Automation 101 Guide

What automation actually is, what to build first, and the mistakes that make people give up in week two.

The mental model

Every automation is three things:

TRIGGER  ->  STEPS  ->  OUTPUT
(when)      (what)     (where it lands)

That is all. Everything else — routers, filters, loops, AI steps — is detail inside "steps".

Choosing the first thing to automate

Not the most annoying thing. The thing with this profile:

• You do it at least weekly
• The steps are the same every time
• You can describe it to another person in under two minutes
• If it runs wrong, nothing bad happens
• It takes at least five minutes each time

Do not start with anything customer-facing, anything involving money, or anything with lots of exceptions. Those come after you understand the tool.

Good first automations: saving form responses to a sheet, sending yourself a daily digest, filing attachments, logging something you currently track by hand.

Map before you build

Ten minutes here saves hours. Answer three questions:

1. Where does the data start? The trigger. Be specific: not "an email", but "an email to support@ with the label 'refund'".
2. Where does it end? The output. A row? A draft? A notification?
3. What changes in between? Formatting, a lookup, an AI judgement, a filter.

Write it as a line before you open the tool:

New form response -> AI classifies urgency -> if urgent, Slack me;
otherwise add row to sheet

If you cannot write that line, you are not ready to build.

The filter is not optional

The single most common and most expensive mistake: no filter on the trigger.

Trigger: new email                     <- runs on EVERY email
Trigger: new email WHERE label = support AND sender not internal

Without the filter, your automation runs hundreds of times a day, costs money on every AI step, and buries you in notifications. Add it before the first test run, not after the first bill.

Variables, not hardcoding

Bad:  Send email to "[email protected]"
Good: Send email to {{Email from the trigger}}

Every value that changes between runs must come from the data, not be typed in. This is what makes one automation serve a thousand records instead of one.

Test properly

• Run it once with real data and watch every step's output
• Run it with a blank or missing field — what happens?
• Run it with something that should be filtered out — does the filter hold?
• Run it ten times — is the result the same each time?

That last one catches the temperature problem. If an AI classification step gives different answers on identical input, the temperature is too high. Set judgement steps to 0–0.2.

Error handling from day one

Automations fail silently by default. That is the worst property software can have.

Every automation needs:

• An error path. If a step fails, something tells a human.
• A known cost per run, and a monthly estimate at expected volume.
• A log. At minimum, what went in and what came out of the AI step.

If any step errors -> send me a message with: which step, what the input
was, and the error text.

Five minutes to set up, and it is the difference between noticing a failure today and noticing it in March.

The build order

1. Build the trigger. Test that it fires.
2. Add the filter. Test that it excludes correctly.
3. Add one step. Test.
4. Add the next step. Test.
5. Add the error path.
6. Run it live, watching, for a week.
7. Then stop watching.

Building all six steps and then testing means debugging six things at once.

The mistakes that make people quit

Starting too big. A 20-step workflow on day one. Start with two steps.

No filter. Covered above.

Automating a broken process. If the process is wrong, automating it produces wrong results faster. Fix the process first.

No documentation. It breaks in three months and you have no idea what it did. One paragraph per automation: what it does, what triggers it, what breaks if it stops.

Automating something you do monthly. The build cost exceeds the saving. Automate weekly things.

Not knowing how to turn it off. Know this before you turn it on.

What not to automate

• Anything where being wrong is expensive and the error is invisible
• Anything requiring judgement about a person
• Anything you do rarely
• Anything with so many exceptions that the exceptions are the work
• Sending. Draft-first, always — a human presses send

Your first automation

What it does
Trigger, precisely
Filter
Steps
Output
Runs per week, expected
Cost per run
What happens if it fails
How I turn it off
Documented where