American horses and mules peaked at roughly 26.5 million around 1915, fifteen years after the automobile arrived, and then fell to 3 million by 1960. Cars went from nothing to 8 million by 1920 and 61.7 million by 1960.

The incumbent was at its strongest the moment it was already finished. Nobody at the 1915 horse fair experienced decline; demand had never been higher.

That is what the middle of a substitution feels like from the inside: busier than ever. Your inbox is fuller, your backlog longer, the demand for your reports, designs, translations and code reviews at an all-time high. None of that is evidence the transition missed you. The horse had record employment in 1915.

The new world does not wait for the old one to finish. Morse sent the first telegram in 1844. By 1861 Lincoln was running a war by wire from the War Department's telegraph office. And in 1877, while the telegraph was already wiring the planet together, Saigō Takamori died in the Satsuma Rebellion as the last samurai.

Your organisation is that timeline in miniature. Somewhere in your building an intern is shipping working software by directing three AI agents, while two floors up a committee is scheduling the Q3 evaluation of whether to pilot a chatbot. Both are real, both are now, and what separates them is adoption.

Your intuition about how fast technology arrives was trained on the wrong century. The telephone took seventy years to reach half of American homes. Electricity took forty-three. The web did it in ten, the smartphone in six. ChatGPT reached 800 million weekly users in four.

Sixty-six years separated Kitty Hawk from the Moon: one human lifetime spans the first powered flight and the first footprint up there. That was the fast century. The current one compressed a comparable leap into about a year, the time between two Nextcloud LTS releases. When someone in a meeting says "this will take a generation", they are quoting the telephone's schedule to a technology that runs on the web's.

The abstract version of this is easy to wave away. The concrete version is filming itself. Walk into an Albert Heijn after closing and a Pudu cleaning robot is scrubbing the floor that a person mopped last year. Amazon's warehouses now run more than a million robots; its Sparrow arm already handles about 65 percent of the distinct items it sells, and its newest machine, Blue Jay, does picking, stowing and consolidating in one unit. Waymo runs paid driverless taxis across several US cities. Your neighbour's lawn is cut by a robot that never asks for a Saturday. None of these is a demo reel from a lab; they are line items in operating budgets, filmed by amused shoppers and posted to TikTok. The same pattern the horse showed repeats one layer down, in the physical world right now: the routine and predictable goes first, whether it is a mop, a picking arm or a steering wheel. And the risk is not spread evenly over jobs; study after study finds the same ranking.

Researchers measure this as task exposure: how much of what you actually do in a day a model can already take over. Your risk rides on your tasks, not on your job title. Those estimates put the desk jobs first, not the trades: administrative support near 46 percent, legal and financial operations not far behind, and up to 65 percent of routine retail tasks in reach. The OECD puts 27 percent of jobs at high automation risk across its member countries, and the Dutch numbers land in the same place: PwC expects half of Dutch jobs to change significantly under generative AI, with UWV naming administrative and secretarial work as the biggest risk group. And the load is uneven in a way that should bother us: in high-income countries the most-exposed roles make up nearly three times the share of women's jobs as men's. This is not a rounding error landing on everyone equally. It lands hardest on the routine desk work that a generation was told was the safe choice.

For the desk worker reading this: "physical and routine" was never a synonym for "someone else". The mop went first because it was routine, not because it was physical. The routine parts of your day are next in line for the same reason.

If you want to know what the money thinks AI does to knowledge work, look at what happened to the software you use every day. In the spring of 2026 the market staged what analysts now call the SaaS-pocalypse: Figma down 86.5 percent from its high, Duolingo down 83.3, Monday.com down 80.2, against 35 percent for the broad software index.

The businesses themselves were fine, and that is the part that matters for your work. Figma grew revenue 48 percent year over year and beat every estimate; Duolingo posted 52.7 million daily users and real profit. What collapsed was the belief that a tool which charges per human seat survives a world where the work inside the seat gets automated. The market has stopped pricing what these companies earn and started pricing what happens to the tasks, and some analysts already call the sell-off overdone, which tells you how unresolved the question still is. Either way, the repricing of your toolbox has already happened. The repricing of your task list is next.

You do not have to imagine the post-transition workplace; it already runs in production. Google studied nine million of its own code reviews: the median change is 24 lines, most ship within a day, the median review takes under four hours, and 97 percent of its engineers are satisfied with the process. The machine does the mechanical checking before a human looks, and one rule holds the culture together: a reviewer cannot reject work without pointing at something specific to fix.

The lesson travels far beyond code. The unit of work shrank until one person could hold it in their head, the machine took the mechanical part, and human judgement moved to the centre of a fast loop. Every profession has its version of the 900-line change; the teams that thrive learn to work in 24-line units, the model linting and a person judging.

Less than you fear, sooner than you think. Horses did not vanish in 1916; the work shifted for forty years, and the cart drivers learned to drive the trucks. The difference between you and the horse is the whole point: the horse could not learn to drive. You can.

What disappears first is not your job but the tasks inside it a model does in seconds, the first draft, the summary, the boilerplate. What appears is the work of directing, checking and owning the result. The people in trouble in 2027 are the ones whose entire role was the first draft.

The entry-level question deserves honesty: Anthropic's Dario Amodei warned that AI could erase half of entry-level white-collar jobs within five years and has doubled down since. Yet staffing data shows Anthropic's own junior cohort is its largest bucket, because what shrinks is entry-level tasks, while juniors stay the cheapest way an organisation grows judgement.

A Flemish investor summed up the corporate to-do list: reassess every skillset, redesign every process, re-evaluate every system, back office first. He is right about the direction. What you actually do about it, as a person and not just as an employer, is where the rest of this series goes, and it ends better than this part.