---
title: "Chapter 4: A Motor Where the Steam Engine Stood"
book: "Collapse or Collapse"
author: "Vinay Pasricha"
url: https://vinaypasricha.com/books/collapse-or-collapse/chapter-4-a-motor-where-the-steam-engine-stood
status: "On Kindle: https://www.amazon.com/dp/B0HNL53J5R. The full text is shared on vinaypasricha.com with the author's permission."
---

# Chapter 4: A Motor Where the Steam Engine Stood

In 1895 an electrical engineer from Columbia University, Professor F. B. Crocker, went to Taftville in Connecticut to see the newest thing in the world at work.

The Ponemah mill at Taftville had run its looms from engines in its basement. Now a hydroelectric station had been built at a dam upriver, at Baltic, and the power came down to the mill along overhead lines, which was the new thing, and it was remarkable: power made in one place and used in another, miles away, with nothing between the two but a wire on poles. Crocker followed the wire to where it arrived. The motors, he reported, were “located in the basement of the mill, near the engines which they replace.” They were belted to pulleys, and the pulleys were clutched to the mill’s shafts. One motor drove twelve hundred looms. The other drove five hundred.[^20]

To see what that meant you have to see a mill of the time, and the historian Warren Devine has drawn one. A single engine stood at the centre. It turned iron or steel line shafts, about three inches thick, which hung from the ceiling and ran the whole length of each floor; belts ran up through holes in the ceilings to carry the power to the floors above; the line shafts turned shorter countershafts, and every machine in the building was belted to a countershaft and stood, of necessity, in a row parallel to it. The whole network turned from the moment the engine was started in the morning until it was stopped at night, no matter how many machines were in use. If a shaft broke, a room stopped. If the engine broke, the mill stopped.

At Taftville the engine had been taken out of that picture and a motor put where it had stood. Everything else was left as it was: the shafts, the belts, the rows, the building shaped around them. The Ponemah mill had the power of the twentieth century and the floor of the nineteenth. Nobody there was foolish. The shafts were paid for and they worked. Devine’s judgement on the whole period is one sentence, and it is the subject of this chapter: “Replacing a steam engine with one or more electric motors, leaving the power distribution system unchanged, appears to have been the usual juxtaposition of a new technology upon the framework of an old one.”[^21]

I have stood in that basement more often in the last three years than I can count. It is not in Connecticut, and the machine in it is not a motor.

## The basement today

The first three chapters of this book argued that the department was a workaround for one constraint, that the constraint has gone, and that the outcomes of the old company leak at every joint between its desks. Almost every owner I have said this to agrees with some of it, and then tells me what his company is doing about AI. What he describes, nearly always, is a motor where the steam engine stood. It comes in four forms.

**A copilot at every desk.** Take the recruitment firm of chapter 2, with its seven desks between a client’s need and an engineer who joins. Give every desk an assistant. The account manager’s assistant turns the client’s call into a written brief in a minute. The sourcer’s assistant searches faster and more widely than she ever could. The screener’s assistant reads a thousand CVs an hour. The coordinator’s finds interview slots, and the offer desk’s drafts the offer. Every desk is faster, and every desk can prove it, with a number for its own step. And the client still waits three weeks, because the three weeks were never at the desks. They were between them, in the queue at each desk, in the brief that lost a little of the client at every pass, in the coordinator’s wait for the interviewer’s diary. Michael Hammer quoted an insurer, in 1990, which had estimated that an application spent twenty-two days in process and was actually worked on for seventeen minutes.[^22] A copilot at every desk turns the seventeen minutes into eight. It leaves the twenty-two days where they were.

**A chatbot in front of the same desks.** Take the bank of chapter 2. A conversational agent now takes the small business’s application at midnight, in the owner’s own language, checks that every document is attached, and tells him in a second that his application has been received. Then the application goes to the relationship manager, the credit analyst, the risk committee, legal and operations, as it did before, six desks and six weeks, and the order the owner needed the money for is lost as it was before. The bank has rebuilt its front door and left the house as it stood.

