After the Triangle — Part 3/5, AI and organisational structure

When One Person Stands on All Three Levels

Last spring I sat at my computer one evening with a task that, a few years ago, would have taken three people and a week. I had to work out which way a client’s project ought to go, build a rough plan for it, and produce a first version of the materials. Before, this would have travelled down the staircase: someone thinks out the direction, someone breaks it into tasks, someone does it. Now I did all three myself, in one sitting, with AI beside me. I asked why, planned how, and produced what — one after another, within the same evening, without having to send anything to anyone and wait for it to come back.

I’m not telling this to boast. I’m telling it because something happened there that is the heart of this whole series. The pyramid assumed that those three levels are held by different people, and that work moves between them slowly and laboriously. That evening the three levels were in one head, and the work didn’t move anywhere — it just happened. It’s time to talk a bit about AI and organisational structure.

AI strikes at the foundations of the pyramid

In the previous part I described why the pyramid was once sensible: information moved slowly and expensively, so it paid to gather it at the top, make the decisions there, and send them down the staircase. The whole structure rested on this one assumption. AI strikes at exactly that — and from three directions at once.

The first blow lands on the cost of moving information. Organisation researchers have long shown that the layers are fundamentally machines for compressing and filtering knowledge: the lower level gathers and condenses, the upper level receives a ready-chewed summary rather than raw data. Think of an ordinary weekly report: someone collects the figures, someone condenses them, someone turns them into a slide, and only then do they reach the leadership team’s table — and along the way some of the information is filtered out, some is distorted, and all of it takes time. The more expensive this gathering and moving was, the more layers it paid to have doing it. But that is exactly what language models now do almost for free. They read, structure, and summarise information in seconds. When compressing and moving information no longer costs much of anything, the very reason for having a layer in the middle to do it disappears too. A filter whose existence used to be expensive becomes the press of a button.

A pace you don’t set yourself

The second blow lands on speed. Bound up with this is an idea put forward by the American fighter pilot and strategist John Boyd. Boyd was an odd genius: he barely wrote, presenting his ideas instead as briefings he refined over decades. His best-known idea is the OODA loop — observe, orient, decide, act. Boyd’s point was that the side that cycles this loop faster gets inside the opponent’s decision cycle: it makes the other’s observations obsolete before they have even oriented, and produces a confusion that eventually paralyses.

The important thing to notice is which stage Boyd considered most essential. Not speed in itself, but orientation — the moment in which observations turn into understanding. Boyd said that even perfect information is of no use if it isn’t connected to a sharp sense of what it means. What decides is judgement, not the amount of information. And it’s here that AI does something interesting: it compresses observing and acting — the gathering of data and the making of outputs — but orientation, judgement, still falls to the human.

The flip side of this is uncomfortable. When the machine accelerates the loop, the pace is no longer set by you. It’s like cycling in the lead group: the speed is dictated by the front, and if you can’t keep up, you’re dropped — and once dropped, you’re hard to reel back in. AI raises the pace of the whole group. That means a person too has to move information and make decisions faster than before, not because it’s pleasant, but because the slow one ends up outside the loop of their own field. The Spanish company Zara is an old example: it takes a collection from design to the shop floor in two weeks, where the industry traditionally takes around five months. When someone cycles the loop that much faster, the others aren’t competing by the same rules — they’re competing in a time already gone.

The layer that gets eaten first

The third blow lands on what is done at the base of the pyramid. The operational WHAT level — routine collation, reporting, first drafts, information retrieval, preparing the analysis — is exactly what AI takes over first. This isn’t a guess. Harvard and MIT researchers followed over five thousand customer-service agents and found that AI raised productivity by around 15% on average — and the ones who benefited most were the least experienced, whose work improved towards the level of the best. The study has since progressed to peer review, which makes it weightier than usual. Stanford researchers, for their part, found in a large payroll dataset that the employment of young, entry-level workers has fallen precisely in those occupations where AI replaces codified knowledge — the kind learned from books, with which a beginner usually starts. In Finland, the same pattern has so far not appeared. When ETLA, the Research Institute of the Finnish Economy, replicated the Stanford design on Finnish data, the employment of highly exposed young people tracked everyone else’s almost identically; the researchers’ interpretation is that the Nordic labour-market model and strong dismissal protection dampen technological shocks. A shock absorber doesn’t remove the blow, though; it softens it. The difference is not that the shift will pass us by — it’s that we have a moment more time to decide how to meet it.

This includes an example that is constantly misread in both directions. The Swedish company Klarna announced in early 2024 that its AI assistant had handled, in its first month, two-thirds of customer-service chats — a workload equivalent to about seven hundred full-time people. Resolution time dropped from eleven minutes to two. The hype camp draws its conclusion from this: AI replaced the people, full stop. The opposing camp seizes on the fact that Klarna later started hiring people again for the more demanding cases, and proclaims: look, AI doesn’t replace the human after all.

Both readings are dishonest. The honest picture is more tangled. Klarna now genuinely makes larger revenues with a clearly smaller staff — that is not an illusion. At the same time the nature of the work changed: what came in its place was not the same kind of full-time, place-bound work, but more flexible gig work that can be done by, say, a student or someone living in the regions, for whom full-time office work would not have been on offer. Some lost steady work; some gained work that didn’t exist before. So AI neither replaced the human nor failed to replace them. It ate the repetitive part of the operational layer — that routine WHAT — and left the human what the machine can’t do: the awkward case, the judgement, the responsibility. Two sides, and neither extreme.

