There’s a part of every important decision I can’t hand to AI. Not because the tool falls short. Because the part that matters was never words to begin with.
I’ve spent the better part of my life learning to read a room, and you pick things up you could never write down. The pause before someone says yes when they mean no. The sense that a plan is wrong before you can say why. The read on whether a team is tired or checked out. Whether someone needs an answer or just needs to be heard. None of that arrived as information. It came from being there, again and again, until the pattern lived in me somewhere below language.
That’s the thing about the knowledge that matters most. You know more than you can say. A tool works only with what can be said, so when I hand it a decision, I’ve already left out the part that would actually decide it. It hands me a clean, confident answer built on everything except the thing I couldn’t put into words.
I’ve felt the pull to fix that by feeding it more… more context, more background, longer and longer prompts, until surely it has enough to tell me what to do. It never quite works, and when I’ve leaned on it anyway, I can feel what I gave up. You can turn judgment into a list of rules, but the moment you do, it stops being judgment. The whole value of experience is that it handles the case the rules didn’t see coming.
So I’ve stopped asking the tool to carry the parts it can’t. I use it for the explicit work. Draft this, summarize that, show me what I’m missing, check my reasoning. And I keep the in-between work where it belongs, with people.
Here’s what I’ve noticed about that in-between work. It’s the same ground as making someone feel they matter. Being seen is not a task you can delegate. The hard conversation, had honestly and made safe, has to come from a person who’s actually in it. The judgment call with your name on it stays yours precisely because your name is on it. The apology, the thank-you, the moment you tell someone the truth because you respect them, all of it depends on a person being present, not a good draft.
That’s not a limitation AI will grow out of. Better models will get better at what can be said. They won’t get better at what can’t, because that isn’t a modeling problem. It’s the difference between information and presence.
Knowing which work is which is its own skill, and I’m still learning where the line sits. Some days I hand over too much and feel the work go flat. Some days I hold onto something the tool could have helped with. But the direction is clear enough. The easy, explicit parts can go to the machine. The parts that need judgment, that need someone to feel seen, those stay with us. That isn’t AI’s weakness. It’s the whole point of being the human in the loop.
So I’ll ask you the question I keep asking myself. What’s the work you’d never hand to a machine, even if it could do it? That answer is usually the work that matters most.