AI automation is software that does a repeating job on its own, where one step in that job needs something read, understood or judged.
Ordinary automation already handles the rest. When this happens, do that. The AI part earns its place at the step where a rule can't work out the answer.
The difference, in one example
Say enquiries come into one inbox and somebody sorts them every morning.
Ordinary automation can look at the subject line, match it against a list of words you wrote, and forward it. That works right up until a customer writes "quick question" and means an emergency. A rule only knows the words you thought of.
AI automation reads the message. What is this about, how urgent is it, who handles this? Then the same forwarding happens, the ordinary way.
One step out of four changed. That's usually the ratio, and it's why the phrase is misleading: most of an AI automation is not AI.
What it looks like in practice
The jobs that suit it look alike. Something arrives, a person has to read it before anyone knows what happens next, and it happens often.
- Paper coming in. Work orders, delivery notes, timesheets. A photo goes in, the fields come out into whatever system you already use.
- Enquiries needing triage. Sorted by subject and urgency before anyone opens them.
- Recordings nobody writes up. A site visit or a call becomes a summary and a list of what to do next.
- Documents somebody has to search. You ask a question, it points at the paragraph and quotes it.
When a rule is the better answer
A lot of what gets called AI automation is ordinary automation. Rules are cheaper, they're faster, and they do the same thing every single time.
The test is one question: does a person currently have to read something and decide, before anyone knows what to do next?
If the answer is no, you want a rule. Moving a form submission into a spreadsheet, sending a receipt, reminding somebody on a Tuesday, updating stock when an order comes in. None of those need anything read. Adding AI to them costs more and makes the result less predictable.
If the answer is yes, that's the step worth handing over, and only that step.
What it costs to run
There are two costs and they behave differently.
- Building it. A one-time project. Custom software here starts at $5,000, and automation work on its own is often less. Larger or more custom projects are quoted separately.
- Running it. AI charges per use, not per month. Reading one document is a fraction of a cent. Reading nine hundred a week is a line on your books, and it moves with your volume.
That second one is worth modelling before you build. Ask what it costs at today's volume and at three times today's volume. If nobody can tell you, ask again before you sign anything.
How to know it's working
Two numbers. Take both before you start.
How long the job takes now, and how often it happens. That's the hours you're buying back.
How often it gets the answer wrong, once it's running. Not zero, and it shouldn't need to be. What matters is whether a person can spot a wrong answer quickly and fix it, and whether anybody is actually looking.
Anything that runs unwatched will eventually be wrong for a month before somebody notices. Build the check in at the start, not after that happens.