**A programme run by the department being transformed.** Take the insurer of chapter 2, with its five desks between a damaged car and a paid claim. It appoints a head of AI transformation, who reports to the head of operations, and an office to support her. Each department is asked for its use cases. The registration desk pilots an agent that reads the claim form. The assessment desk pilots one that reads the photographs. The survey desk pilots one that drafts the survey report. Each pilot is chosen by the department that will run it, built inside that department’s systems, and judged by the number that department already reports, and a steering committee of the department heads meets once a month to review them. The pilots succeed, every one of them, because a pilot is designed to succeed: it is chosen so that it can be done inside one department, without asking another department’s permission, and so it cannot touch the joints between departments, which is where the thirty days were lost. The minutes of the steering committee are now taken by an agent. The meeting is still held.

**An agent where the people were.** The fourth form is the boldest and fails in the most instructive way: an agent put into a department in place of the people who worked there, and measured by the number the people were measured by. Chapter 19 tells it at full length, from the payments company that did it in 2024, as I read the public record, and I will not tell it twice. What matters here is only its shape: the department stayed standing around the agent, and no human stood at a gate owning what the customer felt.

## Why the joints eat the gain

The four forms fail for the same reasons, and the reasons are not in the technology. The agents work. They work at the desk, and the desk was never the problem.

The first reason is that the joints stay. Every handoff that existed before the copilot exists after it, and so does every approval, and so does every meeting called to carry an outcome across a joint that nobody owns. Chapter 2 counted how much of the working week goes on communicating about the work rather than doing it, and chapter 3 showed what leaks at each joint. A copilot does not remove a joint; it makes the work on either side of it quicker, and the joint, which was a wait, is still a wait.

The second reason is that each department optimises its own piece. Give a department an agent and it will use the agent to improve the number it is measured by, which is the only rational thing for it to do. The screening desk screens three times as many CVs an hour, so three times as many screened CVs arrive at the coordinator, whose queue was already the slowest in the chain, and the client waits a little longer than before for a little more paper. Every desk is now better at its piece. The pieces still have to be carried from one desk to the next by people whose job is carrying, and their work has increased.

The third reason is that nobody owns the outcome. The agents in a paved company are bought by functions, configured by functions and judged by functions, and the outcome the customer feels belongs to none of them, exactly as it belonged to none of the desks before. Chapter 3 put it plainly: cut an outcome into seven pieces and give each to a specialist, and you have told each specialist that the outcome is not his. The paved company has given each of the seven a faster tool and told him the same thing.

The fourth reason is the one the evidence now shows most plainly: the pilots succeed and the company does not change. When McKinsey surveyed nearly fifteen hundred organisations in the summer of 2024, most respondents whose business units used generative AI reported cost reductions inside those units; and more than eighty per cent said their organisations were seeing no tangible effect on enterprise-level profit. Of the twenty-five attributes the survey tested, the one with the largest effect on whether a company saw any profit at all from generative AI was whether it had redesigned its workflows, and only twenty-one per cent of the companies using it had fundamentally redesigned even some.[^23] When a team of economists including Nicholas Bloom asked nearly six thousand senior executives in America, Britain, Germany and Australia, sixty-nine per cent said their firms actively used AI, and nine in ten said it had had no effect on their employment or their productivity over the previous three years.[^24] And when two economists linked surveys of twenty-five thousand workers in seven thousand Danish workplaces, in eleven of the occupations most exposed to AI chatbots, to the country’s records of what those workers were paid and how many hours they worked, they found that the users reported saving about three per cent of their working hours, that most of them spent the saved time on other tasks in the same job, and that two years after the launch of ChatGPT the effect on earnings and recorded hours was zero, measured precisely enough to rule out anything above two per cent.[^25]

These are three different kinds of evidence: a consultancy’s survey of its own respondents, executives’ reports of their own firms, and a national register. Each has its limits, and the Danish authors draw their own conclusion, that the change shows first in how work is reorganised and only later in pay and hours, which I do not dispute. But the three point the same way, and I read them as one finding. The time a desk saves stays at the desk. The gain is real, and it is spent where it was made, and the company, which is the sum of its joints, does not feel it.

## The computer age everywhere

None of this is new. It has happened once already, in living memory, and it was named at the time.

In July 1987 Robert Solow, who would win the Nobel prize in economics that October, reviewed a book about American manufacturing for the *New York Times*. The authors believed that a revolution in programmable automation was under way, and Solow noted, a little drily, that what everyone felt to have been a technological revolution had been accompanied everywhere, including Japan, by a slowing-down of productivity growth. Then he wrote the sentence economists have been quoting ever since: “You can see the computer age everywhere but in the productivity statistics.”[^26] A year later Naval and I were riding a scooter from factory gate to factory gate, selling the computer age to men who had not seen it either.

