The three big A’s in Macro - AI, Automation and Ageing

^ Twice a week, my trusty AI assistant sends me essays on economics and demographics. Thursdays are macro-essays, and Tuesday’s are for demographics geekery (and it gets deliciously geeky indeed!). These are fast becoming my favourite reading experiences of the week. I know; #sorrynotsorry etc. The post below is my first attempt to synthesise some of this material from the earlier AI runs—on AI, productivity, ageing and demographics—with some perspective recent from blog posts by Kamil Kovar. There are two kinds of people who use AI. Those that admit it and are open about it, and those who use it but won’t admit to it. I am the former. Take it or leave it.

—

On 19 November, Anthropic began trading on Nasdaq at an IPO valuation of about $1.5 trillion, the most direct and daring public-market bet on AI to date. The offering brought in $93.5 billion in gross proceeds, of which $91.6 billion went to Anthropic. OpenAI, meanwhile, has reportedly filed confidential IPO paperwork and is aiming to list in 2027 or sooner. The Nasdaq’s 10% fall in the same week suggests investors are suffering from an acute case of AI and technology indigestion. I suspect the economic story is just getting started. Share prices can reassess how much of the gain the frontier labs will capture while employers are still figuring out how to use the technology. [1]

The labour-market question is where I would start. My hypothesis is that AI can raise the productivity of less experienced workers by making existing expertise easier to access. It can also make already productive workers faster. A good financial analyst with AI might cover more companies or test more scenarios in the same time. Both effects can operate within the same team. What matters for employment is how the firm responds. If demand is strong, the team can take on more work. If the firm wants roughly the same output, it can fill fewer vacancies.

That second possibility gives rise to a defensive labour market with insider-outsider dynamics. An employer can retain the people who know the business and ask them to do more, while cutting back on entry-level recruitment as AI takes over some of the work that used to justify those hires. The first signs would show up in vacancies, hiring and job-to-job moves. Headcounts and layoffs could look fairly stable. And if routine work is how juniors learn to judge a difficult case, cutting those jobs may also narrow the pipeline of experienced workers later. This is a hypothesis informed by what I see in the labour market. It leaves room for other firms to use AI to train juniors faster and widen access to skilled work.

Ageing makes the question more interesting. The familiar image of automation is a factory line that keeps running with fewer workers. AI offers a version of that in the white-collar economy, just as slower labour-force growth makes it more valuable to economise on human time. But the macro-perspective here can be misleadingly simple. Labour can be scarce in aggregate while particular jobs become harder to get. An economy can have too few care workers and fewer openings for graduate analysts at the same time. To understand what AI does to productivity and employment, we first need to know where and how it is being used.

Adoption is rising. How much work has changed?

About one in four US employer businesses reported using AI in the two weeks ending 15 November, up from roughly one in six in late November 2025. The US Census Bureau broadened its question a year ago from AI used in producing goods or services to AI used in any business function. A manufacturer using AI to draft recruitment ads and a software company that has rebuilt its development process around AI both count as adopters. That tells us how far use has spread across firms. How much of their production has changed is a separate question. [2]

In the Census Bureau’s more detailed supplement, 18% of firms used AI in at least one business function from November 2025 to January 2026. Weight the firms by employment and the figure rises to 32%, reflecting greater take-up at larger employers. Even among adopters, 57% used AI in three functions or fewer. The study also finds that workers sometimes use AI without a formal rollout. An analyst trying out a model to explore an unfamiliar industry is a case of bottom-up adoption. A sales team whose customer database automatically prepares call notes is using AI as part of an established process. A successful experiment can lead to a formal rollout. For occasional, self-contained tasks, individual use may be all the organisation needs. [3]

Different definitions and samples also help explain the gulf between surveys. In the October revision of Firm Data on AI, Ivan Yotzov and co-authors report that 78% of firms in a seven-country executive survey use AI. Executives averaged 2.1 hours of personal use a week. They attributed around 0.2 percentage points of annual labour-productivity growth over the previous three years to AI and expected 0.6 points a year over the next three. These are managers’ own estimates. The comparison I find most revealing is within the survey itself. Use is already widespread in that sample, while much of the expected gain still lies ahead. [4]

Some of the clearest evidence comes from workplaces where output is easy to count. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the rollout of an AI assistant to more than 5,000 customer-support agents. On average, agents resolved 15% more issues per hour, with larger gains for less experienced workers. A repeatable workflow, a queue of incoming requests and a clear output measure helped the firm both realise the gain and see it. The result shows how AI can help newer workers get up to speed. Whether firms will then need fewer new hires depends on demand and how they staff the work. In a separate survey, Salomé Baslandze and co-authors find the strongest reported productivity gains in high-skill services and finance. That tells us where executives see gains. It leaves open which workers account for them. [5]

Suppose AI cuts the time a task takes by 20%. If the task accounts for a tenth of paid hours and the tool covers only a quarter of that work, the initial saving is half a per cent of total hours, before checking and rework. A support team can use the time to clear more tickets. An analyst might write a better report, with a gain that is harder to measure. Elsewhere, a few minutes saved here and there may be difficult to redeploy. Anders Humlum and Emilie Vestergaard find new work in AI oversight and integration in Denmark, alongside little detectable effect on recorded hours or earnings two years after ChatGPT’s launch. Higher output at the same hours remains possible. Their evidence shows why time savings and a lower wage bill are two different outcomes. [6]

The productivity J-curve

Paul David’s history of electrification offers a parallel. A factory could swap a steam engine for an electric motor and keep its existing shafts, belts and layout. The larger gains came when it could power machines individually and reorganise the shop floor. Today, an insurer can give its claims team AI subscriptions or use AI to extract details from documents and then rethink how claims are routed and which ones need a specialist. The payoff depends on the whole process, including the time a specialist spends correcting mistakes upstream. [7]

