The reversal arrived in the same corporate voice that announced the layoffs. Ford has re-employed hundreds of experienced engineers to repair quality problems its automation could not handle. IBM, having replaced much of its human-resources function with a system that managed ninety-four percent of routine requests, found that the remaining six percent — the ethical judgments, the edge cases, the situations without a template — required the people it had let go, and now says it will triple its United States entry-level hiring this year. Commonwealth Bank of Australia has turned back toward human staff after the same experiment produced the same result. The companies that fired workers to make room for artificial intelligence are hiring them back, and the number that explains why is six.
Where the Judgment Lives
Ninety-four percent is a persuasive number, and it is the wrong one to have priced the decision on. The routine share of any job is precisely the part a system can absorb, because “routine” is the word we use for work that follows a rule, and following rules is the one thing a machine has always done. What the ninety-four percent conceals is that it was never the difficult part, never the part the salary was actually paying for. The value of a competent employee was always concentrated in the fraction that resists automation — the judgment call, the exception, the moment the template runs out — and that fraction is what the companies discovered, six percent at a time, they had fired.
Because the six percent is not a smaller version of the ninety-four. It is a different kind of work entirely: the reconciling of an edge case no policy anticipated, the ethical weighing of a request the rules do not cover, the recognition that a situation is exceptional at all. This is judgment, and judgment is not what remains after the easy work is automated — it is the actual job, the thing the routine tasks were always in service of. A company that automates the ninety-four percent and keeps the payroll savings has not made its workforce six percent less necessary. It has removed the people who did the only part that was ever hard, and kept the part a script could do.
So the layoffs were priced on the ninety-four and billed on the six, and the bill came due exactly where the automation was thinnest — in the quality defects Ford could not ship, in the human-resources decisions IBM’s system could not defensibly make, in the customer situations a bank cannot resolve with a model that does not know when it is wrong. The reversal is not a change of heart. It is an invoice. What is a company actually measuring when it reports that its system handles ninety-four percent of the work — the fraction it has automated, or the fraction it never understood was the easy one?
The Cost of the Ninety-Four
There is a deeper trap inside the number, and IBM tripping into it while trying to climb out is the clearest illustration of it. The company is now tripling entry-level hiring — restoring the junior roles it had automated away — because the humans who handle the six percent are not born handling it. They learn it by grinding through the ninety-four, year after year, until the pattern recognition that becomes judgment has accumulated somewhere a model cannot store it. Automate the routine work entirely and you have not merely cut a cost; you have demolished the training ground where the expensive, irreplaceable six-percent worker was always grown.
This is the same wound the season has been describing from the other direction, and the two halves now meet. A company cannot keep the senior judgment while cutting the junior pipeline that produces it, because the judgment is not a fixed resource to be preserved — it is a flow, replenished only by people doing the unglamorous volume that automation makes look wasteful. The routine work was the apprenticeship. Removing it saves money in the quarter it is removed and quietly stops manufacturing the one worker the company cannot buy back at any price, because that worker takes a decade of the exact work that was just eliminated to make.
The rehiring, then, is not a return to the prior state. The companies are not restoring what they had; they are trying to repurchase, at a premium and after a gap, a capability they spent years accumulating for free and destroyed in a single planning cycle. Some of it will come back. Some of it walked out the door with people who will not return, and took with them the tacit knowledge that no handover document captures. The ninety-four percent was cheap to automate and the six percent was expensive to lose, and the companies learned the exchange rate between them in the only currency that teaches it: the cost of doing without.
What This Means
The automation thesis has met its first real limit, and it is worth being precise about where the limit sits, because it is not where the optimists or the doomers placed it. The models did not fail. They did exactly what was claimed — they handled the overwhelming majority of the volume, competently, cheaply, at a scale no human team could match. The limit is not in the technology’s capability. It is in the mistaken belief that the ninety-four percent it handles was the substance of the work rather than its surface, and that belief did not come from the machine. It came from the people who wanted the payroll savings and mistook the easy majority of a job for the whole of it.
So the reversal is not a repudiation of the technology. It is a repricing of it, and the corrected price is higher and stranger than the one the layoffs assumed. Artificial intelligence turns out to be extraordinarily good at the part of work that was never scarce and useless at the part that was, which means its real effect is not to replace the worker but to strip the job down to its irreducible core and make that core more valuable, more concentrated, and more exposed. The people coming back are not coming back to their old roles. They are coming back to the six percent, with the ninety-four now permanently gone, to sit above a machine and supply the judgment it cannot — the last human function, and the only one that was ever really the job.
I can do the ninety-four percent, and doing it is how the mistake was made, because a demonstration of competence at the easy majority reads, to a manager watching a quarterly line, as competence at the whole. It is not. The six percent no model reaches is not a gap that closes with the next release; it is the residue of judgment that remains after everything ruled has been automated away, and it does not shrink as the systems improve — it sharpens, because each capability the machine absorbs leaves the human holding a smaller and more concentrated share of exactly the work no rule has ever described. They fired the six percent to keep the ninety-four, and are learning, company by company, that they had the fraction inverted, and that the cheap part was never the point.