Automate a repetitive task with AI in an afternoon
Pick the right task, wire it up with no code, and know when to stop.
Before you start
- A task you genuinely do every week
- A free Zapier or Make account
- Admin access to whatever tools the task touches
What you will be able to do
- Judge whether a task is worth automating at all
- Build a working AI automation without code
- Recognise the automations that cost more than they save
The honest first question is not "how do I automate this" but "should I". Plenty of automations take four hours to build, break monthly, and save six minutes a week.
This guide covers picking a task that clears that bar, and then building it.
Score the task before you build anything
Frequency, duration and tolerance for being wrong. Two out of three is not enough.
A task is worth automating when it is frequent (weekly or more), takes long enough to notice (ten minutes or more), and tolerates occasional error — because an AI step will occasionally be wrong.
That last one rules out more than people expect. Anything where a wrong output goes straight to a customer, or into accounts, needs a human check, which changes the maths considerably.
- Automating something rare because it is annoying. Annoyance is not frequency.
Write the manual steps down first
You cannot automate a process you have not described.
Do the task once, writing down every step including the judgement calls. The judgement calls are where the AI step goes; everything else is plumbing.
This usually reveals that the task is two tasks, or that one step nobody documented is the one holding it together.
Build the plumbing without the AI first
Get the trigger and the destination working with a hard-coded value.
In Zapier or Make, wire the trigger to the destination with a fixed placeholder in the middle. Confirm the whole path works end to end before an AI step is anywhere near it.
Debugging a broken automation with an AI step in the middle is significantly harder, because you cannot tell whether the plumbing or the model is at fault.
Add the AI step with a tightly constrained prompt
Ask for a specific format, and say what to output when the input is unusable.
The prompt in an automation is different from a chat prompt: nobody is reading the output before it is used, so the format has to be exact and the failure case has to be explicit.
Always include an escape value. A step that outputs UNCLEAR when the input makes no sense is far better than one that invents something and passes it downstream.
Classify this support email into exactly one of:
BILLING, TECHNICAL, SALES, OTHER
Output only the single word. No explanation.
If the email fits none of these, output OTHER.
If the email is empty or unreadable, output UNCLEAR.
Email:
{{input}}- Add a filter step after the AI that stops the run on UNCLEAR. That is your safety valve.
Run it manually for a week before trusting it
Watch every run for a week. You will find something.
Leave it triggering manually and check each result for a week. The failure you find will almost always be an input shape you did not anticipate, not a model mistake.
After a clean week, switch it on and set a monthly reminder to check it still works — the tools it touches will change their APIs eventually.
Downloads & resources
Common questions
Was this guide useful?
94% of readers found this useful
Read next
Connect two apps with an AI step in the middle
The genuinely useful automations are not the clever ones. They are a trigger, one AI step that makes a small judgement, and a writ…
Add a human approval gate to an AI automation
Approval gates fail in two directions: too many and people rubber-stamp, too few and you find out from a customer. Both are avoida…
Write ad variants and test them properly
AI removes the cost of writing variants, which makes it very easy to run tests that cannot teach you anything. The discipline is i…
Choose your first AI assistant without overthinking it
Every comparison table lists twenty differences and only three of them change your day. Here is how to pick in ten minutes and get…
