AI's easy promise is to replace. The one it keeps is to inform.
There is a distinction almost forty years old that remains the most useful for deciding where AI is worth it. It separates two things almost always treated as one: to automate and to informate.
Shoshana Zuboff put it that way in 1988, in a book about work and the machine — In the Age of the Smart Machine. Technology, she argued, can be used to automate or to informate: to replace the person doing the work, or to give that person a view they did not have before.
Almost four decades later, it is still the question that separates AI projects that work from the ones that disappoint.
What happens when you automate what nobody understands
The system starts producing results nobody knows how to check.
Not because people are incapable — because the knowledge of how that used to be decided lived in the head of whoever decided it, and the decision has now left that head. When a result comes out strange, there is nobody to ask: the process became a box, and the box does not explain.
In that arrangement, the error does not show up when it happens. It shows up once it has become a consequence — a customer complaining, a wrong number in a report that already circulated, a decision taken on what the box said.
And there is a cruel detail: the better the automation seems to work, the fewer people check it. Trust grows with time in use, not with verification.
What happens when you inform someone who already knows
The result is simpler and less impressive: the same person, deciding better.
They remain responsible, they still understand the process, and they now get in minutes a reading that used to take hours — or that they could not have produced at all. Judgement stays where it was; what changes is how much information is available at the moment of judging.
This is not a new idea. Douglas Engelbart wrote about it in 1962, in the work he called Augmenting Human Intellect — technology as amplification of whoever already knows the problem, not as a substitute for them.
And there is a practical reason it works better: whoever has the pain knows where it hurts. Eric von Hippel documented in Democratizing Innovation, from 2005, that the best solutions tend to come from those who live with the problem, not from those observing it from outside.
Why the easy promise sells better
Because "it replaces" is a short, measurable sentence. Saves this much, cuts this many, pays for itself in this many months.
"It informs better" does not fit a spreadsheet as easily. The gain is real and shows up in faster, less wrong decisions — but it does not arrive with a ready number for the first meeting.
This is not only a worry for people who work. Erik Brynjolfsson called it the Turing Trap, in 2022: when technology is used to imitate the human rather than extend them, the benefit shrinks exactly where it could be largest.
What to do with this on a Tuesday
Before applying AI to something, one question settles a lot:
Does anyone here understand this process well enough to notice when the result is wrong?
If the answer is yes, informing that person is almost always the best use — and the lowest risk. If it is no, the problem is not about AI: it is that nobody understands the process, and automating it will only make that faster and harder to find out.
The test
Look at where AI has already entered your company and ask: is it replacing judgement, or giving more information to whoever judges?
The two look identical in the first month. They come apart when something goes wrong — and then the difference is between someone who can explain what happened and a system that can only repeat.