The first tool in history that adapts to you.
For ten thousand years we bent ourselves to our tools. A study of five thousand support agents caught the moment that reversed.
Look at the top row of letters on your keyboard. Q, W, E, R, T, Y. The arrangement was designed, in 1868, to slow you down.
A Milwaukee printer named Christopher Latham Sholes built an early writing machine with the letters in alphabetical order, which was the obvious arrangement and the wrong one. Each key drove a metal bar up at the paper, and bars sitting close together jammed when struck in quick succession. Sholes fixed it not with a better mechanism but with a worse alphabet: he pulled apart the letters English pairs most often, so the common combinations swung in from different directions and missed each other. Remington began manufacturing the machine in 1874, and the layout went out into the world.
The jamming bar has not existed in a working office for the better part of a century. Every keyboard now is glass or silent switches, and any letter could sit anywhere. The layout has not moved. Whole industries rose, shifted from cast iron to silicon, and died while the sequence held. Generations trained their hands to a set of deliberate inefficiencies because the machine demanded it, then taught the habit to the generation after them.
That is the deal the human has had with its tools for ten thousand years. The tool arrives with a fixed logic, and the human does the travelling to meet it. The loom, the printing press, the steam engine, the enterprise software system — each demands that the person learn its language and reshape the working day around its constraints.
In 2023 a study put a number on the deal, at the moment it broke.
A team led by Erik Brynjolfsson at Stanford, with Danielle Li and Lindsey Raymond at MIT, tracked 5,179 customer-support agents at a Fortune 500 software firm over five months as they were given a generative AI assistant. The standard expectation was a flat lift: the tool raises output across the board. The data did something stranger. The most experienced agents gained almost nothing. The newest agents resolved 34 per cent more cases, reaching in months the performance that used to take veterans years.
The detail that matters is which part of expertise the tool carried. The veterans gained least because the thing the model could hand a novice was the codifiable part — the accumulated knowledge of how the common cases resolve. What it could not hand over was the rest of senior judgement: the read on the case that does not fit the pattern, or the sense that something is wrong before the evidence says so. Customer support is a domain where most of the expertise is codifiable, which is why the gap nearly closed. Where the senior's edge is texture rather than codified knowledge, the same tool barely moves it.
So the asymmetry is not proof that experience stopped mattering. It is the first precise measurement of which layer of experience the machine takes, and which it leaves.
It absorbed that layer in a way no earlier tool could. For a century the corporation was built out of human keyboards — workers hired to sit between unstructured human reality and structured corporate systems, translating one into the other. Earlier tools narrowed that distance without closing it. The spreadsheet recalculated on its own, the database answered a structured query, the search engine forgave a clumsy phrase — each an improvement, each still met on the tool's terms.
The generative model is the first tool a person can address in plain language and have it travel the formal distance itself. The earlier tools moved toward the worker. This is the first that comes the whole way.
Set that beside the keyboard. For ten thousand years the human did the adapting and the tool held still. Here, at last, is a tool that adapts to the human.
Which makes its arrival inside the corporation strange to watch. A technology that finally moves toward the worker is landing, in most organisations, as the same exercise every earlier technology triggered: buy the capability, train the workforce to the interface, book the efficiency. The consulting layer has always treated adoption as a straight line — purchase, train, yield — and it is treating this one the same way.
But a tool that comes the whole way to the worker is not a faster keyboard. It is a different kind of object, and the question is what an organisation built to make humans adapt to tools does when it is handed a tool built to adapt to humans.
The capability is never the variable. The architecture receiving it is.
The earlier tools moved toward the worker. This is the first that comes the whole way.
Klarna learned the distinction in public. Between 2022 and 2024 the payments company replaced around seven hundred customer-service roles with an assistant that, at its peak, handled two-thirds of customer interactions. The volume metrics looked decisive — resolution rate, time to first response, tickets per hour. Then the quality metrics arrived. Customer satisfaction fell on exactly the interactions the volume metrics could not see: the complex and the nuanced, the cases that needed a human read. By 2025 the chief executive conceded the company had gone too far, that the drive for efficiency had cost it quality and trust, and Klarna began rebuilding human capacity — into a hybrid model where the assistant holds the high-volume routine and people hold the complex and the high-value.
The AI did its job. The volume metrics measured the layer it was good at and said nothing about the layer it was not, and an organisation reading only the flattering numbers will not learn the difference until the quality has already gone.
This is where the keyboard and the support study meet. The same model is two entirely different instruments depending on what the architecture points it at. Aimed at the workforce as a cost to remove, it absorbs the codifiable layer, the volume metrics glow, and the texture — the part of the work that was never codifiable — quietly leaves. Aimed at the workforce as the thing to amplify, the same tool carries the routine and frees the human for the judgement only the human holds. The model does not choose. The architecture chooses, and most architectures were built on the older premise, where the worker is a keyboard to be made redundant rather than an author to be freed.
For ten thousand years a tool could be adopted by being installed. The user met its interface and the shape of the organisation stayed intact. This is the first tool that has to be adopted the other way round — by being built around. A firm that bolts it onto the old structure gets Klarna's first eighteen months: the volume rises, the quality drains, the trust erodes, and the dashboard says everything is fine until the moment it does not.
So the question for an executive is not how capable the model is. On capability the answer is settled, and it improves every quarter on its own. The question is what the organisation has pointed it at. A tool deployed to thin the workforce and a tool deployed to move toward it are the same software and opposite instruments, and the metrics that flatter the first are the ones that hide its cost.
The keyboard can finally be put down. What an organisation builds in the space that opens is the whole of the decision.
The final section — the reframe — is reserved for members. Enter your email to continue reading.