Automating tasks or augmenting decisions

One evening I wrote down everything I had done with AI over six months. Fourteen things, and at first the list looked good. Then I tried to write next to each one what had changed because of it. Five of them got nothing. Not because nothing had changed, but because I never measured it and never went back to look.

Fourteen experiments are not a strategy. That is a hobby, only expensed.

I see the same thing inside companies. Marketing has its own tool, customer support is testing another, and in finance one person has built something useful and told nobody. Ask what changed and the room goes quiet.

Two boxes that look alike

Almost everything we do with AI fits into two boxes. They behave nothing alike, which is why it is worth separating them before buying anything.

Automation fits where a task repeats and has a correct answer. Booking invoices, sorting documents, a first translation draft, moving data from one form into another. The goal is to take the human out of the chain, and you measure it in time and in errors that come back.

Augmentation fits where somebody has to decide. What price to offer. Who to hire. Which project to stop. Here AI decides nothing, it only widens the picture: more options, a counterargument, reading you would never get through yourself. The goal is a better decision, not a faster one.

The difference is easy to spot. If two smart people can look at the same output and disagree, it is a decision. If there is nothing to disagree about, it is a repetitive task.

The expensive mistakes come from mixing the boxes. Automate a decision and you get confident nonsense that nobody checks any more. Augment a repetitive task and you get an expensive toy that offers three options every morning where one button was needed. I made the second mistake myself and paid a subscription for it for three months.

Turning experiments into an operating model

This takes three things, not ten.

1. Two lists. What repeats and what gets decided. Written by the people who do the work, not by the leadership team.
2. An owner for every item. Not a team, a person with a name. No owner, no item.
3. No more than five numbers for the whole company. Two on automation: hours freed per month and errors coming back. Two on decisions: days to decide and how many decisions get reversed later. One on usage: how many people use it every week without being told to.

Five numbers look like too few. But nobody reads twenty numbers, and everybody remembers five.

A warning about that last number: the moment usage becomes mandatory it stops telling the truth. People open the tool so it gets logged, then close it.

How I use this in my own week

When a conversation about a new tool starts, my first question is which box it goes in. If the answer is "both", we do not yet understand what we are buying, and the conversation is better postponed by a week.

I do the same with myself. Repetitive things get handed over and I do not go back to them. Where a decision is needed, I use AI as an opponent: give me three reasons my plan will not work. Once in a while one of those three is good enough that the plan has to change. I still make the call myself, because responsibility cannot be handed to a machine, however tempting that would be.

Automation frees up time. Augmentation decides whether that time was worth anything.