We run an operation with a lot of artificial intelligence in it, and we are asked fairly often what that actually means day to day. The honest answer is less impressive and more useful than the pitch decks suggest: we automated the parts that are repeatable, and deliberately did not automate the parts where being wrong is expensive.
What we automated
The scaffolding. Scheduling and routing work between departments. Indexing footage so a clip from four months ago is findable in seconds rather than an afternoon. Sweeping categories for what is new. Drafting, formatting and publishing the same shape of thing for the hundredth time. Reconciling numbers that have to match.
What these have in common is that they are repeatable, checkable, and boring. When they go wrong, the failure is visible immediately and costs an hour. That is exactly the risk profile you want before handing something to a machine.
What we did not
The verdict. Whether a product is actually good is not a thing we delegate, because the entire value of a review is that a person who tested it is willing to put their name on the conclusion.
The terms with a brand. What a supplied unit does and does not buy is a commitment, and commitments need someone accountable for them.
And what we cover next. Research can tell you what people are searching for. It cannot tell you whether the answer they will get is worth the week it takes to produce.
The rule underneath it
The thing you automate is the thing you stop noticing. That is the whole point when the task is indexing files. It is a serious problem when the task is deciding whether something is true.
So the test we apply before automating anything is not can this be automated. Most things can. It is if this quietly went wrong for a month, how would we find out? If the answer is a dashboard, automate it. If the answer is a reader emailing to say they bought the wrong thing because of us, a person keeps doing it.
Why the boring version is the useful one
Most published advice about AI in business is written by people selling AI to businesses, which makes it directionally optimistic in a way that is hard to act on. The version that has actually worked for us is narrower: find the tasks you do the same way every time, hand those over, measure whether the output still holds up, and spend the time you got back on the judgment calls that were always the hard part.
We talk about this at more length on AI In The World, our podcast and written series on how this technology is landing in ordinary businesses and ordinary lives.