AI automation means making a repeated task run on its own, with an AI model in the step that needs reading or deciding: sorting an email, pulling the data out of an invoice, answering a question. Rules cover the cases you programmed; a model handles text that is different every time, but it can be wrong.
A question from the app. Pick an answer.
Jev is an AI model that does not write: it picks an option or gives a score, with probabilities.
Jev splits a ticket: Sales 34%, Support 33%, Billing 33%. What do you do?
It wins by one point. An almost even split means confidence near 0: a three sided coin.
Almost a tie: that is Jev saying "no idea". Send the case to a person or ask for more details.
Asking again does not make the ticket clearer. If the text is not enough to decide, what is missing is information, not another try.
Automation is not new: a script that copies sales into a spreadsheet every night is already an automation. It follows the conditions you programmed, including the variations you planned for. AI comes in when the step is reading something people wrote: a model sorts by meaning, without a rule for every phrase. "I was charged twice" and "there is a repeated charge on my card" are the same complaint: an AI classifier sends both to Billing even though they share no words.
The usual move is to mix: rules for what is always the same and a model only in the step that needs it. You build it with code (Python and an AI API) or with workflow tools like n8n, Make or Zapier, which connect apps and add an AI step.
If you decide the steps in advance (the email arrives, the model sorts it, a ticket is created), it is a workflow. If the model picks which tool to use and in what order until it is done, it is an AI agent. A workflow is more predictable and easier to check; an agent helps when the steps change from case to case. The difference, with examples, is in workflow or agent.
A model is wrong now and then, and not always in the cases you expect. That is why you decide what it does alone and what goes to a person. Moving a ticket to the wrong team is quick to fix; refunding money or deleting data is not. If the system returns a confidence with the answer, it can act above a threshold and ask below it. That confidence is not a hit rate: the threshold is tested on real cases and adjusted by looking at the mistakes. More in calibrated probability.
In job ads, plain "automation" often has no AI in it: automating tests (QA), servers or networks with scripts, or office tasks with RPA. (RPA means software robots that repeat clicks and typing on screens.) Read the description: look at what task it automates and whether it uses an AI model. Both are good places to start.