AI makes specialist knowledge cheap, so the people and firms that win from here are the ones who can see a whole system, close its loops and prove what it does. For two hundred years the sensible move was to go narrow. That move made sense because connecting things was the expensive part of almost any work. AI has made connecting things cheap, and the reasons to stay narrow are going away faster than most experts expect.
By systems thinking I mean a specific habit: draw the boundary of a problem on purpose, find the stocks that build up and drain inside it, trace the feedback loops between them, and look for the delays that make the loops behave in surprising ways. I don't mean "everything is connected." That sentence is true and useless.
Specialization was a way to avoid the cost of connecting things
Adam Smith's pin factory made the case in 1776: split the work into steps, and each person gets fast at one step. Ronald Coase explained in 1937 why firms exist at all: inside a firm, it is cheaper to coordinate people than to negotiate each handoff in a market. Both arguments rest on the same cost. Getting the output of one person, team or field into the hands of the next one is slow, lossy and expensive.
So we built everything to keep that cost down. Professions drew hard lines. Hospitals split into departments, each with its own measures. Software split into products that each did one job well and left the joining to someone else. Research split into fields with their own journals, and peer review rewarded the narrow claim, because a narrow claim is one a reviewer can check.
Systems thinking existed the whole time. Jay Forrester built working models of industrial systems at MIT in the late 1950s and published them in 1961. But building and calibrating those models took months, so most of what came after were diagrams that predicted nothing. The skeptics who called systems thinking a slide deck with arrows were often right.
AI made connecting things cheap
The work that used to sit between specialists is the work AI does best. A model reads across fields that no single person reads. It translates one department's vocabulary into another's. It writes the adapter between two systems that were never designed to talk. It carries context across a handoff without dropping half of it. Each of those took people, weeks and meetings. Now each takes minutes and costs cents.
When the cost that justified a habit disappears, the habit becomes a liability. Narrow teams still hand off work at the old boundaries, still measure only their own part, and still leave open the loops that used to need a person to close.
Cheap expertise is the reason specialists stop being scarce
The common view is that AI favors specialists: it makes the expert faster and better inside the niche. I think that's right and that it leads to the opposite conclusion. If every team can get expert-level help in every niche on demand, expertise in one niche stops being the scarce thing. What stays scarce is the judgment that decides:
- where to draw the boundary of the problem;
- which feedback loops are running inside it;
- which loop was left open because closing it needed people, and can now close;
- where a handoff loses track of who is accountable.
I see this in my own work. When I asked AI for help on one patent at a time, I got ordinary ideas. When I read our 38 granted patents as one system, they turned out to cover every step of a single loop: sensing, establishing trust, deciding, acting, proving that something happened, paying for it and learning from it. Before AI, every step in that loop needed people, so it never closed. Now it can close in seconds, and the patents are worth more together than one by one. No specialist in any one of those patents would have seen that.
Narrow optimization at machine speed does damage
Goodhart's law says that a measure stops being a good measure once people aim at it. With people, that damage builds up over years. With a thousand AI agents each pushing its own metric, it builds up in weeks. A billing agent that maximizes collections, a scheduling agent that maximizes utilization and a messaging agent that maximizes response rate can each hit their targets while the patient waits longer, the staff burns out and the denials go up.
Someone has to own the whole loop to see that. That's the defensive case for systems thinking, and I think it is stronger than the offensive one.
The systems thinking that wins can be run and can be wrong
The objection I take most seriously is that systems thinking without depth produces confident nonsense, and AI makes confident nonsense cheap. I agree. A systems thinker with a whiteboard and no way to test the drawing is no better off than before.
What changes is that the drawing can now run. A model of the loop can be built in days, fed real data and checked against what the system actually did. Depth still matters: AI supplies it in each part, and a person checks it where a wrong answer is costly. John Gall's rule still holds as well: a complex system that works grew from a simple system that worked. So the practical method is to close one small loop, prove it, and add the next.
Own whole loops, and publish what can be wrong
- If you run a company
- Give someone ownership of a loop, not a function. Ask each team which other team's measure moves when theirs improves, and report both numbers together.
- If you fund or buy technology
- Ask which loop a product closes. A product that improves one step of a loop that stays open is usually worth less than it looks.
- If you do research
- State the system you studied, its boundary and its loops, and predict how it will behave over time in a way that can be wrong. We are starting to publish our own research this way, as system maps, each with a model anyone can run once it is built.
I'll call this position wrong if, over the next five years, teams that optimize one part keep beating teams that own and close whole loops in the same markets, with the same AI available to both. I'll record that in the Journal's register of positions either way.