September 29, 2026 by Patricia Horotan, Co-Founder and Chief Strategy Officer of BlueMatrix.
Progress in AI has always meant the same thing: more. More knowledge. More reasoning. More modalities. More autonomy. We are building toward something that resembles a human mind, only faster and better informed.
Human intelligence did not progress that way.
For most of human history, specialization was a luxury. The doctor was surgeon, diagnostician and pharmacist in one. As knowledge accumulated, that stopped being possible. Medicine split into specialties, then subspecialties. Law, science, engineering, finance - every field deep enough to matter did the same.
We did not become less capable. We became vastly more capable. Civilization was not built by people who knew everything. It was built by people who knew narrow things extraordinarily well, and by systems that let those people work together.
With AI, we are running that experiment in reverse.
The Cost of Knowing Everything.
A single model today can write software, read a balance sheet, translate Mandarin and draft a poem. It is a remarkable achievement. I am no longer sure it is the right one - not for most of what we actually need AI to do.
What if, instead of general intelligence, we built specialized expertise? A narrow model can be trained on a narrow body of knowledge. It can be retrained faster. We can see where it fails. We can limit what it is allowed to do. Most importantly, we can understand its boundaries.
We can wrap our minds around something that has edges.
Much of the anxiety around AI comes, I think, from the fact that it seems to have none. Each generation does something the last one couldn’t, and we describe where all of this is heading in increasingly human terms: intelligence, agency, autonomy. We have set out to build an artificial person, and now we worry about what that person might do.
Most of the World Does Not Need an Artificial Person. It Needs Artificial Expertise.
We trust specialists not only because of what they know, but because we know where their knowledge ends. No one asks a cardiologist to design a bridge. That limitation isn’t a flaw in the cardiologist’s intelligence. It is part of why we trust her to be a cardiologist.
AI may earn our trust the same way: not by proving everything it can do, but by making clear exactly what it cannot.
The economics change too. There is no reason to deploy the computational equivalent of all human knowledge to reconcile a spreadsheet a million times a day. A model doesn’t need to know Renaissance painting to do that job. Why pay for that knowledge? More importantly, why introduce capabilities that have nothing to do with the task?
Every unnecessary capability is unpriced risk sitting on the balance sheet.
Capital Markets Already Know This.
Our industry already works this way. Analysts, portfolio managers, traders, compliance, risk, legal - no one expects any one of them to know everything. Why would we expect the AI underneath them to be different?
One model understands a firm’s compliance rules. Another knows its research history. Another flags when an analyst’s thesis has shifted. None needs to know everything else to do its job well.
And because each has a narrow job, we can govern what it learns from, audit what it produces and test it against a universe small enough to actually understand.
That turns an almost impossible question (can we trust AI inside a regulated institution) into much smaller ones. What can this model see? What is it allowed to produce? How often is it wrong? Who approved it? What changed when it was retrained?
Those are questions institutions already know how to answer.
The Intelligence Was Never In One Place.
Human civilization didn’t solve specialization by eventually producing someone who knew everything. It built networks.
The cardiologist consults the radiologist. Records preserve what happened. Standards allow information to move between specialists. Institutions hold the whole thing together.
The intelligence isn’t in any one person. It is also in the connections between them.
Why couldn’t AI work the same way?
Not one enormous artificial mind, but thousands of specialists. Each knows its job. Each has limits. Each can pass work or information to another when necessary. And each leaves a record of what it contributed.
None Needs to Contain the Whole.
This is, in fact, a much more familiar kind of progress than the one we’ve been pursuing. Give a model a job. Decide what it needs to know. Decide, just as carefully, what it doesn’t. Watch it work. Find its weaknesses. Correct them. Let people get used to it. Then connect it to another specialist.
This is how we learned to trust human expertise. Not because it was limitless, but because it wasn’t.
The first era of AI asked one question: how much can we make a single model do?
The next may ask a better one: how little does a model need to know to do one thing extraordinarily well?
The road to more powerful AI may run in a direction nobody planned for.
Backwards.