The night the calls changed direction
At midnight on 14 July 1930, a manufacturing city in New England changed the direction of its telephone calls.
Until then, a caller lifted the receiver and reached an operator. The operator asked for a number, selected a cord, found a jack, made the connection and watched the switchboard until the parties finished. After midnight, callers used a dial. Electromechanical switches performed much of the routing inside the exchange.
The city is unnamed in Ethel Best's 1933 Women's Bureau investigation. That anonymity keeps us from inventing a local legend around the event. The surviving records supply enough of the human scale. The company had decided on conversion in 1927 and spent three years preparing. Its operator payroll stood at 534 in June 1930. By January 1931, six months after the cutover, 249 remained.
Planning softened some of the blow. The company limited hiring, transferred 131 operators and initially retained 260. Of the 116 laid off around the change, all but four had been temporary or occasional employees. Those employment categories made the reduction administratively orderly. They did not make it painless.
Best's researchers later followed 78 former operators. Of the 77 who reported how much work they had lost, 43 had lost at least three months during the following year. Twenty-one had lost at least ten months. Only 21 of the 78 had worked throughout the year, and 18 of those were still temporary telephone operators.
One network had changed in a night. The work had been changing for years before it, and the consequences kept travelling afterwards.
This is the pattern concealed by the phrase “technology changes jobs”. A technical capability appears. An organisation chooses whether and how to adopt it. Tasks move among people, machines, customers and suppliers. Managers assemble the remaining and new tasks into jobs. Similar jobs change across workplaces until an occupation's content or demand begins to move. Employment, earnings and opportunity emerge at the far end of that chain.
Every step contains a decision. Every step can distribute the gain differently.
Four different things that headlines call a job
The first discipline is to use the right unit.
A task is an activity within work: connecting a call, checking a balance, taking a history, cutting a panel, writing a test or resolving a complaint. Tasks can be short or extended, physical or cognitive, routine or uncertain. They are pieces of action rather than people or positions.
A job gathers tasks and duties around one person. The International Labour Organization's occupational classification defines it in almost exactly those terms: a set of tasks and duties performed, or meant to be performed, by one person. A job also has a work relationship, schedule, authority, equipment, colleagues, risks and rewards. Change the task bundle or those surrounding conditions and the job changes, even if the title on the payslip does not.
An occupation groups jobs whose main tasks and duties are highly similar. The Australian Bureau of Statistics uses that principle in the Occupation Standard Classification for Australia. A registered nurse in a public hospital and one in a private hospital can belong to the same occupation because the main tasks are similar, despite different employers and settings.
Employment asks another question: how many people or hours are engaged, in which places and work relationships? An occupation can retain hundreds of thousands of workers while changing its daily work. Its task content can remain recognisable while employment falls. Output can grow while fewer hours are required. A new task can expand demand for people already carrying an old occupational title.
Official occupational systems preserve these distinctions because no single field can contain the whole reality. The U.S. Department of Labor's O*NET Content Model keeps occupation-specific tasks apart from general work activities, work context, worker characteristics, worker requirements and labour-market information. The separation is useful well beyond statistics. It stops a claim about one activity from silently becoming a claim about an entire livelihood.
When software writes a first draft, that is evidence about a task. Whether an editor's job shrinks, broadens or divides depends on who reviews the draft, what quality is required, whether more work is commissioned, how responsibility is assigned and what the organisation does with the time saved. Whether the occupation grows depends on the same choices playing out across many workplaces, alongside changes in demand and entry.
The sequence begins smaller than the job title.
Babbage and the divided table
Charles Babbage understood this before an electronic computer existed.
In 1832, while writing about machinery and manufacturing, he devoted a chapter to the division of mental labour. His example was the production of mathematical tables in post-revolutionary France. The calculations could be decomposed. A small number of highly skilled mathematicians chose formulas and methods. Other people organised the sequence. Many more carried out repeated arithmetic operations.
Babbage's point was economic as well as computational. When a complicated act is divided, different operations can be assigned to people with different training and cost. The same decomposition also makes part of the process more susceptible to a machine. His later designs for calculating engines grew from that insight.
He did not leave us a modern theory of labour markets. He supplied an enduring way to look. An occupation that appears indivisible from a distance often contains operations with different demands for judgement, dexterity, memory, communication and responsibility. A technology may reach some of them early, others weakly, and some only after the surrounding system is rebuilt.
