How to Build Your First AI Automation (2026)

You are reading Chapter 28 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

To build your first AI automation, write it down as 3 parts - a trigger (when it runs), an action (what it does) and a result (where it lands) - then schedule one small, read-only task in your assistant and check its list of past runs the next morning. This chapter builds that task, a daily brief on 3 things you would otherwise look up by hand, and shows you how to tell an automation that had nothing to report from one that has stopped running.

It picks up where Chapter 27 left off. That chapter gave you two lists: what your assistant may do while you are not looking, and what waits for your hand. Your first automation comes off the first list.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

Most people start with the job that costs them the most time: the weekly report, the client update, the 2 hours that disappear every Friday. That is the right list to be thinking about and the wrong place to start.

The problem is not difficulty. Your first automation is where you learn what these things look like when they work and when they fail. Learning that while trusting it with something other people read is two jobs at once, and it is how people end up with a broken automation they are afraid to touch and too embarrassed to mention.

So the first one has a different job. It is not there to save you time. It is there to show you the 3 parts every automation has, prove that something runs while you are not looking, and show you what a failure looks like from the outside. It should be small enough that you would rebuild it from scratch without complaining.

Imagine it fires at 3am with completely wrong content. If the honest answer is that you delete an email, build it. If the answer involves apologising to a client, it is not your first one.

In practice, a safe first automation has 3 properties:

The valuable one is not lost. It is Chapter 29, which audits your recurring work and automates the job worth the most, using the same 3 parts on work other people see.

Every tool names these parts differently, and the naming is the only confusing part. Zapier calls the whole thing a Zap. Make calls it a scenario. ChatGPT calls it a scheduled task, Microsoft a scheduled prompt, Google a scheduled action, and Power Automate a flow. Underneath, all of them are the same 3 parts in the same order, so once you know them, the next platform change costs you 10 minutes instead of a weekend.

A trigger is a time or an event. A time trigger fires at 7am whether or not anything happened. An event trigger fires when something specific occurs, such as an email arriving from a named sender. Start with a time: a clock is predictable, so you can test it, and an event trigger that never fires looks exactly like one that is broken.

The action is a prompt like any other, and everything Chapter 5 taught you about context, task and format still applies. The difference is that nobody is there to answer a clarifying question. If a person would have needed to check something with you before starting, the automation will guess, in silence, every day.

So write the boundary into the instruction. As well as what it should do, say what it must not do: do not email anyone, do not post anything, do not add anything to my calendar, do not change any file. It feels excessive on day one. Writing the limits down is what lets you leave an automation running without checking on it every hour.

The result is where most first automations fail, because the output arrives somewhere the person never opens. Decide it deliberately: a push notification, an email to yourself, a page in the tool you check every morning. If you have to remember to go and look, you will not.

Before you open anything, write the 3 lines on paper: when it runs, what it does, where it lands. If one of them will not fill, that part is not decided yet, and the automation is not ready to build.

The automation to build is a morning brief on 3 things you currently check by hand: a competitor, a large client, a regulator, or a category you are expected to have an opinion on at the next board meeting. Name them specifically, because vague subjects produce vague briefs. It reads only, it is useful, and a total failure costs you nothing.

Use this, near enough word for word:

"Every morning, search the web for anything new in the last 24 hours about these 3 things: [name], [name] and [name]. Give me at most 5 bullets. One sentence each, with a link. If there is genuinely nothing new, reply with the words nothing new and stop. Do not email anyone, do not post anything, do not add anything to my calendar."

Find the 3 parts in it. "Every morning" is the trigger. Searching and summarising into 5 bullets is the action. The reply that reaches your phone is the result. The last sentence is the boundary, and the sentence before it is there for a reason the next section explains.

Paste the same instruction into an ordinary chat and read what comes back. You are not checking whether the news is interesting. You are checking the format: 5 bullets, one sentence each, with links, and nothing else attempted. This is the step most people skip, and it catches almost every mistake while the mistake is still cheap.

Where you set the schedule depends on the assistant you picked in Chapter 1. If you have not set one up yet, our guide to setting up ChatGPT, Claude and Copilot covers the accounts.