Three years after Solow, Michael Hammer said why. Chapter 2 told what became of the reengineering movement he and James Champy led, and why it failed: the skills still lived in seven heads, and the seven heads re-formed into seven desks. But his diagnosis came first, in an article in the *Harvard Business Review* in the summer of 1990, and his diagnosis did not fail. Heavy investment in computers had delivered disappointing results, he wrote, “largely because companies tend to use technology to mechanize old ways of doing business. They leave the existing processes intact and use computers simply to speed them up.” And then the line that named this chapter’s subject for a generation of managers: “It is time to stop paving the cow paths.”[^27]

His best example is worth telling, because it is the copilot, described and rejected three decades before the word was used. In the early 1980s Ford looked at its accounts payable department in North America, which employed more than five hundred people, and concluded that by rationalising its processes and installing new computer systems it could reduce the headcount by about a fifth. Then it looked at Mazda, whose accounts payable department had five people. The department spent most of its time on mismatches, the cases where the purchase order, the receiving document and the supplier’s invoice disagreed. Hammer put the choice in one sentence: “One way to improve things might have been to help the accounts payable clerk investigate more efficiently, but a better choice was to prevent the mismatches in the first place.” Ford stopped asking suppliers for invoices, matched the order against the receipt in one database, and paid on the match. Where the new process was installed, the headcount fell by three-quarters, not the fifth that the better-equipped clerk would have delivered.

A fifth against three-quarters. That is the difference between paving a cow path and building a road, and Ford found it with the computers of the 1980s. The reorder book had taught me the same lesson a year before Hammer’s article, without my knowing it had a name. The easy sale in 1988 would have been a computer on Ramesh’s desk, so that he could copy the registers faster. We would have been paid for it, and the machine on the assembly floor would still have stood there waiting for its part, because the mistake was never in Ramesh’s copying. It was in two pages that nobody put together.

## The dynamo

The fullest account of why a new machine can be everywhere and invisible was written by the economic historian Paul David, in 1990, and his subject was the motor in the basement at Taftville.

In 1899, David showed, electric motors provided less than five per cent of the mechanical drive in American factories. It took about two decades for the figure to reach half. And factory electrification had no effect on the growth of productivity in manufacturing before the early 1920s, which was four decades after the first central power station opened for business. In 1900, he remarked, contemporary observers might well have said that the electric dynamos were to be seen “everywhere but in the productivity statistics!”

The delay was not in the motor. The motor worked from the beginning. The delay was in the factory. From the mid-1890s to the eve of the 1920s, David wrote, the usual way to electrify an existing plant was to add electric motors to the equipment it already had, turning the old shafting in sections, with the original belts, shafts and even engines left in place. The effect was that the factory carried two systems where it needed one, so the capital it employed for every unit it produced went up, and the productivity it could show went nowhere.[^28]

The gain came when the factory was rebuilt around the new machine. When every machine was given its own motor, a system called unit drive, the building changed. With no heavy shafting to carry overhead, the structure could be lighter. With no long shafts to turn, it no longer had to be several storeys high, and the single-storey factory became possible. Machinery, Devine wrote, “could now be arranged on the factory floor according to the natural sequence of manufacturing operations,” rather than in rows beside a shaft; and, David noted, the whole power system no longer had to be shut down to make a change in one department. And then the numbers moved. In American manufacturing, output per man-hour had grown by 1.3 per cent a year before 1919. After 1919 it grew by 3.1 per cent. David found that about half of the acceleration in the growth of manufacturing productivity in the 1920s, against the decade before, is accounted for statistically by the spread of electric motors through the factories in that decade.[^29]

The line shaft is the organisation chart: the fixed structure through which all power is distributed, which every machine must be lined up beside, and which has to turn, all of it, from morning to night, however little of the work is moving. A department holds its weekly review whether one file is waiting or forty. The mechanical engineer H. C. Spaulding named the fault of the motor-driven line shaft in 1891: the necessity, as Devine reports it, of turning all the shafting in the plant regardless of the number of machines in operation.[^30] Unit drive is the vector, a chain of agents with its own power for each outcome, laid out in the order of the customer’s journey and not in rows beside a function. And the single-storey factory is the company of chapter 2, two layers tall, which will not grow a third, because there is no shafting overhead for a third layer to carry.