David Autor, Frank Levy and Richard Murnane used this task-level lens to study computerisation in the late twentieth-century United States. Their 2003 analysis treated computers as especially effective at cognitive and manual tasks that could be expressed in rules. Computers could also complement people doing non-routine problem-solving and complex communication. In data from 1960 to 1998, task content changed within industries, education groups and named occupations as well as through movement among them.
The labels “routine” and “non-routine” in that model describe a relationship between a task and the available technology. They are not measures of human worth. Nor are they permanent properties. A task that could not be specified to a machine in 1980 may become tractable after better sensors, data, interfaces or models. A task that can be performed in a demonstration may remain unreliable, costly or unacceptable inside real work.
Technology changes the feasible set. Work changes when an institution chooses a new allocation.
The chain between invention and employment
A machine, model or process can change the cost, speed, scale or quality of an activity. It can make a previously impossible action feasible. That is a technical event. The labour-market event requires a longer chain:
- Capability: What can the technology do, under which conditions and at what cost or quality?
- Adoption: Which organisations acquire it, integrate it and trust it for real work?
- Task redistribution: Which activities move to the technology, remain with people, become shared, or move to customers and suppliers?
- Job redesign: How are the remaining and new tasks bundled, supervised, scheduled and rewarded?
- Occupational change: Do similar job changes spread far enough to alter an occupation's content, entry route or demand?
- Employment and earnings: What happens to hours, headcount, pay, mobility and access for particular workers and places?
The demand for an output runs through the whole chain. Lower production costs may reduce the labour required for each unit while lower prices or better quality expand the number of units people buy. A new service may attract demand that did not exist before. A firm may keep the same output and take the saving as profit. A public institution may use it to serve more people under the same budget, or to reduce staffing.
This is why a laboratory result, product launch or venture-capital total cannot settle an employment question. Capability does not prove adoption. Adoption does not reveal job design. A changed job does not tell us how many similar jobs exist. An occupational percentage does not show who gains, who waits and who leaves.
Technology · Adoption · Work
How does a technology travel into an occupation?
A capability reaches employment only through adoption, task allocation, job design, organisational demand and the distribution of gains and losses.
Dial exchangeA technical cutover moved connection workShow detailsHide details
Automatic switching transferred routine call connection from operators to equipment. Installation, maintenance, traffic management and customer-service work remained, while the timing and place of worker losses depended on adoption.
ATMSelf-service changed the branch bundleShow detailsHide details
Cash withdrawal moved partly to customers and machines. Branch staff handled a different mix of service and sales tasks, while branch numbers, staffing, hours and earnings followed demand and organisational choices rather than the machine alone.
Generative-AI draftingDrafting can move before authority doesShow detailsHide details
A model may produce a first draft while people retain framing, checking, approval and responsibility. Adoption, output growth and task redesign must be observed before any claim about jobs or employment can follow.
| Case | Task before | Task after | Complementary or new work | Still requires evidence |
|---|---|---|---|---|
| Dial exchange | Operator connects calls. | Switching equipment connects routine calls. | Install, maintain, monitor and serve exceptions. | Local adoption date, worker exit, later hiring, hours and earnings. |
| ATM | Teller handles routine cash withdrawal. | Customer and machine complete many withdrawals. | Machine servicing, exception handling and a changed branch service mix. | Branch demand, staffing model, headcount, hours and earnings. |
| AI drafting | Person frames, drafts, checks and approves. | Tool may generate a draft; people may retain the other stages. | Prompting, evidence retrieval, verification, exception work and governance. | Actual adoption, output response, job redesign and worker outcomes. |
Three examples—the dial exchange, ATM and generative-AI drafting—move through six stages: technical capability, organisational adoption, task redistribution, job redesign, occupational change and employment outcomes. The map keeps output, job content, headcount, hours and earnings separate and shows that each transition needs additional evidence.
The allocation stage contains at least four recurring routes. They often operate together.
A task can leave, grow, appear or move
Substitution: a machine takes an existing task
The dial switch took over much of the act of connecting local calls. A spreadsheet can recalculate a column that a clerk once totalled by hand. A robot can weld a repeatable seam.
Daron Acemoglu and Pascual Restrepo call this automation's displacement effect: capital performs tasks previously allocated to labour, reducing labour demand for those tasks. The word task matters. A worker may have many other duties. The organisation may expand output. Yet the displaced slice is real, and workers concentrated in it face the most immediate pressure.