A free ChatGPT account can do this today, delivered in a rough morning window rather than at an exact minute, so you need to buy nothing to finish this chapter. If your assistant puts scheduling behind a plan you are not paying for, Zapier's free plan teaches the same 3 parts: unlimited automations, each limited to 2 steps and 100 runs a month. Two steps is exactly one trigger plus one action, and a once-a-day automation fits inside 100 runs. You describe what you want in a sentence and it drafts the automation for you to check.

Once the schedule is saved, read what the tool says it will do and when. If that is not the trigger you wrote on paper, fix it now.

A result that lands somewhere you never look is the same as no result. Switch on notifications for the task, so each run arrives on your phone or in your inbox without you going to find it.

After the first scheduled run, open the tool's list of past runs and confirm one is there. That list is how you know it is alive, and the next section explains why it is better evidence than your inbox.

The build takes about 4 minutes. Do it now, before you read on: the next section makes more sense once something of yours is scheduled.

If you cannot check an automation, you cannot rely on it. Two habits fix that, and both apply to every automation you build, in any tool.

Running the instruction by hand in step 2 is 30 seconds of work. If it does not produce the format you asked for while you are watching, scheduling it only repeats the mistake every morning.

An automation that says nothing when there is nothing to report looks exactly like one that stopped 3 weeks ago. Both produce an empty inbox, and you will assume the quiet one is working until the day you need it.

That is why the instruction ends by asking for the words nothing new. Now something arrives every day, and a day with no message is a day to go and look. With that one clause, a stopped automation shows up the next morning as a missing message instead of going unnoticed for weeks.

When a run surprises you, read down this list until a line is true:

From your inbox, the last two look identical, which is why the instruction has to say something even when it has nothing to say.

Each of these tools keeps a list of past and upcoming runs: ChatGPT on its Scheduled page, Claude on the Scheduled page in Cowork's sidebar, Copilot in the Automations tab. Look there rather than relying on whether you remember a notification, because notifications get swiped away half asleep.

What stops these things is rarely dramatic. ChatGPT pauses a task that goes unused for long enough, one that needs a decision from you, and one whose original chat you delete. A revoked connection, a changed password or an expired sign-in will stop any of them, and none of that announces itself.

Do not let this first one act in the world yet. Sending, posting and filing come later, with rules: stringing several apps together is Chapter 33, supervising an automation that acts is Chapter 34, and what to do when one breaks overnight is Chapter 35. For now you want one small read-only automation running, and tomorrow you want to check that it did.

Before this chapter, everything in the AI 303 track still needed you in the room. You connected your assistant to your work in Chapter 26 and decided what it may do in Chapter 27, both while watching. After it, one automation has run without you, you can describe any automation in any tool as a trigger, an action and a result, and you can tell one that had nothing to report from one that has stopped.

You take two things into Chapter 29: the automation you built, with the run list that proves it ran, and a rough list of the jobs you repeat every week, however untidy. Chapter 29 audits that list and builds the one that pays for the rest.

If you want to go further into how multi-step AI systems are put together, our guide to building AI agents covers when to build one and what every agent needs, and more guides like this one live in the AI Central Library.

Write it as 3 parts: a trigger (when it runs), an action (what it does) and a result (where it lands). Start with a read-only daily brief that only you receive, run the instruction once by hand in a normal chat, then schedule it and check the list of past runs the next day.

Yes. The free ChatGPT plan includes scheduled tasks: up to 3 active at once, each running once a day in a morning, afternoon or night window. Paid plans add hourly runs and exact times.

The trigger is when the automation runs, either a time or an event. The action is the instruction it carries out. The result is where the output lands. Every automation tool uses these 3 parts, whether it calls the whole thing a Zap, a scenario, a flow or a scheduled task.

One that reads only, writes nowhere and reaches only you, such as a morning brief on 3 subjects you would otherwise look up by hand. If it fires at 3am with completely wrong content, the worst outcome is that you delete an email.

Check the tool's list of past runs rather than your inbox. Tell the automation to reply "nothing new" when there is nothing to report, so a day with no message always means something stopped it, never that it was a quiet day.