Every firm now holds a stake in goodwill it never wrote down. The knowledge your next model needs sits in your employees’ heads, and the stopwatch no longer works.
The head of operations at a manufacturer I work with has been running his weekly reports through a language model since the spring. The output is good, his afternoons are quieter, and, as far as he knows, nobody has formally asked how the work is now being done. Two floors up, the board of the same company has been discussing how to capture what its most experienced people know — the pricing instincts, the supplier folklore, the feel for when a delivery promise is safe to make — into systems that would survive their departure. Neither party has mentioned any of this to the other. Both consider themselves rather clever.
Versions of this standoff are visible. Employees adopt AI privately and keep the productivity gain, boards plan to mine what employees know and keep the model. Each side is convinced it discovered the value first, and each is accumulating quietly while the terms remain unwritten.
The second half of this game has been played before. In the early twentieth century Frederick Winslow Taylor sent men with stopwatches and notebooks onto American factory floors to extract what he called the “mass of rule-of-thumb or traditional knowledge” inside workers’ heads. Once codified, the knowledge changed hands: under scientific management the worker was told what to do, how to do it, and exactly how long it should take. Sarah O’Connor drew the parallel in the Financial Times this week — the knowledge that would make AI genuinely useful in most workplaces is locked inside employees’ heads, and employers have begun to see it. It is worth remembering how Taylor’s version ended: strikes, a congressional investigation, and a congressional ban on stopwatch-based time studies and premium systems in federal arsenals, which remained in force until 1949. The workers understood that a knowledge grab is a power grab, and priced it accordingly.
The instinct today is the same as Taylor’s
The difference is that extraction no longer works, and the reason is mechanical.
A stopwatch could time an arm movement whether the worker consented or not, because movement is observable by third parties. Judgment is observable only by the person who holds it. Fine-tuning a model on expert knowledge means asking the expert to produce examples of that judgment: this invoice is anomalous, that one is noise; this deadline is safe, that one will burn us. To verify the examples you would need to know the right answers — and the only person who knows the right answers is the expert you are asking. The ground truth lives in the same head that produces the data.
Withholding does not even require sabotage. Seventy per cent effort is enough: hand over the obvious cases, the ones any trainee could label, and keep the boundary cases, the home of everything that made you worth asking. A lazy label is plausible, coherent and defensible. Nobody can distinguish “I gave you my best criterion” from “I gave you the criterion anyone would give”. The failure surfaces months later, when the model turns out mediocre, and the autopsy cannot assign blame: thin data, wrong architecture and defensive experts all leave the same corpse.
Bridgewater Associates ran the experiment that shows what is at stake. The firm tried general-purpose models on a task its analysts perform daily — reading news and filings for information relevant to investment decisions — and found the best of them right about half the time. Better prompting did not move the needle; as the team observed, a prompt can only carry the intuition an expert manages to put into words, and the judgments that matter most are the hardest to articulate. Fine-tuning a smaller model with its own experts and data took accuracy to about 85 per cent. The jump from coin-toss to useful lives almost entirely in the hard cases — precisely the ones a wary expert withholds. The quality of a training signal depends on the goodwill of the person producing it, and goodwill cannot be audited.
The stopwatch is no longer needed
The obvious rejoinder is that nobody needs to ask any more. Microsoft’s Work IQ, generally available since June, builds what the company calls a real-time model of how your organisation operates by continuously processing email, meetings, chats and files — Taylor’s stopwatch, reissued as a subscription. Watching does capture more than it used to; meetings that once evaporated are now transcripts. But telemetry records the work that got written down, and boundary judgment mostly never is: the deal not pursued, the number that smelled wrong, the reason an exception was waived. Bridgewater swims in recorded data and still needed its experts to label by hand. What passive capture genuinely changes is the ownership question — whose layer holds the resulting model of your firm — and that negotiation has a different counterparty, in a different city.
Seen from the employee’s side, wariness is rational. Tacit knowledge is the one asset a professional owns that does not transfer: the degree transfers, the CV transfers, the feel for how this firm actually decides does not. Handing it to a model is an irreversible transfer of a bargaining asset. The people who hold this kind of knowledge are, in my experience, competent and loyal, and their defence of it is honest — anyone would protect the thing that makes them hard to replace. A firm that asks for the download without stating the terms is asking its best people to act against their own interest and hoping they will not notice. They notice.
The conditions under which people hand such knowledge over willingly are documented, because that has happened before too. In 1984 General Motors reopened its Fremont plant — by most accounts the worst in its network, with absenteeism above 20 per cent and cars occasionally sabotaged on the line — as a joint venture with Toyota, staffed largely by the same workers. The venture opened with a commitment GM had never made: the company would not use the workers’ own improvement suggestions to eliminate their jobs. Under the Toyota system the same workforce began proposing process improvements at rates GM had never managed to buy from it, and within a few years Fremont was building some of the best-quality cars the group sold. The commitment came before the collaboration, in that order and no other.
Which gives the practical rule its shape. If the next phase of AI in your firm involves tuning systems on what your people know — and outside software it mostly will, because the public internet has already been eaten — then the agreement on who keeps the gains has to be visible before you ask for the knowledge, since the quality of what you receive depends on it. There is a Monday-morning test. Announce the programme, and watch what your best people bring you. If it is the boundary cases, you have a contract. If it is the cases a trainee could have labelled, you have a Taylor problem, and no tooling budget will fix it.
One question survives all of this, and it is harder. The tacit knowledge every firm now wants to mine was formed over years of doing precisely the work the machines are beginning to absorb — the reconciliations, the drafting, the triage. Extract the seam and retire the miners, and the next model has nobody left to learn from. Where the following generation of judgment is supposed to come from is a different question, and the next piece takes it up.
If this trade-off is already playing out in your organisation — people quietly fine-tuning models on what they know, while the terms stay unwritten — we help firms work out where that knowledge sits and negotiate for it properly. Get in touch with Swell.
For the judgment calls specifically financial — the kind Bridgewater’s data above was measuring — we built FinModeler to close that exact gap: not a generic model guessing at your numbers, but one that works from your actual assumptions. Build your first model free →