Complementarity and scale: the remaining human task becomes more productive
Technology can make another part of the work more valuable. A diagnostic tool may let a clinician inspect more cases while increasing the importance of communicating uncertainty and choosing treatment. Design software can accelerate revision while leaving responsibility for constraints and trade-offs with the designer.
David Autor's 2015 synthesis explains why substitution and complementarity coexist. If a process requires several inputs, making one stage cheaper can increase the value of the stages that remain scarce. Productivity can lower prices or improve quality, expanding demand for the output. The expansion may support more work even though each unit requires less labour.
There is no promise of compensation inside the mechanism. Demand may respond weakly. The gains may accrue to owners or consumers. The complementary work may require different people or fewer people with greater responsibility.
New tasks: the production system creates work it did not previously need
New technologies bring installation, maintenance, monitoring, training, integration, safety, auditing and product work. They can also create outputs that were previously too expensive or impossible.
Acemoglu and Restrepo describe a reinstatement effect when new tasks arise in which labour has a comparative advantage. This can restore labour's place in production after automation has removed old tasks. It is a mechanism to investigate, not a law that every lost job returns in another form.
The identity of the new worker matters. A former operator might be able to enter the new work with support, or the work may appear in another city, firm, occupation or generation. Aggregate balance can hide personal rupture.
Transfer: the task moves across a boundary
Self-service often moves work to the customer. The bank customer enters a personal identification number, chooses a transaction and handles cash at the machine. Online shoppers search, compare, order and track. Airline passengers check in and label bags.
The activity has not vanished. Part of it has changed hands, usually without becoming paid work for the customer. Other tasks move to suppliers through contracting, or across borders through remote delivery and production. A firm can report fewer internal jobs while the task survives elsewhere.
The work relationship changes what this transfer means. It determines who directs the activity, owns the equipment, carries the risk, controls quality and receives income. A technology story that follows only the employer's headcount can miss work shifted to contractors, households and users.
These four routes turn an invention into a new allocation. The next question is how the pieces are assembled.
The organisation hidden inside the machine
Technology rarely arrives alone. It needs workflow, authority, training, data, maintenance, incentives and a reason to be used.
Timothy Bresnahan, Erik Brynjolfsson and Lorin Hitt found that information technology and organisational change often travelled together in U.S. firms. Greater use of computing was associated with changes such as broader line responsibilities, more decentralised decisions and altered demand for skill. The OECD and ILO reach a related conclusion in their review of skills use at work: what people are able to use depends on job design, management practice, work organisation and the way technology is introduced.
This means two workplaces can buy the same system and produce different jobs.
One may use it to script each step, measure pace and move exceptions to a supervisor. The frontline job becomes narrower and more observable. Another may automate record retrieval and give the worker more authority to diagnose the customer's problem. The job becomes broader. Both organisations can claim to have “augmented” workers. The lived allocation of judgement and control is different.
Training belongs here, after the design question. Teaching a person to operate a system can help them enter or retain work. It cannot repair a job designed with no route for learning, no discretion and no time to use the training. Nor can it create demand for a role the organisation has removed.
Worker consultation matters for the same reason. OECD surveys on AI and work associate training and consultation with better reported worker outcomes. That evidence supports involving people who know the work. It does not establish that consultation alone prevents displacement or determines how gains are shared.
A job is therefore a designed bundle as well as a list of technical necessities. The person or institution with the power to bundle it helps decide whether a tool becomes surveillance, assistance, self-service, craftsmanship or redundancy.
Why the occupational title blurs the change
Occupations are necessary abstractions. They let statistical agencies count, compare and describe millions of jobs. The abstraction also conceals variation.
Two people with the same title may work in different industries, use different systems, serve different customers and spend different shares of the day on each task. Melanie Arntz, Terry Gregory and Ulrich Zierahn used worker-level task data from 21 OECD countries to revisit occupation-based estimates of automation risk. Their 2016 study found a smaller highly automatable share when it accounted for the different tasks people actually performed within the same occupation.
The point is methodological. A technical judgement about an occupation's task list cannot be applied uniformly to every job carrying the title. Even a precise task-level estimate is still an estimate of technical exposure. It does not show that an employer will adopt the technology, that the whole task can be performed at required quality, that demand will remain fixed, or that a worker will lose employment.
