There’s a reasonable objection that deserves to be taken seriously: every wave of automation has arrived accompanied by predictions of occupational extermination that never came true. ATMs were supposed to wipe out bank tellers, and yet for forty years, from 1970 to 2010, American teller employment never entered a lasting decline: James Bessen documented that ATMs cut the tellers a branch needed from about twenty to roughly thirteen, but made branches so cheap to open that banks opened many more of them, and overall the jobs held (the decline did come, after 2010, but at the hands of mobile banking, not the ATM). The spreadsheet was supposed to wipe out the people who did the sums, and in the United States some 400,000 bookkeeping clerk jobs have indeed disappeared since 1980, while 600,000 were created among accountants proper, because analysis had become so cheap that clients started asking for far more of it. Anyone who dismisses today’s AI anxiety as the latest Luddite panic has a long historical record on their side, and an embarrassing one for the prophets of doom. David Autor, the MIT economist who shaped the way we study these transitions, has quantified the phenomenon with his coauthors: about 60% of American employment in 2018 was in occupations that didn’t even exist as a job title in 1940. Work doesn’t end. It moves.
And yet this consoling reading, true in aggregate and over horizons of decades, turns false the moment you use it to reassure a single person about a single decade. The typist who was the fastest in the office in 1985 was not saved by the statistic that work, on the whole, would recompose itself elsewhere. That specific job evaporated all the same, and understanding why it evaporated is the most useful thing we can do today, because the mechanism is repeating itself with almost textbook fidelity.
The Typist’s Story, Told Properly
That story is worth telling properly, because the version in circulation is poorer than the original. In 1975, at Xerox PARC in Palo Alto, Larry Tesler and Tim Mott developed a text editor called Gypsy. It wasn’t meant for a mass market: it was built for Ginn & Company, a textbook publisher owned by Xerox, based in Lexington, Massachusetts. Gypsy introduced things we now take for granted to the point of no longer seeing them: selecting text by dragging the mouse and double-clicking a word, along with the commands Tesler christened cut, copy and paste. There’s a detail in its genesis I find almost moving. To design the interface, Tesler asked his new secretary, who had only ever used high-end typewriters, to sit in front of a blank screen and describe her ideal way of taking notes and composing documents. What she described became Gypsy. The profession the tool would dissolve helped design it, and neither of them, the secretary or the engineer, could see where the whole thing was going to end up.
Because that is exactly the point: Gypsy and its descendants didn’t make typists faster. They made an operation that used to be expensive nearly free, namely editing a text that had already been written. On a typewriter every correction carried a steep price, and that price justified an entire organizational architecture: the executive dictated or wrote by hand, the draft went to the typing pool, came back corrected, went out again. When editing became free, typists didn’t start working faster. What happened is that executives started writing on their own, badly and gladly, because they could correct themselves endlessly at no cost. The typing task wasn’t automated: it was pulverized and redistributed across millions of people, inside a process that hadn’t existed before. The American Bureau of Labor Statistics still counts the survivors of the category “word processors and typists”, and steadily projects it near the top of the fastest-declining occupations. Not because a machine types better than they do, but because the world stopped being organized in a way that needed a dedicated typing function.
The Bundle and the Goal-Seeker
This is where Autor’s “task approach” becomes a tool and not just an academic label. The idea, formalized in 2013 but rooted in his 2003 work with Levy and Murnane, is that technology never meets a profession: it meets individual activities, and reassigns those activities between human labor and capital according to a principle of comparative advantage. A profession is a package of tasks held together by an organization, and the package rarely disappears all at once. Sangeet Paul Choudary, picked up in Italy by Vincenzo Cosenza in the article that prompted this reflection, calls the package a “bundle” and observes that so far human beings have survived automation because they kept the goal-seeker function: technology ate the executive tasks, but someone still had to decide the goal, break it down, handle the exceptions, reassemble the result. The professional title stayed on the door while the contents of the room changed.
