Automation follows a path you decided in advance. When this happens, do that. It's a rule, and it does the same thing every single time.
AI handles the step where a rule can't work out the answer, because something has to be read, understood or judged first.
The one question
Does a person have to read something and decide, before anyone knows what happens next?
If no, you want a rule. Send the receipt, remind them the day before, move the form into the spreadsheet, reorder when stock drops below ten. Nothing in any of those needs reading. A rule is cheaper, it's faster, and it behaves identically every time.
If yes, that's the one step worth handing to AI. What is this email about. What are the fields on this photograph. Is this claim complete. None of those can be written as a list of conditions, because the answer depends on what the document says.
Why it matters what you pick
Getting this wrong is expensive in both directions and the two failures look nothing alike.
Using AI where a rule would do costs more to build, costs more per run, and makes a job that was completely predictable slightly unpredictable. Ask it the same question twice and you can get two phrasings. For sending a receipt, that's a downgrade you paid for.
Using rules where AI is needed produces the automation everybody has met: it works on the examples somebody thought of and falls over on the rest. The list of exceptions grows every month until somebody quietly goes back to doing it by hand.
Most useful builds are mostly rules
Take sorting an inbox. The AI reads each message and says what it's about and how urgent it is. That's one step. Everything around it is ordinary code: fetch the messages, apply the routing, write it to the system, log what happened, retry if the connection dropped, tell somebody if it failed twice.
One step in seven or eight is common. It's why the phrase AI automation is a bit misleading, and why a quote that's all about the AI is usually missing most of the work.
What each costs to run
They're billed completely differently, and that catches people out at scale.
- Rules are billed per step, or not at all. Zapier's entry paid plan is $19.99 US a month billed annually and includes 750 tasks. Something you own outright is free to run, beyond the hosting.
- AI is billed per use. A fraction of a cent per document. Trivial at ten a day, a real line at two thousand a day, and it grows exactly as fast as you do.
Model both at today's volume and at three times today's volume before you commit. A build that's affordable now and painful at scale is worse than one you didn't do.
How to decide in five minutes
Take the job you have in mind and write down every step as it happens today.
Go through them one at a time and mark each step with whether a person has to look at something and form a view. Most steps won't need it.
The marked steps are your AI. Everything else is rules. If nothing gets marked, you don't need AI for this, and that's good news.