Why AI that learns your business can deliver better work at lower cost—and give Adam Smith’s division of labor a new lease on life.
In 1767, Adam Smith, a Scottish philosopher in his 40s, moved back in with his mother so he could focus on his writing. Britain's Industrial Revolution was just getting started. Nearby, the Carron cannon factory was fast becoming one of the biggest in Europe. But it was something unexpectedly humble that inspired the defining idea of the industrial age: the pin.
Here’s how Smith put it in his founding treatise of modern economics, The Wealth of Nations:
"One man draws out the wire, another straights it, a third cuts it, a fourth points it, a fifth grinds it at the top for receiving the head," he wrote. Split into about eighteen steps, he observed, ten workers could make upwards of 48,000 pins a day. A worker doing every step himself "could not… have made twenty, perhaps not one pin in a day."
Two and a half centuries later, his idea that workers are far more productive when they specialize is still one of the foundations of economics. But today, as we add agents to our division of labor, it's worth asking whether his idea still holds when a single worker is capable of every step.
AI agents aren’t constrained by the limitations of Smith’s pin factory workers. No one human can be superb at all 18 steps; but researchers can train a model that knows every step perfectly, and make a whole virtual factory of them. And for a while, it seemed Smith had met his match. Early attempts to build specialist models couldn't keep up, because they were frozen on top of AI base models that quickly became outdated. They were like the sat-navs, digital cameras and MP3 players of the 2000s: excellent at one job, until the smartphone came along and got so good, so fast, so generally, that it did each of their jobs better than they did. Companies would spend months building their own custom models that were really good at a specific task, only to watch new general models from the big labs overtake them even at their specialization.
But that’s changing, because now we can make specialists that learn on the job. Much of AI’s recent progress comes from reinforcement learning, in which a model practices a task again and again and is rewarded each time it does well, until it gets reliably good at it. That training is now cheap enough to run on open models in days rather than months, using the signals a business already produces: every edit, retry and correction its people make. The gains it brings, however, favor narrow, specialized AI. Researchers have found that a model trained this way gets much better at the kind of task it practiced, but those gains don’t carry over to other kinds of work, especially in less structured fields like law and finance. And because that specialized training is now cheap and fast, it can now happen continually, not once-off: platforms like Trajectory turn each day’s corrections into ongoing training for a model the company owns. And when a better open model comes out, those lessons carry over with little more than a settings change. That makes today’s specialists less like sat-nav gadgets and more like apps, which get better every time the phone they run on updates.
And when specialists can keep up, they win.
Specialists have two built-in advantages over one do-everything model.
The first is cost. Building on Smith, the mathematician Charles Babbage explained why dividing work among specialists is cheaper. If one worker made the whole pin, you’d have to pay a skilled craftsman’s wage even for the simplest step, like pressing finished pins into paper. Dividing the work, however, lets you “purchase exactly that precise quantity” of skill each step needs. A do-everything model is built for the hardest problems, so you pay frontier prices even for routine tasks. A specialist agent is the opposite; it’s trained with the “precise intelligence” for its job, and can beat a do-everything model on its specialization at a fraction of the price. In legal tests run by Trajectory, a trained open model matched the leading frontier models at under a tenth of the price, and an even smaller one outscored Anthropic’s Opus 4.6. But won’t that edge vanish as do-everything AIs keep getting cheaper and better? Not really, because specialists ride that same cost curve down as well. And as agents take on more and more of a firm’s work, AI becomes a bigger share of its costs, and paying only for the skill each job needs saves more and more.
The second is quality. Specialists learn what matters to your business. Smith's first explanation for the pin factory's output was the "increase of dexterity in every particular workman": do one step thousands of times and you get very good at it. Specialist agents get good the same way, by doing your work over and over and learning from every correction. A specialist trained on your own data learns how your company actually does the work. AI labs pay industry experts billions to teach their models what good work looks like, which shows how valuable that know-how is. But they can only teach what's typical across an industry. It sometimes works, but will still be out-competed by the highest-quality results coming from leveraging data existing only inside your company; in your people's edits, judgments and past mistakes.
Because of this, AI labs and firms alike are converging on specialists. Smith's deeper point was that the bigger the market, the more finely work gets divided. In a remote village, he noted, every farmer "must also be butcher, baker, and brewer" for his own family; only a large town has enough work for a full-time porter. AI is being put to work in almost every industry and on almost every kind of task, which makes it one of the largest and most varied markets a technology has ever had. By Smith's logic, that's exactly where specialization goes furthest.
So AI doesn’t make specialization obsolete. Even when one worker can do every step, it still pays to give each step to the worker trained for it. The most efficient division of labour remains a system like Smith’s pin factory: specialists learning their craft through continual learning on the job, getting better with every correction, at a fraction of a frontier model’s price. This is where the next leaps in AI cost efficiency and quality will come from. Rather than proving Smith wrong, all that has changed in the 250 years since he wrote is the worker, and how it gains its expertise.
The future of AI at work isn’t one genius. It’s a team.