Agents Move the Boundary
Agents move the boundary exactly there. An assistant that speeds up every single step of a process leaves the process as it was: that’s how most companies are using AI today, and it’s also the mode that produces the most disappointing results, because it compresses task time without touching the dead time between tasks, which in any real organization is the dominant line item. An agent that starts from the goal, builds a plan, hunts down sources, produces a first version of the deliverable and submits it for review isn’t speeding up five tasks: it’s proposing a different workflow, one in which some intermediate positions lose their function and others gain one. The slogan that AI won’t replace workers, only workers who don’t know how to use it, assumes the process stays as it is, and that the race is run on how fast you execute the existing steps. The fastest typist on Word had won exactly that race. It didn’t help, because by then the track was gone.
The View from the Provinces
So much for the analysis, which I owe largely to Autor, to Choudary and to Cosenza’s synthesis. What I’d like to add comes from the particular vantage point of someone who runs the technical side of a small provincial software house, about ten people working mostly on healthcare and public administration, with the occasional foray into legaltech, and which over the past two years has reorganized the way it builds software around agentic tools. It’s not a theoretical vantage point: it’s the place where these dynamics arrive first and without shock absorbers, because a ten-person structure has no organizational redundancy to absorb a wrong workflow.
The first thing I’ve learned is that in our craft the recomposition has already happened, only we call it by technical names that hide its scale. When code becomes cheap to produce, like typing in the eighties, value migrates to the artifact that precedes it: the specification. I’ve written elsewhere about specification debt, and every passing month I’m more convinced. An agent that generates in an afternoon what once took two sprints hasn’t accelerated development: it has moved the bottleneck onto the ability to say precisely what needs to be built and with which verifiable properties. Whoever can write that specification, and can recognize when the output doesn’t honor it, is doing the work that counts. Whoever executes tasks the specification has already fully determined is doing the work the process can redistribute, and sooner or later it will.
The Seed Corn
The second thing is more uncomfortable, and it concerns the pyramid. In a small team, junior work was never just low-cost production: it was the mechanism by which seniors were formed. If agents absorb precisely the kind of task on which a junior built their judgment, the company that settles for celebrating the savings is eating its seed corn. I don’t have an elegant solution. I have a practice: we don’t take tasks away from juniors, we change their object. The junior no longer writes the CRUD, they supervise the agent that writes it, and they’re evaluated on the quality of the review, not on typing speed. It’s a bet that judgment can also be formed by examining someone else’s work, as has always happened in newsrooms and law firms. Whether the bet is a good one we’ll know in a few years, but the alternative, teams with no entry-level positions, is an answer that speaks for itself.
The Map Written into the Regulations
The third thing is the reason I’m writing this piece. Cosenza closes his article on the level of the system: who defines the goals, who coordinates people and agents, who holds the client relationship, who takes responsibility for the result. And it’s true, but put that way it sounds like a matter of individual positioning, almost of career strategy. I think it’s a structural matter, and that in Europe it already has a written address. Because while we debate which tasks AI can perform, the European legislator is compiling, article by article, the list of functions no process will ever be able to redistribute to machines, for the simple reason that the law nails them to a legal person. The AI Act demands effective human oversight for high-risk systems, and assigns precise obligations to those who provide them and those who deploy them. The Cyber Resilience Act makes the manufacturer responsible for the security of the digital product regardless of who, or what, wrote its code, while the new Product Liability Directive extends that liability to software as such, and NIS2 goes as far as requiring that the management body itself, not some delegated office, approve the security measures and answer for them. You can read all of this as bureaucratic ballast, and that’s the most common reading in our industry. Or you can read it as a map: the topography, drawn in advance and carrying the force of law, of the points in the workflow where human value is not negotiable because responsibility cannot be delegated.
What No Tool Can Sign
The typist disappeared when the process stopped needing a dedicated typing function. The right question, for anyone who works with knowledge, is not which tasks a model can perform, and not even just whether those tasks will stay organized the way they are today. It’s which function the process, however much it gets recomposed, will never be able to expel. In my industry the answer has stopped being an opinion: the function the process cannot expel is responsibility, and in Europe responsibility is written into the regulations before it’s written into the org charts. Treat it as a cost and you’re getting ready to compete on execution speed, which means training to win the typist’s race. Treat it as architecture and you’re choosing, today, the level of the system you’ll inhabit when the recomposition is done.
