The shape of the organisation after AI

The expectation is that AI makes organisations smaller. Fewer people, flatter structure, less middle. That may yet happen. It has not happened so far, and the case for acting as though it has is weaker than the confidence with which it is usually made.

The headcount case is running ahead of the evidence#

Three measurements, none of which supports cutting on the strength of AI today.

• The macro gains are modest. Acemoglu’s task-based model, applying Hulten’s theorem to existing estimates of task exposure and task-level cost savings, puts total factor productivity gains at no more than 0.66 per cent over ten years, revised to under 0.53 per cent once hard-to-learn tasks are accounted for. That is an order of magnitude below the headline value estimates in circulation.
• The labour effects are not showing up. Humlum and Vestergaard, linking adoption surveys to administrative records for roughly 25,000 workers across 7,000 Danish workplaces in eleven exposed occupations, found precise null effects on earnings and hours two years after ChatGPT, ruling out effects larger than two per cent.
• Displacement is concentrated, not general. Brynjolfsson, Chandar and Chen, in ADP payroll microdata covering millions of US workers, explicitly rule out widespread economy-wide displacement, while finding employment among 22 to 25 year olds in highly exposed occupations about 19 per cent below counterfactual, through reduced hiring.

Each carries its own limits. Acemoglu’s model would not capture gains running through new products or new tasks. Denmark is high-trust, high-wage and heavily unionised, and two years is early. Payroll data cannot establish causation. But an organisation cutting headcount today on the strength of AI is acting in advance of every measurement available, and should at least know that it is doing so.

What actually changed: the work, not the headcount#

The Danish study is the interesting one, because the null on pay sits alongside substantial task reorganisation and new tasks in AI oversight and integration. The structure of the work moved considerably while the numbers a board watches did not move at all.

That should change what an organisation measures. If the visible indicators are headcount and cost, an organisation will conclude nothing is happening for the two to three years during which the thing that matters is happening. Pay is the slowest available indicator, and most AI reporting is built on indicators slower still.

Redesigning a job, and the question that actually decides it#

Autor and Thompson give the sharpest lens available. Across four decades of task data covering 303 US occupations, automation that removed the less expert tasks raised wages and reduced employment, while automation that removed the expert tasks lowered wages and increased employment. Which tasks you remove decides whether a role appreciates or commoditises, and the two paths run in opposite directions on both pay and employment.

So job redesign is not primarily an exercise in removing effort. It is a choice about which half of a role to keep, made role by role, with a consequence that shows up in the labour market rather than in the process map. Their data ends in 2018, so this is a question to ask rather than a forecast to rely on.

Bainbridge’s Ironies of Automation supplies the second half. Automating the routine parts of a task leaves the human with the hardest residue, monitoring and exception handling, while removing the routine practice that built the competence to do it. A redesigned job is usually a harder job with a shorter runway for learning it, which is not how redesign is normally sold internally.

Smaller but more fragile?#

This is the most interesting version of the question and the least evidenced, so the sourcing needs care.

The mechanism has been named. Rohde describes capability masking followed by capability erosion: AI output creates a persuasive appearance that organisational capability has been replaced, while dependence on skilled human labour remains, supporting hiring restraint and deferred structural reform while costs accumulate. That is a sole-authored conceptual synthesis, nineteen pages, a preprint, with no new empirical data, and the author says so in the paper. Treat it as a hypothesis somebody has articulated clearly rather than as a finding.

What gives the hypothesis weight is arriving from elsewhere. Dauth and colleagues, in German administrative data from 1994 to 2014, found incumbents kept their jobs and moved into higher-quality tasks while the cost fell on young entrants who left vocational training altogether. An organisation that thins its intake looks identical to one that has not, for years, and then does not.

The counter-case deserves equal billing. Lee, Iizuka and Eggleston, using regional robot subsidies as an instrument across Japanese nursing homes, found adoption raised employment, improved retention, moved effort towards direct care and improved quality on hard measures, with less use of physical restraint and fewer pressure ulcers. The condition was an acute labour shortage, so the robots substituted for vacancies rather than for people. Fragility is a consequence of choices about which tasks to remove, rather than a property of adopting the technology.

Middle management#

Almost everything written on this is assertion, and this research holds no direct measurement of what AI does to management layers. What can be said comes from reasoning about the function rather than from data about the role.