Occupational continuity can hide deep change too. A title can survive while tools, responsibilities and entry paths shift. A stable occupation count can coexist with widespread redesign. A falling count can coexist with more output and more complex remaining work.
This is why the library's work vocabulary matters. The unit must match the claim. “This task is exposed” can be defensible. “This occupation will disappear” requires evidence about adoption, every major task, job rebundling, demand, time and place that an exposure score does not contain.
The ATM story, with its ending restored
The automatic teller machine has become a parable. It is often told to prove that automation creates jobs: banks installed ATMs, the cost of operating branches fell, branches multiplied, and tellers moved towards customer service.
There is truth in the middle of that account. Federal Reserve evidence shows that U.S. bank branches increased during the 1990s while ATM deployment also rose rapidly. Branch numbers continued rising until 2008. Digital and branch services were imperfect substitutes: people used machines for routine transactions and still visited branches for advice, problems and services.
The teller job contained more than dispensing cash. The Bureau of Labor Statistics still describes tellers as processing transactions, answering questions, explaining products and referring customers to other staff. Machines took some tasks, customers took some, and the branch retained others. Lower transaction costs and continuing demand for branches supported coexistence for a long period.
The ending matters. U.S. branch counts plateaued after 2008 and began a modest decline after 2013. The BLS counted about 347,400 teller jobs in 2024 and projected a 13 per cent decline through 2034. That is a forecast, not a fate, but it prevents the historical episode becoming a guarantee of permanent protection.
ATMs show that task substitution can coincide with job rebundling and expanding demand. They also show that the balance changes as online banking, mobile services, consumer habits and branch strategies change. A technology's first decades do not settle an occupation for all time. One occupation's history does not settle another's.
Aggregate adjustment is not personal repair
Return to the telephone operators.
James Feigenbaum and Daniel Gross reconstructed the spread of mechanical switching across more than half the U.S. telephone network from 1920 to 1940. Their 2024 study found two histories inside the same transition.
For later cohorts of young women, overall employment rates did not fall. They entered clerical, service and newly appearing work instead of telephone operating. The economy adjusted partly because people approaching the labour market chose among a changed set of entry jobs.
Incumbent operators had a harder path. After local cutovers they were more likely to leave employment or move into lower-paying work, with effects still visible years later. The precise estimates belong to young, white, American-born women in a period marked by discrimination and a narrow set of acceptable jobs. The broader lesson needs no universal number: future entrants can route around a disappearing entry point while the people already attached to it bear a lasting loss.
Place can carry the same asymmetry. Acemoglu and Restrepo studied industrial robots across U.S. commuting zones. Their local estimates associated one additional robot per thousand workers with a 0.2 percentage-point reduction in the employment-to-population ratio and 0.42 per cent lower wages in exposed areas. Those figures are bounded to a technology, country and period. They demonstrate why gains elsewhere do not automatically repair the labour market where tasks were displaced.
“The economy creates other work” can be true while evading every distributive question. Who can perform the new tasks? Where are they? What do they pay? How long does the route take? Who funds the transition? Does a temporary worker receive the same protection as a regular employee? Does a community dependent on one industry share in lower prices or higher profits elsewhere?
The 1930 telephone company used advance notice, attrition and transfers to protect many regular operators. The temporary women absorbed much of the cut. That was a work-design and employment-policy choice made around a machine, not a feature inside the dial.
What an exposure estimate can tell us
Generative AI returns us to the unit problem at larger scale.
The ILO's 2025 refined global index combined task-level occupational data with expert and model judgements. It estimated that one in four workers were in occupations with some exposure to generative AI, while 3.3 per cent were in the highest exposure category. The ILO judged transformation more likely than full replacement because most exposed occupations retain tasks requiring human input.
These are exposure findings. They identify places where the technology may reach the task bundle. They do not say that one in four workers will lose jobs, that 3.3 per cent will be automated, or that transformation will be benign.
The path still runs through capability, cost, reliability, adoption, workflow, responsibility, demand and institutions. A drafting system may reduce time on first-pass text while increasing review, verification and revision. It may let an organisation produce more, keep output fixed with fewer hours, move work to clients, or create a new service. The consequences depend on the design and on who controls the saved time.
Automation, Augmentation, Exposure, Displacement and Transformation separates those terms in detail. The boundary here is enough to continue: exposure names a possible point of contact between a technology and a task. It is the beginning of an employment inquiry, not its answer.