The same is true of the software the departments run on. It was built for people who read screens, fill forms and wait in approval inboxes, and an agent set to work on it is a motor belted to the old shaft. The vector’s agents write to the record directly; the screens were the shafting.

David added a warning that I have not been able to stop thinking about. The old factories were rebuilt slowly because it did not pay to replace a building that still worked; the industries that grew fastest built new electrified plants first, and the rest waited for their buildings to wear out. An owner today might hope to wait the same way. He cannot, David wrote in 1990, because a firm’s information structures do not wear out on their own: “one cannot depend on the mere passage of time to create occasions to radically redesign a firm’s information structures and operating modes.”[^31] The departments will not fall down. Somebody has to decide to take them down.

## The dip

Thirty years after David, Erik Brynjolfsson, Daniel Rock and Chad Syverson gave his history a shape, and the shape has a name. A general purpose technology, they argued, such as electricity, the computer or AI, needs large complementary investments, the co-invention of new processes, products, business models and skills, and most of those investments are intangible and poorly measured. So in the early years a company pays for the reorganisation, its measured productivity falls or stalls because the reorganisation shows up nowhere as output, and later, when the reorganisation is done and its benefits are harvested, measured productivity rises faster than the real gain. Drawn on a chart, it is a J. Adjusting only for the intangible investment tied to computer hardware and software, they estimated that American productivity at the end of 2017 was 15.9 per cent higher than the official figures showed.[^32]

I will put their finding in the language of this book, and the translation is mine, not theirs. The bottom of the J is the reorganisation. It is the period in which the company is spending on the new way of working and still running the old one, and nothing yet shows for it.

That explains the paving company exactly. It never makes the complementary investment, because a copilot at every desk is designed to need none: the desks, the handoffs, the approvals and the meetings stay as they were. So it never climbs the J. It stays at the bottom of it, with the motor and the shafts both on its books, paying for the agents and for the departments, which is the most expensive company there is.

The collapse takes the other path. It does the reorganisation on purpose, at once, and whole. Chapter 13 shows it at Pravaha, the invented pump maker whose year is the method’s worked example: every department sorted by one test, the agents built and run beside the departments for two weeks, the whole revenue chain collapsed on one Monday, every department in it. By the end of month one, in that chapter, the win rate on quotes has gone from eighteen per cent to twenty-three, and the plant, which has not changed, ships two hundred and five pumps, because the order book had been the thing standing still. Pravaha is a composite and its figures are a design, not a measurement. But the design is built on the J. There is a dip, and it is month zero, in which the agents run beside the departments and the company pays for both. It is short because it is taken in one go, and the doctrine sets a limit on the whole collapse: twelve months, or a quarter more if the machines are late. Longer is a reorganisation, not a collapse.

There is one honest objection to all this, and it is my own. In 1989 Mr Malhotra made us run the program beside the registers for six months, and I have called it the most important decision of the whole project. David noted that the offices of his own day were doing the same with the computer, keeping old paper procedures alongside the new methods, sometimes to the detriment of both.[^33] The difference is the date. Running both for six months, with a day on which the registers stop, is a test. Running both with no such day is a motor in the basement, and the shafts still turning overhead.

## The cow-path test

So here is a test, and it takes a minute. Describe any AI initiative in your company in one sentence, by saying what it makes faster. If your answer begins with the name of a department, you are paving.

“The claims desk now reads documents with an agent.” Paving. “HR screens CVs with AI.” Paving. “Finance closes the books in three days instead of ten.” Paving, however good the three days are, because the sentence is about a department’s own piece. Now say it the other way, beginning with what the customer feels. “The car is repaired and the owner paid in the seven days he was promised, and one chain holds the registration, the assessment, the survey, the approval and the payment.” That is not a paved path. It is a vector, and its sentence does not mention a department because there is none in it.

**The discharge.** “The pharmacy prepares discharge medicines in ten minutes with an agent” begins with the pharmacy, and the patient who was well at nine still leaves at six, because the file, the account and the car still wait on each other. “The patient who is well at nine is home by lunch” begins with her, and it can be true only if one chain holds the order, the file, the medicines, the account and the car, and one human stands at the gate where she is let go.

**The licence.** “The ministry has put a chatbot on its website” begins with the ministry, and the citizen still stands in the queue you know. “The citizen has the permission the day she applies, or a named official’s reason why not” begins with her, and no ministry has yet said it, which is why chapter 21 expects a small state to be the first.