Middle management does at least three separable things: routing information upward and downward, allocating and sequencing work, and developing people. The first is the most automatable and the most often cited when the layer is declared finished. The third is the least automatable and, on the entry-level evidence above, becoming more valuable at the moment the pipeline thins. An organisation that removes the layer because the first function got cheap will discover it also removed the third. Whether that is happening at scale is unmeasured, and anybody who tells you otherwise is extrapolating from anecdote.

Small organisations#

The structural advantage is real: fewer approval layers, less legacy process, and a founder who can make the augmentation-or-replacement call directly rather than through a committee. See who should own AI strategy for why that call is the one that matters.

The specific risk is different from the one large organisations face. A small organisation has no bench. If three people hold all the judgement and two of them let it decay because the tool is handling the work, there is no depth behind them and no formal process that would surface it. Large organisations lose capability slowly and visibly. Small ones lose it suddenly, when somebody leaves.

What good adoption looks like a year in#

Not high usage. The measurable things worth checking after twelve months, in rough order of how much they tell you:

• Somebody can name which decisions are now made differently, and who is accountable for each.
• There is a written position on augmentation or replacement, by domain, that matches what the business cases actually count as the benefit.
• Somebody has checked whether people can still do the core work unaided, and the check has a date and an owner.
• Juniors are getting deliberate repetitions that the job no longer supplies by accident.
• At least one deployment has been stopped or constrained, which is evidence that the thresholds set at the start were real.

Adoption rates, licences issued and hours saved tell you what was bought rather than what changed. An organisation that can only report those has measured its procurement.

The three-year exercise#

The exercise is simple to describe and slow to do. Take the work your organisation does and sort it, on a three-year horizon, into three piles. What is farmed out to the machine entirely. What is augmented, meaning a person stays in it and the machine changes how they do it. What stays human-retained, meaning a person does it and the reason is written down.

The value is in being honest about each pile rather than in the sorting itself. Most teams pad the third pile with work they have simply never examined, and pad the first with work they would be horrified to actually hand over. Both are comfortable. The test for the third pile is whether you can say what is lost if a machine does it, in a sentence somebody outside the room would accept. The test for the first is whether you would be content to lose the ability to do it at all, because that is what handing it over entirely means in five years.

Then redesign the roles from those three piles, and not from the org chart. This is the step that gets skipped. A team sorts its work thoughtfully and then maps the result onto the job titles it already has, at which point the exercise has told it nothing it can act on. Roles are the unit that has to change, because a role is what a person understands themselves to be doing, and drift is what happens when the work moves and the role does not.

The alternative end state should be said plainly. An organisation that does not do this arrives, in three years, at one of two places. Either the work has been quietly hollowed out, so the output holds up and the people inside can no longer tell you what their job is or check what the machine produced. Or somebody in finance does the sorting instead, in a spreadsheet, under a different name, and calls it headcount. Neither is a prediction. They are what is left when nobody decides.

Where this sits in my own argument#

My argument is that the organisational risk is not headcount, it is capability debt: a cost taken on now and paid later, invisible on every measure a board currently watches. The Danish null on pay alongside substantial task reorganisation is the cleanest illustration of that I know. The balance sheet says nothing happened. The work says otherwise.

This is why I argue the decision belongs upstream, in who owns AI strategy, and why the missing measurement on middle management bothers me more than the confident claims about it. The layer that develops people is the one the missing rungs argument says is becoming scarcer.

What I have observed in organisations#

The change I see most clearly is in the shape itself. Organisations that were pyramids are becoming inverted triangles, or diamonds with very little arriving at the bottom. The junior work is being done by the tools, so the junior people are not being hired.

In the short term that reads as efficiency, and on any current measure it is. The problem is arriving later and somewhere else. An organisation with no junior intake has no bench, and in ten years it has nobody to promote. Leadership development and succession both assume a supply that is quietly being switched off, which is the missing rungs argument seen from inside the org chart rather than from the labour market.

Smaller organisations are not ruled out here#

It does not claim organisations will not get smaller. It claims the measured evidence does not yet support it, that the strongest single macro estimate is modest, and that acting ahead of measurement is a decision rather than a deduction.

It does not claim the fragility argument is established. The clearest statement of the mechanism is a preprint with no empirical content, which is said here rather than glossed, and the supporting evidence is drawn from a different technology and a different era.

And it offers nothing measured on middle management, because nothing measured exists in this corpus. That absence is the most striking thing found while writing this page: the layer everybody says is disappearing is the one nobody has counted.