Tesler’s secretary described her ideal way of writing to a blank screen, and from that description came the tool that would dissolve her profession. More than half a century later we’re all describing our ideal way of working to a screen. We’d better do it knowing the description will be taken literally, and that in the end the only part of that work left standing will be the part no tool can sign in our place.
Key takeaways
In aggregate the optimists are right: ATMs didn’t shrink teller employment for forty years, and the spreadsheet created more accounting jobs than it destroyed among bookkeeping clerks. But aggregate statistics don’t save a single person in a single decade: the typist’s job evaporated all the same.
Gypsy, the editor Larry Tesler and Tim Mott built at Xerox PARC in 1975, didn’t make typing faster: it made editing nearly free. The task wasn’t automated but pulverized and redistributed across millions of people, and the dedicated function disappeared along with the organizational architecture that justified it.
Autor’s task approach and Choudary’s bundle explain the mechanism: technology never meets a profession, it meets individual tasks. So far humans survived because they remained the goal-seeker; agents, which start from the goal and propose entire workflows, move the boundary onto that very function.
In a software house the recomposition has already happened: when code becomes cheap, value migrates to the specification and to the ability to recognize when the output doesn’t honor it. And juniors shouldn’t lose their tasks but have their object changed to supervising the agent, otherwise the company celebrating the savings is eating its seed corn.
The function no process can expel is responsibility: the AI Act, the CRA, the PLD and NIS2 nail it to legal persons and management bodies. Read as a map instead of bureaucratic ballast, it is the topography of the points in the workflow where human value is not negotiable.
Questions & answers
Why isn't the historical optimism about automation enough to reassure workers today?
Because it’s true in aggregate and over horizons of decades, and turns false the moment you use it to reassure a single person about a single decade. ATMs didn’t push American teller employment into decline for forty years, and the spreadsheet created more accounting jobs than it destroyed among bookkeeping clerks, but the fastest typist in the office was not saved by the statistic that work, on the whole, would recompose itself elsewhere. That specific job evaporated all the same, and the mechanism that made it evaporate is repeating itself.
What does the story of Gypsy and the typists actually teach?
That technology doesn’t need to type better than a typist to make the typist disappear. Gypsy, built by Larry Tesler and Tim Mott at Xerox PARC in 1975, made an operation that used to be expensive nearly free: editing a text that had already been written. Once correcting cost nothing, executives started writing on their own and the organizational architecture that justified the typing pool dissolved. The task wasn’t automated: it was pulverized and redistributed across millions of people, inside a process that hadn’t existed before.
What is David Autor's task approach, and why do agents change the picture?
It’s the idea, formalized in 2013, that technology never meets a profession but individual activities, and reassigns them between human labor and capital according to comparative advantage. A profession is a bundle of tasks held together by an organization, and so far humans survived because they kept the goal-seeker function: deciding the goal, breaking it down, handling exceptions. An agent that starts from the goal and produces the deliverable isn’t speeding up five tasks: it’s proposing a different workflow, in which some intermediate positions lose their function. The boundary moves onto the very level that seemed guaranteed.
How does junior work change in a team that uses agents?
Junior work was never just low-cost production: it was the mechanism by which seniors were formed. If agents absorb precisely the tasks on which a junior built their judgment, the company celebrating the savings is eating its seed corn. One workable practice is to change the object of the tasks instead of removing them: the junior no longer writes the CRUD, they supervise the agent that writes it, and they’re evaluated on the quality of the review. It’s a bet that judgment can also be formed by examining someone else’s work, as has always happened in newsrooms and law firms.
What role do the AI Act, the CRA, the PLD and NIS2 play in this recomposition?
While we debate which tasks AI can perform, the European legislator is compiling the list of functions no process will ever redistribute to machines, because the law nails them to a legal person. The AI Act demands effective human oversight for high-risk systems, the CRA makes the manufacturer responsible for product security regardless of who wrote the code, the PLD extends product liability to software as such, and NIS2 requires the management body itself to approve security measures and answer for them. You can read all of this as bureaucratic ballast, or as a map of the points in the workflow where responsibility cannot be delegated.
The author
