and an output before you build anything.
• Build a workflow in Zapier or n8n with one AI step and a manual approval before anything is sent.
• Test the workflow with real data, log every run and plan for what happens when a step fails.

Triggers, steps and where AI fits

Every automation has the same anatomy. A trigger starts it: a new form response, an incoming email, a file added to a folder, or a time of day. Then a series of steps runs: look something up, copy it to a spreadsheet, send a message.

Most steps are ordinary software. They do exactly the same thing every time. An AI step is different. It can read messy text and summarize, classify, extract or draft, which ordinary steps can't. But it's probabilistic: the same input can give slightly different outputs, and sometimes the output is wrong.

So use AI only for the step that needs it, usually the one where a person currently reads free text and makes a judgment. Keep everything else, like dates, lookups, routing on exact values and arithmetic, in plain steps. A workflow with one well-placed AI step is easier to trust and debug than one where every step is a model.

Map it on paper first

Start with a chore you already do by hand every week. If you can't describe how you do it, you can't automate it.

On paper, write down:

• The trigger: what starts the task?
• Each step, in order, including the ones you do without thinking.
• The decisions: where do you read something and choose what happens next? That's the likely AI step.
• Where data goes: every app the data passes through.
• What “done” looks like, and who sees the result.

Then ask whether it's worth it. Count the minutes it takes each week against the time to build, test and maintain it.

Good first projects are low stakes and internal: summarizing club form responses into a sheet, or sorting your own reading list. Poor first projects send messages to many people, touch grades or money, or handle other people's sensitive information.

Make the AI step predictable

The next step in the workflow can't read a friendly paragraph. It needs an output in a fixed shape. Write the AI step's prompt like a specification:

• Fixed choices: “Classify the message as one of: dietary, accessibility, schedule, other.”
• An out: “If none fit, or the message is empty, answer other.”
• Structured output: ask for set fields, like a label and a one-line summary, so later steps can pick out each part. Many tools can request structured (JSON) output directly.
• Examples: two or three sample inputs with the correct label.

Then add a check after the AI step. If the output isn't one of the allowed labels, or a required field is blank, route it to a person instead of carrying on.

This is the same problem Microsoft's course describes for developers: free-form model answers vary from run to run, and consistent structure is what lets other systems use them.

Keep a human in the loop

An automation acts without anyone watching, so a mistake can repeat on every run before you notice. Put a person in front of every consequential step: anything sent to other people, anything touching grades, money or records, and anything that deletes data.

The simplest pattern is to have the workflow prepare, not act. Save a reply as an email draft, or post it to a channel where you approve it, rather than sending it straight away. You keep most of the time saved and all of the judgment.

Be open with people about what's automated. If a message was drafted by AI and approved by you, it's reasonable to say so, especially if people might reply expecting a person read their message closely.

Once a workflow has run correctly for weeks, you might remove approval from its lowest-risk steps. Make that a deliberate decision based on your logs, not a shortcut you take because approving got tedious.

Test, log and plan for failure

Test with real data before you rely on it. Rerun ten past examples through the workflow, including awkward ones: a blank field, a very long message, a message in another language, one that fits no category. Compare the results with what you'd have done by hand.

Log every run, for example as a row in a spreadsheet: time, input, AI output, action taken, and whether it was approved. Logs show you patterns, like one category the model keeps getting wrong.

Plan for failure, because it will happen. A connected app changes its form, a service is down, you hit a usage limit, or an account's access expires. The worst failure is a silent one. Both Zapier and n8n can alert you when a run errors; turn that on.

Finally, mind the data. Every connected app, and the AI provider, is another place it flows. Keep student records, health information and other people's private data out of workflows on personal accounts.

A club sign-up workflow with one AI step

A student runs sign-ups for a club's workshops. Each week they read form responses, answer questions and update a spreadsheet by hand.

The map: trigger (new form response), read the question field, decide what kind of question it is, draft a reply, log it, send.

The build in Zapier:

1. Trigger: new response in the sign-up form.
2. AI step: classify the question as dietary, accessibility, schedule or other, and draft a two-sentence reply. Blank or unclear questions are labeled other.
3. Filter: if the label isn't one of the four, stop and flag it.
4. Log: add a row to a sheet with the time, question, label and draft.
5. Approval: save the reply as an email draft. The student reviews and sends it.

Testing: they replay twelve old responses. The labels are right on most, but two blank responses got a cheerful reply to a question nobody asked. They add a rule that skips the AI step when the question field is empty, and switch on error alerts. Sign-up admin drops from an hour a week to about fifteen minutes of reviewing drafts.

Try it

A paper map of one weekly chore, a working Zapier or n8n workflow with one AI step and a manual approval, and a log of at least ten test runs with notes on what failed and what you changed.