Two kinds of knowing.

The book argues that institutions are drawing down an asset they do not count, and that the drawdown shows first in the middle of the organisation. There is an objection to that. We think it is right, and it changes what the framework has to do.

It runs like this. Much of what a middle layer knows is navigational. Who in the company has done this before. Which approval unblocks what is stuck. Whose sign-off is real and whose is ceremonial. That knowledge takes years to acquire and it is worth money. It is a different capacity from doing the work well and from telling when the work is wrong. Much of it accumulated because coordination used to be expensive. Make coordination cheap and some of that layer becomes surplus. Thinning it is a correction.

The distinction has a name and twenty-four years behind it. Harry Collins and Robert Evans drew it in “The Third Wave of Science Studies: Studies of Expertise and Experience” (Social Studies of Science 32:2, April 2002), where the definitions are framed around fieldwork. The general form arrives five years later, in Rethinking Expertise (Chicago, 2007): contributory expertise “enables those who have acquired it to contribute to the domain to which the expertise pertains,” and interactional expertise is “expertise in the language of a specialism in the absence of expertise in its practice.”

Interactional expertise is much of what a good general manager sells. Daniel Wegner’s transactive memory (1986) is its group-level form: an organisation remembers by knowing who holds what, and the middle layer is where that index has lived.

The economics are older. Ronald Coase’s answer to why firms exist at all is that using the market costs something, and a firm appears where internal coordination comes cheaper; Oliver Williamson built transaction-cost economics on that footing. Push coordination costs down far enough and flatter firms are what the theory predicts.

The thinning is on the record. Citigroup’s 2023 Form 10-K tells shareholders the bank “announced plans to reduce management layers from 13 to a median of eight as part of organizational simplification initiatives” — and files it among the risk factors, as a threat to attracting and retaining people. Gap’s 2022 10-K describes “increasing spans of control and decreasing management layers to improve quality and speed of decision making.” 3M’s 2023 10-K puts “reduce layers of management” inside the audited restructuring note, against roughly 6,000 positions and a pre-tax charge of $437 million.

A framework that reads every flattening as a drawdown of judgment capital will escalate a healthy correction, and deserve the reception it gets. The register carries a rule about this: every entry states, in advance, the condition under which it comes back off, because a sheet that records only worsening stops being read. A sheet that fires on every reorganisation stops being read sooner.

The concession stops at one place. Inside a firm, the two kinds of knowing have been carried by the same activity. Command of a specialism’s language is picked up by sitting close to the work, and sitting close has meant doing the junior work — the first-pass draft, the file nobody senior wanted, the meeting you sat in because somebody had to take the note. That is where a person learns who to call, and where they learn to do the thing. Remove the work and both go.

The part we could not answer is this. Nothing a human-resources function runs separates the two. The nine-box grid rates performance against potential. Bench strength counts ready-now successors per critical role. Competency frameworks list behaviours, and succession slates name people. None of them asks whether an individual can produce the work with the tools switched off.

Here the nine-box does something more troubling than stay silent. Potential ratings anchor on observed output, and when a machine carries more of the output the observed work improves. The bench reads stronger on the instrument in the same period it thins underneath. An organisation running that grid through an AI deployment is being told the opposite of what is happening, by a tool it has trusted for decades.

The asymmetry is what puts this on a board agenda. Cut coordination overhead wrongly and you spend a quarter putting it back, at a price you can name. Cut formation wrongly and you learn of it in a decade, in a seat you cannot fill, with no supply to buy because your competitors cut in the same years you did. A director who cannot separate the two beforehand still faces two very differently sized downsides.

How much time you have depends on where you sit. The US Bureau of Labor Statistics put median tenure with a current employer at 3.9 years in January 2024, the lowest since January 2002. The OECD puts average job tenure in dependent employment across the EU27 at 9.98 years for 2025, and Latvia at 9.10. A median and a mean measure different things, so no ratio between them means much. What compares is the short-tenure band: 22.2 per cent of US wage and salary workers had been in post twelve months or less, against about 14 per cent of EU27 employees. An American firm gets fewer years with a person, so formation there has to happen faster.

What the objection changes is the instrument. Before a flattening reaches the succession line, the register owes three answers about the work being removed. Which of these steps was somebody learning on. Who currently senior acquired their eye by doing them. And once the layer is gone, what is left that a new person could form on.

An institution flattening a layer that coordinated things is doing arithmetic, and it should. An institution flattening the layer where its people were made is spending an asset, and it should know that it is. The difference is visible in advance, once somebody builds the test. That belongs in the appendix with the other methods, priced and graded. The objection stays in the book.