The Unautomatables
Everyone I speak to is running AI training right now.
These trainings are designed to teach people how to use the tool, how to prompt more effectively and how to better integrate AI into their work.
Unfortunately, none of it teaches them when to stop trusting it.
There is a study that reveals how expensive that gap is:
758 BCG consultants, given GPT-4 and a business case. They did worse than the consultants working without it. And their wrong answers came out more coherent and more persuasive than the right ones. That is scary.
Same tool, different task, and those same consultants were 25% faster and produced better work.
The researchers call this a 'jagged frontier' or a boundary of what AI does well.
Idea generation feels hard to us. It sits comfortably inside the boundary of what AI does well.
Reconciling a spreadsheet against interview notes feels basic, but it falls outside that boundary.
We cannot tell which side we’re standing on by how difficult the task feels to us.
More prompt training won’t fix this. The skill was never operating the tool. It’s knowing what to outsource to AI and what to do with the output that AI hands back.
I’ve been trying to name the skills that get more valuable as AI gets better. I have six groups on my list so far:
Framing. Working out what the real problem is, before deciding what to do about it.
Judgement. Knowing what good looks like, and being able to say why.
Orchestration. Deciding what you hand to the machine and what stays yours.
Experimentation. Acting before you’re certain.
Synthesis. Turning a mess into a position.
Influence. Getting people to move on it.
I’ve started calling them the 'Unautomatables'.
I’ve been sitting with this list for a while, and I’m not convinced it’s finished.
What am I missing? Which skills or sets of skills become increasingly important and harder to replace as AI becomes better and faster?
(Study link.)
