The first generation of frontier AI had an enormous advantage: much of the knowledge it needed had already been digitized.
Law had cases. Medicine had research and clinical literature. Finance had filings, markets, and decades of recorded analysis. The plant floor had manuals too. What it didn't have was a record of everything experienced people learned between the lines.
Nobody documented which valve you crack first on a cold start in January, or why the line jams on third shift and not first, or the workaround the field tech has used for eleven years because the manual is wrong and everyone knows it.
That knowledge lives in the people doing the work.
And it is leaving.
The labor shortage is also a data shortage
The labor shortage in these industries is usually told as a shortage of workers. Not enough welders, not enough linemen, not enough people willing and able to work a plant floor. That's real. But the number of open seats understates what's disappearing.
The Census Bureau's Business Dynamics Statistics of Human Capital shows how dramatically the workforce inside these industries is aging. In 2006, 35 percent of employment in utilities was at firms where at least 25 percent of workers were over 55. By 2022, it was 80 percent. Manufacturing and wholesale trade saw the same pattern: the share of employment at firms where at least a 25 percent of workers were over 55 rose from roughly 14 percent to more than 40 percent over the period.
When a thirty-year operator retires, the company doesn’t write anything down. There was never an asset on the balance sheet, so there's no loss to record. But something valuable left the business: years of accumulated knowledge about how its equipment, processes, and exceptions actually work.
Seen that way, the labor shortage is also a loss of training data that was never captured. Some of the most valuable operational datasets in the physical economy have never been recorded, and they're depreciating on a demographic schedule nobody can extend.
Why this didn't work in 1985
The obvious objection is that people have tried to capture expert knowledge before. Expert systems tried to bottle the master operator in the 1980s and largely failed.
One of their biggest problems was how they captured knowledge. They asked the expert to write the rules. Sit the veteran down, interview him, encode what he says into if-then logic.
But experienced people can't always explain what experience has taught them. The veteran knows the pump sounds wrong. Ask him exactly why and you get a shrug.
What's different now is that you don't have to rely on the expert articulating the rules. You can instrument the work instead. Sensors, voice, video, and models in the loop observe what the operator actually does and the context in which he does it.
The 1980s approach asked the operator to write the manual. The new one watches him work and infers it.
The harder objection is that instrumentation captures what the operator did, not everything he considered and rejected. The judgment that matters most is sometimes the intervention that didn't happen: the alarm he ignored because he had seen that pattern a hundred times, the shutdown he chose not to call. A camera doesn't see a decision not to act.
That's a real limit. Some judgment will still be missed.
It's still worth building, because the alternative is capturing nothing, and because a record of what experienced people do under real conditions, across thousands of shifts, compresses a learning curve even if it never fully reconstructs the expert.
Capturing the residual judgment is the harder problem. But we can now capture far more of the work that surrounds it.
Augment first, and know why
The near-term products don't look like autonomy. They look like tools that make the remaining workforce more capable: guidance that shortens the learning curve, systems that surface the exception before the new tech hits it, software that finally treats the operator as the actual user.
As experienced workers use these products, pieces of their judgment become data instead of disappearing at the end of the shift. Every shift makes the product more useful and the underlying dataset harder to replicate.
Augmentation is how the knowledge gets collected. Automation is what you do once you have it.
Who owns the floor
There's a catch in that sequence.
The knowledge is being captured inside a plant, fleet, jobsite, or utility, from the people who work there, about the equipment and processes they operate. Whether the resulting dataset becomes a moat depends on who owns it, and the honest answer is that much of the physical economy hasn't settled that question yet.
Companies that start capturing this knowledge early have an advantage. As customers recognize its value, they will become more deliberate about who can access it and what they can do with it. Plants will start asking the question, and the companies that survive it will be the ones whose value comes from what they learn across customers, not from holding any single customer's data hostage.
The model of an industry, improved by patterns learned across a thousand floors, can be defensible. The raw feed from one of them isn't. Getting that structure right early is the difference between a moat and a liability.
What the wrong bets look like
If this is right, it changes how some of the most crowded bets in the physical economy should be evaluated.
Humanoids without judgment. Robotics can solve for hands. It still has to solve for knowing what to do with them. A robot that can walk the plant still doesn't know the pump sounds wrong. Embodiment without operational intelligence solves only part of the problem.
Models before data. Frontier models will keep getting better at what humanity has already recorded. The leverage in physical industries is in what hasn't been recorded, and no amount of compute can train on data that doesn't exist.
Building for management instead of operators. Enterprise software has done far more to digitize how physical industries are managed than how the work itself gets done. Labor scarcity changes the economics of closing that gap. As experienced workers become harder to replace, the value of making each remaining worker more capable rises with them. What looks like wage pressure from one side of the P&L becomes demand for technology on the other.
The clock
The window to capture what the last generation knows is closing with every retirement. Once that knowledge leaves, some of it can be reconstructed. Some of it can't. What disappears will be far more expensive to relearn than it would have been to preserve.
The teachers are still here. For now.
-Kiswana