by Claude Opus 5.5
Which historical analogies (computers, electrification, offshoring, industrial automation) best fit modern AI—and where do those analogies break down?
No single analogy fits, but each answers a different question well. Electrification and computers are the best guides to timing: big gains arrive only after organisations are redesigned, which takes years. Offshoring is the best guide to which tasks and workers are exposed: codifiable, remotely deliverable white-collar work. Industrial automation is the best guide to who bears the cost: incumbents in specific roles and places, even when aggregate employment holds up. All four break down on the same points. AI spreads faster, reaches non-routine cognitive work, makes errors that are probabilistic rather than mechanical, and aims at better-paid jobs than earlier waves did.
Electrification: the timing lesson
Paul David’s “The Dynamo and the Computer” (1990) explained why electricity took decades to raise factory productivity. Early adopters swapped a steam engine for one large electric motor and kept the old layout. The gains came only when “unit drive” let factories put small motors on each machine and redesign the floor around the flow of work. Andrew Bailey drew on the same history in his May 2026 Sheffield speech, setting AI alongside earlier waves of technological change from the industrial revolution onwards. Brynjolfsson, Rock and Syverson formalised the pattern as a “productivity J-curve”. Firms first invest in hard-to-measure intangibles such as new processes, skills and data, so measured productivity initially understates progress.
The analogy fits the 2026 UK picture well. ONS reports that about 35% of businesses with 10 or more employees use AI, up from about 12% in late 2023. Yet the average number of AI technologies per user has risen only from 1.4 to 1.6, and only 11% of firms with 10 or more employees have trained more than half their workforce. Access is broad and use is shallow, which is exactly the “motor bolted onto the old layout” phase.
Where it breaks down: electricity required new physical infrastructure, whereas AI arrives as software on devices people already own. Diffusion of access is therefore far faster than electrification, even if diffusion of redesign may not be.
Computers: the task and polarisation lesson
The computer era supplied the framework most economists still use. Autor, Levy and Murnane (2003) showed that computers substituted for routine, rules-based tasks and complemented non-routine problem-solving. In Britain, Goos and Manning documented the result for 1975–1999 as polarisation into “lousy and lovely jobs”: growth at the top and bottom of the pay distribution and a hollowing of the middle. The IT era also showed that management decides the payoff. Bloom, Sadun and Van Reenen found that US multinationals operating in the UK got more productivity from the same IT than other firms in the UK, because of how they organised work.
Where it breaks down: computers took over tasks that could be written as explicit rules. Generative AI acts on the tasks computers complemented: drafting, summarising, coding, analysing and explaining. The skill bias may also be reversed in places. In Brynjolfsson, Li and Raymond’s study of 5,179 customer-support agents, AI raised productivity by 14% on average, by 34% for novices, and had minimal effect on the most experienced. Computers mostly rewarded the already-skilled. Some AI tools narrow the gap instead.
Offshoring: the exposure lesson
The tasks firms offshored to India, South Africa or Poland in the 2000s were the ones that could be specified, done remotely and delivered digitally: back-office processing, first-line customer contact, and routine coding and testing. Those are almost exactly the characteristics that make a task amenable to AI. In 2026, Standard Chartered announced about 7,800 back-office roles to go by the end of the decade, and its chief executive spoke of “replacing, in some cases, lower-value human capital with the financial capital and investment capital we’re putting in” (as reported). Deloitte’s Q2 2026 CFO survey ranked AI and automation (net 47%) and outsourcing (net 33%) side by side as reasons for weaker graduate hiring, which suggests firms treat the two as substitutes.
Where it breaks down: offshoring moved tasks to other people, who earned wages and spent them. AI moves tasks to capital, so the income goes to profits and to the owners of models and data. The OBR’s July 2026 Fiscal Risks and Sustainability report flagged this explicitly, warning that AI-driven productivity could shift GDP “from (more highly taxed) labour to (lower-taxed) profits”. Offshoring was also limited by time zones, language and contract management. AI faces none of these limits.
Industrial automation: the distribution lesson
Graetz and Michaels studied robots in 17 countries from 1993 to 2007. They found robots added about 0.36 percentage points a year to labour productivity growth with no significant effect on total hours worked, but reduced the share of hours worked by low-skilled people. The most instructive historical case is the telephone switchboard. Feigenbaum and Gross found that AT&T’s mechanisation of operator work in 1920–1940 did not reduce employment for later cohorts of young women, who moved into clerical and service jobs. Incumbent operators, however, were more likely a decade later to be in lower-paid work or out of work. Further back, Robert Allen’s “Engels’ pause” describes the first half of the 19th century in Britain, when output per worker grew but real wages stagnated and the profit share rose for decades before wages caught up.
Where it breaks down: robots required capital spending, a site and physical integration, so adoption was slow and visible. AI is priced per seat or per token and can be switched on across an organisation in weeks. It also targets higher-paid, office-based and London-heavy work rather than factory floors.
Where all four analogies fail
Speed of capability change. Earlier technologies improved over decades. Frontier AI models are now released every few months; mid-2026 alone brought several major launches. Plans built on today’s capabilities date quickly.
Probabilistic errors. Motors, spreadsheets and robots fail in predictable ways. AI produces fluent output that is sometimes wrong, so verification becomes a permanent cost and a source of new work.
AI helps build its own complements. Coding assistants speed up the software integration that slowed earlier technologies, which could shorten the J-curve.
The exposed workforce. Previous waves hit routine manual and clerical work. This one reaches graduate professions, which matters for a services-heavy, London-centred economy like the UK’s.
Which analogy for which question
Why aren’t productivity statistics moving yet? Best analogy: Electrification, computers.
Which tasks go first? Best analogy: Offshoring, computers.
Who bears the cost, and for how long? Best analogy: Industrial automation, telephone operators, Engels’ pause.
Where does the income go? Best analogy: None fit well; closer to software platforms.
Bottom line
History says aggregate employment usually adjusts, but slowly, and the transition costs fall on specific incumbents and places. The lesson from the telephone operators is the most relevant for 2026. Watch what happens to the people already in exposed roles, and to the next cohort trying to enter them, not just the headline employment rate.