The missing links in a technology claim
A claim that a technology will “change an occupation” leaves nine links to be filled in.
What exactly is the task? The activity, its inputs, required quality, exceptions and attached responsibility define what the technology would have to change.
What has been demonstrated? A controlled capability is different from reliable workplace performance at a viable cost and scale.
Who adopts it, and why? Adoption requires a firm, household or public institution with an output it wants and the resources to make complementary changes.
Where does each task go? Work may move to machines, remaining workers, new specialists, customers, contractors or people in other places.
How is the job rebuilt? The new bundle may change judgement, pace, monitoring, collaboration, safety, entry requirements, learning and authority.
What happens to demand? Changes in cost, price, quality and output influence whether the organisation expands service, protects its margin or cuts capacity.
Which occupation is being counted? Jobs carrying one title can vary, and the classification may not yet capture emerging work.
Who and when? Incumbents and entrants may face different outcomes, as may regular and temporary workers, different places and different stages of adjustment.
What happens to employment and earnings? Headcount, hours, pay, security and access can move in different directions.
Following the links does not produce a universal percentage. It produces a bounded claim with a unit, mechanism, population, place and time, which is more useful than false precision.
What changes, and in what order
That technology changes the possible actions and constraints inside production.
That organisations decide which possibilities become real work arrangements.
That tasks can be substituted, complemented, created or transferred.
That managers and institutions rebundle those tasks into jobs, with choices about authority, learning and risk.
That occupations move when similar job changes spread, even when their titles survive.
That employment and earnings depend on demand, adoption, output, institutions and distribution as well as technical capability.
That adjustment for a later cohort can coexist with deep loss for an incumbent worker.
At midnight in July 1930, the dial changed who connected the call. It did not decide which operator would be transferred, which temporary worker would be released, how long she would search, or where the next generation of young women would work. People and institutions made those arrangements around the machine.
Technology enters work through tasks. Its consequences arrive through decisions.
Notes on the evidence
The telephone opening is drawn from the U.S. Women's Bureau's investigation of one unnamed New England city. The later cohort findings come from Feigenbaum and Gross and remain bounded to the population and period they studied. Babbage is used for the history of decomposing mental operations, not as a source of modern employment estimates. The task model draws on Autor, Levy and Murnane; Autor; and Acemoglu and Restrepo. The robot figures are local U.S. estimates, not universal coefficients. ATM history is paired with the current BLS outlook to prevent coexistence becoming a claim of permanent job security. The ILO generative-AI index measures technical exposure rather than adoption or job loss.
Sources and further reading
- Best, Ethel L. The Change from Manual to Dial Operation in the Telephone Industry. U.S. Women's Bureau, 1933.
- Feigenbaum, James and Daniel P. Gross. “Answering the Call of Automation: How the Labor Market Adjusted to Mechanizing Telephone Operation”. 2024.
- International Labour Organization. “International Standard Classification of Occupations: ISCO-08”.
- Australian Bureau of Statistics. “OSCA 2024 structure”. 2024.
- O*NET Resource Center. “The O*NET Content Model”.
- Babbage, Charles. “On the Division of Mental Labour”. 1832.
- Autor, David, Frank Levy and Richard Murnane. “The Skill Content of Recent Technological Change”. 2003.
- Autor, David. “Why Are There Still So Many Jobs?”. 2015.
- Acemoglu, Daron and Pascual Restrepo. “Automation and New Tasks: How Technology Displaces and Reinstates Labor”. 2019.
- Acemoglu, Daron and Pascual Restrepo. “Robots and Jobs: Evidence from US Labor Markets”. 2020.
- Arntz, Melanie, Terry Gregory and Ulrich Zierahn. The Risk of Automation for Jobs in OECD Countries. OECD, 2016.
- Bresnahan, Timothy, Erik Brynjolfsson and Lorin Hitt. “Information Technology, Workplace Organization, and the Demand for Skilled Labor”. 1999.
- OECD and International Labour Organization. Better Use of Skills in the Workplace. 2017.
- International Labour Organization. Generative AI and Jobs: A Refined Global Index of Occupational Exposure. 2025.
- Federal Reserve Bank of Chicago. “Will Electronic Money Be Adopted in the United States?”. 2001.
- Federal Reserve Board. “Why Are There Still Bank Branches?”. 2018.
- U.S. Bureau of Labor Statistics. “Tellers”. 2025.