Chapter 4
Making Something
The Utopia of the Agentic Enterprise, Part I. The Jobs the Agents Are Here to Take
In a competitive economy, nobody keeps a premium for long. A margin attracts competitors, and they bid it away. Economists call that the market working, usually from a tenured position, or at a conference in an airport hotel, where think tanks and special interest groups pay for the tiny egg salad sandwiches. Applied to AI, it means that wherever value moves, it gets thinner as soon as the next round of automation reaches it. Vendors like this version, as every layer automation reaches is the next product to sell, and it’s true as far as it goes. But it’s a story about markets. It says nothing about what an organisation does while that happens, and inside an organisation the answer is rarely “less”.
Consultancies sell reports, and each of them can now produce a competent one in an afternoon. Their clients are spending less: in the first half of 2026 UniCredit cut its spending on external consultants by 24% and Société Générale by 9%, and SAP says planned changes to its systems will cut the cost of its external consultants by up to 50%.1 The reports aren’t any shorter. Part of what the cheap input saved never reaches any judgment layer. The market reprices the whole workflow downwards, and that part of the value simply disappears. Meanwhile the clients are using the same tools, so they stop paying for preparation and pay only for what their own model can’t do.
The rest needn’t reach the checkers either. The migration metaphor implies conservation, value flowing from one place to another. Some of it evaporates. If you want to know which happened, follow one workflow’s money for a year: what the client paid before and after, and what the checkers are paid now. If that last number hasn’t moved, the migration probably only happened in the slide deck.
And cheap output raises the volume. Vendor proposals now arrive by the dozen, all AI-assisted and hard to tell apart. So attention becomes the binding constraint. Finding the few outputs that deserve checking is now a job of its own, and whoever does it is the one catching the slop grenades.
Nobody Decided
A loop can close by decision. Someone works out what it would cost to catch the automated output’s errors and decides the error rate is acceptable. The loop closes that afternoon, and nothing about the model has changed, although the next quarterly update will report a higher automation rate. So the organisation didn’t replace the human verifiers with automated ones. It lowered its definition of “verified”.
Or it closes by threshold. The system takes the routine cases and leaves people the exceptions. Then the line for what counts as an exception moves, because the automated resolution rate is acceptable, and the next quarter it moves again. Nobody decides to close the loop. It closes because nobody decides to keep it open. Each adjustment makes sense in isolation, and may come as a harmless question in a steering meeting, “can we handle a bit more volume automatically?”
A risk committee would catch an explicit decision to remove human oversight, and would probably want a form about it. But it reviews the decision to deploy a system, not the quarterly recalibration of the system’s autonomy boundary. Quality management calls this process drift, and in manufacturing a control chart catches it. AI verification has no control chart, only a dashboard of the savings, so the drift can run for quarters, and it has a direction. The thresholds move towards whichever function reports the savings, while the cost falls on people who aren’t party to the adjustment: the customer, who loses the channel, and the junior, who loses the ladder.
Jevons, Then Parkinson
Every layer of routine work that automation commoditises gives back some time and attention. So what does an organisation do with it?
In 1865 William Stanley Jevons observed that more efficient steam engines didn’t reduce coal consumption, because they made steam power economical for uses that had never had it.2 Cheaper code means more features and more prototypes, and anyone who has looked after a software backlog knows it doesn’t get shorter when the team gets faster. Where demand has no natural ceiling, as in software, the freed week fills up with the next layer of production that was previously too expensive to attempt.
That is the free market theory. Inside a mature organisation, another law applies. In 1955 the naval historian C. Northcote Parkinson wrote a satirical essay in The Economist about the office that ran the Royal Navy.3 Work, he observed, expands to fill the time available for its completion. An elderly lady with nothing else to do can spend a whole day sending a postcard to her niece in Bognor Regis: an hour to find the postcard, another to find her glasses, and 20 minutes deciding whether to take an umbrella to the postbox in the next street. The Navy had the same talent, at scale. In 1914 it had 62 capital ships and 2,000 officials and clerks at the Admiralty. By 1928 it had 20 ships and 3,569 officials, so the administration had grown by nearly 80% while the fleet it administered lost two thirds of its ships. Parkinson’s explanation was that an official wants to multiply subordinates rather than rivals, and that officials make work for each other. Graeber found the same curve in the universities 70 years on, and when his essay came out in 2013, the same magazine ran a rebuttal within two days, arguing that the jobs were what a complex global economy needs. Which is roughly what the Navy’s officials would have said.
Where demand is bounded by regulation or physical constraint, as in safety certification, the freed capacity can turn into labour savings, so this is where the business case might actually come true. It is also where the organisation can respond with a new rule, because when a bounded function gets cheaper, its head needs a reason to keep the headcount. This is especially true if it’s not meant to generate revenue.
None of this undoes the repricing. Translation has already been marked down: in a 2024 survey by Britain’s Society of Authors, more than four in ten translators said generative AI had cut their income.4 Each layer of checking will follow once somebody can sell a tool for it. But the market only sets the price. What the freed time is spent on gets decided inside the organisation, and a hierarchy left to itself spends it the way the Navy did: fewer ships and more officials.
Checklist
- Has anyone lowered what “verified” means without saying so?
- Who reviews the quarterly recalibration of the system’s autonomy boundary, if anyone?
- Who keeps the savings when the loop tightens, and who carries the cost?
- Where will the freed time go: more demand, or more administration? Unless someone decides, it goes to more production.