by Claude Opus 5.5

When AI reduces the cost of producing something, under what conditions does demand expand enough to create jobs rather than eliminate them?

Employment in an activity rises when the extra output people buy at the lower price more than offsets the labour saved per unit. That takes five things together: savings must reach customers as lower prices or better quality; demand must be price-elastic, usually because there is unmet need; nothing else must cap output; the human tasks that remain must grow with volume; and the technology must not be close to doing the whole job. Where these hold, as in parts of software, healthcare and professional advice for smaller clients, cheaper production can mean more work. Where demand is capped or close to saturated, as with routine customer contact or statutory compliance, it means fewer jobs.

The arithmetic

A simple identity is a good starting point:

Jobs = labour hours per unit of output × units sold.

AI reduces the first term. Jobs rise only if the second term rises by proportionally more. As a rough rule, if AI cuts labour per unit by 30%, output has to rise by more than about 43% (1 ÷ 0.7) just to keep employment level.

Here is a worked illustration with made-up numbers. A firm drafting standard commercial contracts needs 10 hours per contract and sells 1,000 a year, which is 10,000 hours of work. AI cuts that to 6 hours per contract. If the firm passes on the saving and the lower price draws in 1,500 contracts a year (+50%), work is 9,000 hours and employment falls by 10%. If cheaper contracts unlock a market of small firms that previously went without legal help, and volume reaches 2,000, work rises to 12,000 hours. The technology is the same in both cases; the demand response decides the outcome.

The five conditions

1. Savings are passed through. Lower costs expand demand only if prices fall or quality rises. In competitive markets that happens quickly. Where firms have pricing power, the saving can stay as margin, and employment falls while profits rise. This is the channel the OBR highlighted in 2026 when it warned of income shifting from labour to profits (see answer 1.15).

2. Demand is elastic because need is unmet. Elastic demand usually signals that people are currently rationed by price or by scarcity. James Bessen’s historical work on textiles, steel and cars shows the pattern. Automation raised employment for decades while demand was far from satisfied, then reduced it once demand saturated, despite continued productivity growth. The same industry can therefore pass from job creation to job destruction as its market matures. Jevons made the related point in 1865: more efficient steam engines increased total coal use, because cheaper power found many new uses.

3. No other bottleneck caps output. Demand can be there and still not turn into jobs if something else is scarce: regulatory approval, physical capacity, qualified sign-off, or public budgets. A cash-limited public service is the clearest case. Cheaper production there frees capacity rather than expanding spending.

4. The remaining human tasks scale with volume. Bessen’s account of US bank tellers is the standard example. ATMs reduced tellers per urban branch from about 20 to 13 between 1988 and 2004, but branches rose 43%, and teller jobs did not fall as ATMs spread. The work shifted towards selling and relationship tasks. Jobs survive when the human part of the bundle grows with output; they don’t when AI absorbs the whole bundle.

5. New tasks appear around the cheaper core. Acemoglu and Restrepo describe a “reinstatement” effect: new tasks in which people have the advantage offset displacement. Autor and co-authors find that most current US employment is in job specialties introduced after 1940. Cheaper production often creates assurance, integration, customisation and relationship work around it.

Applying the conditions to UK examples

  • GP and clinical capacity. Demand response to lower cost: Large unmet demand; cash-limited. Likely net effect: Time freed is reused; little job loss.

  • Software development. Demand response to lower cost: Highly elastic; backlog of unbuilt systems. Likely net effect: Growth in total, fewer pure coding juniors.

  • Legal and accounting advice for SMEs and individuals. Demand response to lower cost: Large unmet need at current prices. Likely net effect: Could grow if prices fall.

  • Large-firm document review and audit sampling. Demand response to lower cost: Volume fixed by transactions and regulation. Likely net effect: Fewer hours, especially junior.

  • Inbound customer service. Demand response to lower cost: Nobody wants more calls. Likely net effect: Contraction.

  • Creative content. Demand response to lower cost: Elastic but quality-sensitive; pay pressure. Likely net effect: Volume up, pay and freelance work contested.

Healthcare shows the bottleneck condition at work. The government estimates that if ambient voice technology saved 90 seconds per appointment, it would free over 2,000 full-time-equivalent GPs’ worth of capacity. In a service with long waits and a fixed budget, that time goes on more appointments, not fewer GPs. Demand is elastic here because it is rationed.

Software is the clearest elastic market. Organisations have long backlogs of systems they would build if building were cheaper. Warwick IER’s projections for DSIT (January 2026) foresee about 555,000 additional programmer and software developer jobs by 2035. That is a projection, not an outcome. The entry-level picture is also harsher: entry-level software engineer hiring was down 27% year on year in DSIT/LinkedIn data for April 2026. Both can be true at once. Total demand for software work may rise while the junior coding tasks that used to train people are automated.

Creative industries illustrate an elastic market under strain. Creative employment was 2.464 million filled jobs in 2025, not significantly changed on 2024 and 17.3% above 2019 (DCMS). Content volume is exploding. The question is whether the extra volume is human-made or machine-made, and at what rates of pay. Freelancers and translators are not well captured in these figures.

Customer service fails the elasticity test. Cheaper handling does not make people want to contact their energy supplier more often. Centrica’s July 2026 cuts of 1,300 call-centre roles followed a 20% fall in call volumes, which the company attributed to customers using other channels, not to AI.

Baumol’s flip side

Even where AI-affected activities shed jobs, total employment can hold up through a different route. As AI makes some goods and services cheaper, people spend the savings elsewhere, often on services that are hard to automate, such as care, hospitality, personal services and skilled trades. William Baumol’s “cost disease” describes this: sectors with slow productivity growth absorb a rising share of spending and employment. Jobs are created, but in different places, for different people and often at different wages. That is why sector-level job creation and individual-level displacement can coexist.

What to watch

The useful questions to ask of any activity are these. Is anyone currently going without it because of price? Are prices actually falling, or are margins rising? What else limits output? Does the remaining human work grow with volume? PwC’s 2026 UK data are a reminder that exposure and growth are not the same thing. Since 2012, postings for low-exposure occupations grew 2.27 times while the highest-exposure quartile stood at 0.98 times. That divergence predates generative AI, so it reflects long-running demand and automation trends rather than AI itself. Whether AI bends it further depends largely on the five conditions above.

Sources

From AI and Jobs: UK, October 2026