Writing
What AI Automation Actually Is, and What It Is Not
Most explanations of AI automation are written by people selling a tool. Here is the distinction that actually decides whether it will work in your company.
Published · Updated · 7 min read
The short version
AI automation is the replacement of work that requires judgement. Traditional automation moves data along a rule you wrote; AI automation makes the decision the rule could not express, based on patterns in your own history.
There is a useful line running through everything labelled automation, and almost nobody draws it clearly because drawing it makes half the tools on the market look narrower than their marketing suggests.
Traditional automation moves things. AI automation decides things.
Traditional automation runs a rule you wrote. When a form is submitted, create a record. When an invoice is paid, mark the job closed. It is reliable, cheap, and it has been available for twenty years.
It has one hard limit: you have to be able to state the rule in advance. The moment the correct action depends on context that varies case by case, a rules engine needs a human to look at it and decide.
AI automation is what happens when the deciding itself can be done by a system, because the pattern behind those decisions is present in work your company has already done.
An example that makes the line obvious
A distributor writes estimates from price lists, freight tables, and a salesperson's memory of what that account paid last time. Two days per estimate is normal.
Traditional automation can route the request, create the record, and notify the salesperson. The two days do not move, because the two days are the salesperson deciding.
AI automation trains on three years of quotes actually sent, the catalogue, and the freight rules. It drafts the estimate; a person approves it. Two-day estimates go out in four hours. The difference is not speed of data movement. It is that the decision moved.
The three things that make it work
- History. If your company has done the work a few thousand times, the pattern exists whether or not anyone wrote it down.
- Boundaries. Clear rules for what the system must never decide alone, and what happens when it is unsure.
- An owner. One named person inside the company who runs it, corrects it, and can turn it off.
Miss the first and you are prompting a model to guess. Miss the second and one bad output costs more than the project saved. Miss the third and it quietly dies within a year, which is the most common outcome by a wide margin.
What AI automation is not
It is not a chatbot on your website. That is one possible output, and usually a low-value one.
It is not your team using ChatGPT more. That makes individuals faster, which is real but does not remove an operation. Eight people who are twenty percent faster is not a headcount saved, it is eight people with more time and the same cost structure.
And it is not a strategy document. A roadmap describing where AI could be applied is a description of a problem you could already describe.
How to tell if you have work worth replacing
One test, and it is short. Try to explain the work to a new hire in a single sentence. If you find yourself saying it depends, that is judgement work, and judgement work is what this is for.
- It happens many times a week, by several people
- Explaining it requires the phrase it depends
- Your company has done it enough times to leave a record
- Getting it wrong costs money, so it currently waits for a person
Work that scores on all four is where the money is. Work that scores on none of them should be automated with a $20 tool and no consultant.
What it costs, honestly
The tool is never the expensive part. The expensive parts are deciding what to build, preparing the data, testing against real cases, and getting a team to trust it.
Companies that skip the first of those build something impressive attached to a cost line that was not the problem. That is the most common failure in this category, and it has nothing to do with the technology.
Questions
Is AI automation the same as RPA?
No. RPA repeats a sequence of interface actions exactly as recorded. AI automation makes a decision based on patterns in your data. RPA breaks when the screen changes; AI automation handles cases nobody wrote a rule for.
Do we need a lot of data?
You need history rather than volume for its own sake. A few thousand real examples of the work, such as past quotes or support replies, is usually enough.
What size company does this make sense for?
Roughly $3M to $10M in revenue. Below that it is usually a people or process problem. Above it, somebody inside should already own this.
Read next
If you are comparing options
Find out what your most expensive operation is
The assessment covers the whole company: how you get customers, how you deliver, and how the place runs. You get it in writing, and you can build from it with or without us.