The machine doesn’t eat everything, though, and it doesn’t eat evenly. Harvard and BCG researchers gave well over seven hundred consultants tasks to do with AI and without. On tasks that fell within AI’s strengths, the results improved dramatically — work advanced a quarter faster and quality rose markedly. But when the task fell outside that invisible boundary the researchers called the jagged frontier, those who used AI produced the correct answer markedly less often than those with no machine at all. So the same tool both helped and harmed, depending on which side of the boundary you were on. And the whole crux is this: recognising where the boundary lies is the human’s work. It is Boyd’s orientation — the judgement of when the machine can be trusted and when it can’t. That is exactly what the machine leaves to us.

The levels begin to melt

When these three constraints break down at once, the thing I started with happens. At Procter & Gamble in 2025, an experiment had nearly eight hundred professionals solving real product-development tasks. The most surprising result: one person with AI reached roughly the same level as a two-person team without AI. And that’s not all — without AI, the technical people proposed technical solutions and the commercial people commercial ones, but with AI both produced balanced proposals that crossed their own silo. The fences between functions began to dissolve.

Here we return to the observation of the series’ first part. Robert Anthony, regarded as the father of the management pyramid, did not originally speak of layers but of kinds of work, and said that the same person does all of them. The pyramid arose when those kinds of work had to be divided among different people, because one person didn’t have the time or capacity for everything. AI dismantles exactly that necessity. It restores the situation in which one person can again stand on all three levels at once — asking why, planning how, and doing what within the same sitting. The levels don’t disappear. They stop being different people.

In practice this is already visible in the way a small, AI-equipped unit can do what used to need a whole department. A two-person team can now operate like ten, because the collation, preparation, and routine work of that ten-strong crew is handled by the machine. And this doesn’t only concern start-ups. The same logic applies to any team that gets the tools and the permission to use them: it no longer has to send every question up the staircase and wait for an answer; it can grasp the whole itself. The pyramid assumed that the few at the top see far and the many at the bottom see near. AI gives the one at the bottom a telescope too.

But — the same force can also harden the layers

Here it’s only fair to stop. The same technology that can melt the levels can also harden them. If AI gives leadership a real-time view of everything happening at the edge, it can tempt them to concentrate power even more at the top, not less — the same pipe that carries information up cheaply carries control down cheaply too. Algorithmic surveillance is already everyday reality in many industries. And there’s another possibility: perhaps the levels don’t melt, but the base simply disappears — the operational layer is amputated and the strategic and tactical are left on top. Then an awkward question arises: if the beginner’s work vanishes, where do the experienced people come from, whose tacit knowledge is born from exactly those years at the base? On top of that, the models and the server capacity are owned by a few: the three largest cloud providers already control about two-thirds of the enterprise cloud-infrastructure market. The individual’s freedom to operate on all levels takes place on rented capacity, whose terms are set by someone else.

So I’m not arguing that the pyramid is automatically dead. I’m arguing that the assumption it was built on is breaking — and that the outcome depends on what we choose to do with the breach.

Because if information is cheap, the decision cycle fast, and the person on the spot suddenly sees the same whole picture as the central management, an inevitable question arises: where does decision-making power actually belong? In the previous part I left Hayek’s crack open — the fact that the most important knowledge has always been below, at the edge. Now the crack is turning into a fissure. And out of the fissure, power drains. Where it drains to is the subject of the next part.

Lenni Laukkanen helps leadership teams turn AI into a competitive advantage through people. Invite me to speak at your event or to spar with your leadership team — I reply within a working day.

Next part: Authority Follows the Value — and Finds It at the Edge: where power drains to, once information no longer has to climb.

Sources

Durable theory

John Boyd, Patterns of Conflict / A Discourse on Winning and Losing (briefings 1976–1995). Background: Frans Osinga, Science, Strategy and War: The Strategic Theory of John Boyd (2007); Robert Coram, Boyd (2002).
Garicano, Luis. “Hierarchies and the Organization of Knowledge in Production.” Journal of Political Economy, 2000. (Layering as a response to the cost of moving information.)

Recent evidence

Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey. “Generative AI at Work.” The Quarterly Journal of Economics 140(2), 2025, pp. 889–942 (NBER WP 31161, 2023). Peer-reviewed: +15% on average, most of all for the least experienced.
Dell’Acqua, Fabrizio et al. “Navigating the Jagged Technological Frontier.” Harvard Business School WP 24-013 (2023) → Organization Science, 2025. Peer-reviewed.
Dell’Acqua, Fabrizio et al. “The Cybernetic Teammate.” HBS WP 25-043 / NBER WP 33641, 2025. Working paper; P&G field experiment (n≈776): individual + AI ≈ team level, breaks silos.
Brynjolfsson, Erik; Chandar, Bharat; Chen, Ruyu. “Canaries in the Coal Mine?” Stanford Digital Economy Lab, 2025. Working paper; decline in young workers’ employment in exposed occupations.
Kauhanen, Antti & Rouvinen, Petri. “AI has not impacted the youth labor market in Finland.” Etla Working Papers 135 / ETLA Muistio 173, 27 January 2026. Replication of the Stanford design on Finnish data.
Klarna/OpenAI (press release 2/2024) and Klarna’s later updates in 2025 (change in the profile of the work).
Synergy Research Group, Q3 2025 (cloud-infrastructure concentration). Algorithmic management: Kellogg, Valentine & Christin, “Algorithms at Work”, Academy of Management Annals, 2020.