Three questions follow, and an owner can ask them of any project on his desk. How many handoffs will the outcome cross after this goes in, and how many did it cross before? Who signs for the outcome the customer feels, by name? And which desk will not exist a year from now? If the count of handoffs has not changed, if the signature belongs to a function, and if the answer to the third question is none, you have bought a motor for the basement. It may be a very good motor. The mill still has the floor it had.

## What stays human

The paving company keeps a human at every desk, because it has kept every desk, and it asks each of them to do, a little faster, work the machine could already hold. The collapsed company keeps humans too, fewer and for a different reason, and the reason is Tenet IV of chapter 10: machines hold the baseline, and humans supply the peaks, which are intention, urgency and judgement. It is worth saying here, briefly, where the collapsed company spends them, because the rest of the book is built on it.

It spends them at the **signature**: a commitment is made by a named human at a gate and owned by her until the next one, which is accountability, and the subject of chapter 17. It spends them in the **Keeper’s** seat, deciding what the company exists to produce, which Outcomes serve it and which should be born or allowed to die, which is direction, and chapter 8. It spends them with the **customer**, who is promised in chapter 2 that when something must be signed, a human will sign it, and he will know her name, which is the whole of trust. And it spends them at the **connected boundary** of chapter 6, the edge between the work that reaches a chain in real time and the work that does not yet: the hands on the shop floor and in the surgery, and the know-how nobody has written down. Dhawan’s afternoons in Verma’s corner, taking a plant out of his head one assembly at a time, were that edge moving. Everything else a human did at a desk, the machine now holds, and a company that keeps the desk to do it has kept the shafting.

## The words for the new floor

Devine tells the history of the factory in three stages: line shaft drive, group drive, unit drive. The first two describe the old floor with a new machine in it. Only the third describes a new floor, and it could not be told in the words of the first two, because those words were about where the engine stood, and the new floor was laid out by how the work flowed.

A motor in the basement needs no new words. A new floor does. Part Two gives them.

[^20]: F. B. Crocker, V. M. Benedikt and A. F. Ormsbee, “Electric Power in Factories and Mills,” *Transactions of the American Institute of Electrical Engineers* 12 (June 1895), 413–14, as quoted in Warren D. Devine Jr., “From Shafts to Wires: Historical Perspective on Electrification,” *Journal of Economic History* 43, no. 2 (June 1983): 347–372, at 356: the Ponemah mill at Taftville, Connecticut, powered from a hydroelectric plant at Baltic; “motors are located in the basement of the mill, near the engines which they replace. They are belted to pulleys which are connected to their respective shafts by friction clutches … one of these motors drives 1,200 looms requiring an expenditure of about 155 horsepower. The other drives 500 looms.”

[^21]: Devine, “From Shafts to Wires,” as above: on line shaft drive, the shafts “usually 3 inches in diameter,” suspended from the ceiling and running the length of each floor, machinery “arranged, of necessity, in rows parallel to the line shafts,” and the network rotating “from the time the steam engine was started up in the morning until it was shut down at night—no matter how many machines were actually being used” (pp. 350–52); and, on the first electric drive, “Replacing a steam engine with one or more electric motors, leaving the power distribution system unchanged, appears to have been the usual juxtaposition of a new technology upon the framework of an old one” (p. 357).

[^22]: Michael Hammer, “Reengineering Work: Don’t Automate, Obliterate,” *Harvard Business Review* 68, no. 4 (July–August 1990): 104–112. In his account of Mutual Benefit Life’s old application process, Hammer adds: “Another insurer estimated that while an application spent 22 days in process, it was actually worked on for just 17 minutes.”

[^23]: McKinsey & Company (QuantumBlack), “The state of AI: How organizations are rewiring to capture value,” McKinsey Global Survey, 12 March 2025; online survey of 1,491 participants, 16–31 July 2024. For most business functions using generative AI, a majority of respondents reported cost reductions within the units using it; “more than 80 percent of respondents say their organizations aren’t seeing a tangible impact on enterprise-level EBIT from their use of gen AI”; “out of 25 attributes tested for organizations of all sizes, the redesign of workflows has the biggest effect on an organization’s ability to see EBIT impact from its use of gen AI”; 21 per cent of respondents reporting generative-AI use said their organisations had fundamentally redesigned at least some workflows. The figures are respondents’ own reports.

