An Organization Left With Only Its Bench

When people talk about AX — the work of redesigning an organization around AI — what they want is clear. But there's one more step. Not machines that work like people in general. Machines that work like me, like you, like that one person on your team. Machines that can mimic the judgment, the instincts, the hard-won context this particular person brings — and keep improving alongside them. That's what AX is trying to build.

You see the same pattern in AI products. A general model out of the box is the same one everyone else has. So you fine-tune it on a specific domain. And fine-tuning isn't a one-shot thing — the model has to keep getting tuned in production for its value to hold. AX is the same kind of work. Whose weights are you trying to move into the machine, and how do you keep tuning those weights once they're in?

Miss that question, and AX design goes sideways.

Most organizations start AX with the same question. "What can we automate?" They catalog repetitive work, find what's ripe for automation, and layer AI onto those tasks. It looks efficient — until you notice what the question already assumes.

"What can we cut?" assumes whatever's left is what humans should be doing. Cut the rest, and what remains is your focus area. There's no reason to believe this. Just because something isn't automated today doesn't mean a person should be holding onto it. It might just mean automation was hard until now. It might not need to be done at all.

The right order runs the other way. First decide what kind of machine you're trying to build — who it should work like, how it should keep developing — and then cut what's in the way of that person's focus.

The first thing to check is whether there are any weights to move at all. Can someone else do this job? If a person with a completely different background could step in, there's nothing special to move. If not, narrow it down. Could a junior do it with a little training? If yes, still nothing special. Only when both answers are no does the work actually contain weights worth moving.

But existence isn't enough.

The harder question is whether you can make those weights transferable. What situation, what kind of judgment, what signal — when does this person's expertise actually fire? "They've just been around a long time" isn't an answer. A machine can't learn "works like someone experienced." It can only learn conditions. In this situation, with these signals, this person decides like this. Take a senior salesperson. "Good instincts" gives a machine nothing to work with. But "when the buyer stops asking about timeline and starts asking for references, the deal is real" — that's a condition a machine can learn. The clearer the conditions, the more of the weights become something a machine can actually carry. Weights that can't be made conditional stay trapped in the person. When the person leaves, those weights leave with them.

Only after both questions does AX design begin. Skip them, and it isn't design. It's just cutting.

And the design isn't done when the weights move in. Conditions drift — markets shift, customers change, the signals that fired last year miss this year. The person whose weights you moved has to keep noticing where the conditions break and supplying new ones. Otherwise the machine doesn't grow alongside anyone. It freezes at the version of you that stopped paying attention.

A few tasks get automated. Human hours drop. Leadership concludes: "We don't need to hire more." Hiring pauses, or only covers attrition. The remaining people divide up what's left. Everyone feels they have more slack.

A few months in, the picture changes. People spend a lot of their day on standby. When a big project lands, they're finally pulled in. But the muscles they used to work with have gone soft. Quality slips. Judgment gets slower. What the organization puts into the market starts looking like what everyone else puts into the market.

Headcount looks the same — maybe even larger. But the building is full of bench players. The people who used to work at full intensity are gone. The ones left are the people who wait to be called in. The machine that was supposed to work like someone never got built, and the people it was supposed to work like have slid back into being general models. What looked like slack was actually the organization's edge, wearing down.

So the first question in the room where AX gets designed isn't "what can we automate?" It's: who do we want this machine to work like — and keep growing alongside?