[^24]: Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom and others, “Firm Data on AI,” NBER Working Paper 34836, February 2026: nearly 6,000 senior executives at firms in the US, UK, Germany and Australia; 69 per cent of firms actively use AI; “nine-in-ten reporting no impact on employment or productivity” over the previous three years; the same executives expect AI to raise their firms’ productivity by 1.4 per cent over the next three.

[^25]: Anders Humlum and Emilie Vestergaard, “Large Language Models, Small Labor Market Effects,” NBER Working Paper 33777, May 2025, and later revisions: surveys of 25,000 workers in 7,000 Danish workplaces across eleven highly exposed occupations, linked to administrative records; adopters report saving about 3 per cent of their work hours, and most users report reallocating the time saved to other job tasks; precise null effects on earnings and recorded hours at both the worker and the workplace level, ruling out effects larger than 2 per cent two years after the launch of ChatGPT. The authors find the change instead in the reorganisation of tasks.

[^26]: Robert M. Solow, “We’d Better Watch Out,” review of Stephen S. Cohen and John Zysman, *Manufacturing Matters: The Myth of the Post-Industrial Economy*, *New York Times Book Review*, 12 July 1987, p. 36.

[^27]: Hammer, “Reengineering Work,” as above. The cow-path passage is on p. 104: “It is time to stop paving the cow paths. Instead of embedding outdated processes in silicon and software, we should obliterate them and start over.” The Ford case: accounts payable in North America “employed more than 500 people”; management expected to “reduce the head count by some 20%”; Mazda’s accounts payable “consisted of a total of 5 people”; “where it has instituted this new process, Ford has achieved a 75% reduction in head count, not the 20% it would have gotten with a conventional program.”

[^28]: Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” *American Economic Review* 80, no. 2 (Papers and Proceedings, May 1990): 355–361: electric motors under 5 per cent of factory mechanical drive in 1899, about two decades to reach 50 per cent, and the dynamos seen “everywhere but in the productivity statistics!” (p. 356); no impact on manufacturing productivity growth before the early 1920s, “four decades after the first central power station opened for business,” and the retrofitting of existing plants under group drive, with “the original belt and shaft equipment (and the primary engines that powered them)” remaining in place, raising the capital-output ratio (p. 357). David attributes Solow’s remark in the form “We see the computers everywhere but in the productivity statistics” (p. 355); the wording in the text is Solow’s own.

[^29]: Devine, “From Shafts to Wires,” as above: output per man-hour in manufacturing grew at an average 1.3 per cent a year before 1919 and 3.1 per cent after (p. 348); under unit drive, “Machinery could now be arranged on the factory floor according to the natural sequence of manufacturing operations, minimizing handling of material” (p. 364). David, “The Dynamo and the Computer,” as above: lighter, single-storey factories, and “the entire power system no longer had to be shut down in order to make changes in one department or section of the mill” (pp. 357–58); “approximately half of the 5 percentage point acceleration recorded in the aggregate TFP growth rate of the U.S. manufacturing sector during 1919–29 (compared with 1909–19) is accounted for statistically simply by the growth in manufacturing secondary electric motor capacity” (p. 359).

[^30]: Devine, “From Shafts to Wires,” as above, p. 357, reporting the mechanical engineer H. C. Spaulding on “the necessity of turning all the shafting in the plant regardless of the number of machines in operation,” and citing Spaulding, “Electric Power Distribution,” *Power* 11 (December 1891), 12.

[^31]: David, “The Dynamo and the Computer,” as above, p. 360. On the delay in rebuilding plant, “the unprofitability of replacing still serviceable manufacturing plants,” and the fastest-growing industries building new electrified plants first, p. 357.

[^32]: Erik Brynjolfsson, Daniel Rock and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” *American Economic Journal: Macroeconomics* 13, no. 1 (January 2021): 333–372, doi 10.1257/mac.20180386: general purpose technologies such as AI require large complementary investments, “including co-invention of new processes, products, business models and human capital,” which are intangible and poorly measured, so that productivity growth is underestimated in a new technology’s early years and overestimated later; adjusting for intangibles related to computer hardware and software, the authors estimate total factor productivity 15.9 per cent above official measures at the end of 2017 (11.3 per cent at the end of 2004). The reading of the dip as the reorganisation, in the text, is the author’s.

[^33]: David, “The Dynamo and the Computer,” as above, p. 357: in the computer’s data-processing and recording applications, “old paper-based procedures are being retained alongside the new, microelectronic-based methods—sometimes to the detriment of each system’s performance.”