Brynjolfsson, Daniel Rock and Chad Syverson describe a productivity J-curve. Early on, cleaning data, building interfaces and training staff can take time away from current production. If some of that work builds durable organisational assets and the accounts miss the investment, measured productivity can also understate what the firm creates during the build-out. Later, measured productivity can grow unusually fast as those unrecorded assets start generating visible output. The J-curve combines a real cost of reorganising work with a measurement issue around how the investment is recorded. A slow start is consistent with a J-curve. The eventual payoff still has to be demonstrated. [7]

Tania Babina, Alex He and Renhao Jiang use job postings and worker histories to track internal AI investment at US listed companies. They find a positive relationship with productivity after 2018 and link it to the organisation capital firms accumulate by learning to use AI. Some businesses may be able to buy part of that capability in the software they already use. Others will still have to redesign processes themselves. And the aggregate gain can come partly through reallocation. If more productive firms win customers, economy-wide output per hour can rise even while the typical firm changes slowly. That route is easier to imagine in a national software market than in home care, where adoption depends on local providers and commissioners. [8]

The studies above show that we are still in the explorative phase of AI adoption. Adoption is rising but its effects are uneven, and in any case difficult to measure accurately, especially when results are self-reported. Add diffusion dynamics and a J-curve and we still lack the exact framework to understand how AI will impact work and firm-productivity over time. Two frameworks, however, are likely to survive in my view. The distinction between informal (bottom-up) and formal (top-down) adoption as well as the effect on AI on low-performance workers vs high-performance workers are both likely to be key to understanding the impact of AI over time.

From the micro to the macro

The hardest part of the macro arithmetic is working out how much of the economy can realise the gains we see in individual tasks. Daron Acemoglu estimates an increase in the level of total factor productivity of up to 0.66% over a decade, using early estimates of task exposure and cost savings. An OECD study estimates a contribution of 0.4 to 1.3 percentage points to annual labour-productivity growth over ten years in highly exposed G7 economies such as the US and UK, depending on adoption. The time units differ, and labour productivity also captures the effect of more capital per worker. The divergence is a reminder that the scope and speed of diffusion can do as much work in these estimates as the results of the task-level studies. [9]

If hours worked per person fall by half a per cent a year, output per hour must grow roughly half a percentage point faster to keep GDP per person on its previous path. A shrinking population also means that a given growth rate in output per person delivers slower growth in total GDP. That distinction matters to anyone forecasting the size of domestic markets or the revenues available to service an existing stock of debt.

Acemoglu and Pascual Restrepo give us a reason to expect firms to respond. Where middle-aged workers became relatively scarcer, countries adopted more industrial robots and developed more automation technologies. Their earlier cross-country work found little evidence that faster-ageing economies grew more slowly in their sample. But is it really that easy? Kamil Kovar is sceptical of the inference that automation will leave growth largely unaffected by ageing. My reading of the growth arithmetic is that a wave of adoption can permanently lift the level of productivity while speeding up growth only during the transition. If the demographic drag persists after the wave, sustaining the offset requires further productivity improvements. [10]

There is also an inflation angle. Charles Goodhart and Manoj Pradhan argue that the reversal of a long period of abundant labour will bring faster wage growth and more inflation. I have never been persuaded that ageing is inherently inflationary. A smaller workforce can put pressure on wages in particular sectors while slower household formation and weaker demand for investment weigh on spending elsewhere. AI could ease the supply constraint where firms can substitute capital for scarce labour. In the build-out, it can also raise demand for power, equipment and the people who install it. The direction of aggregate inflation depends on how those forces play out across the economy and how monetary policy responds. Even if AI brings down costs in some sectors, the relative price of care can still rise. [11]

That last point is central. Iñaki Aldasoro, Sebastian Doerr and Daniel Rees at the BIS find more scope for AI and robots to substitute for labour in sectors such as finance than in health and agriculture, where their data show older workforces. Ageing also raises demand for care. Better scheduling can cut travel time between home visits, and AI can reduce the hours spent on documentation. Those savings will translate into more care where staff time is the binding constraint. Where the limits are beds, clinical supervision or cover for particular shifts, the effect on capacity will be smaller. [12]

The fiscal gain is also conditional. Masao Fukui, Emi Nakamura and Jón Steinsson show how making workers more substitutable can raise productivity while putting pressure on wages. If more of the gain accrues to profits, a pension system that relies on payroll contributions will capture less of it directly. Higher productivity helps an ageing society produce more with the workers it has. How far that eases the pressure on public finances depends on where the income goes and how it is taxed. This, almost certainly, will be a hugely controversial and contentious public debate in many countries over the next decade as the structural fiscal squeeze gets worse. [13]

What would count as an offset?

It’s possible that AI, and more broadly automation, can take a meaningful bite out of the demographic drag, but I want to see it to believe it. I would expect the first gains where output is easy to measure, work can be standardised and firms have the resources to reorganise it. An economy-wide offset is a higher bar. The gains have to spread into sectors where labour is scarce and demand is rising, and they have to arrive fast enough to keep up with the fall in hours worked per person. More software shipped with the same team is a productivity gain. More people safely looked after per paid hour is the kind of gain that would make me more confident about the demographic case.

In the meantime, we should watch hiring alongside output. If a firm produces more and hires fewer juniors, we may see the productivity gain well before we know what it has done to the supply of experienced workers five or ten years later. If firms use AI to train people faster and give them more responsibility sooner, the same technology could ease that constraint. Entry-level hiring and progression will tell us more about which path employers are taking than stable headcounts on their own.

That brings me back to Anthropic. The IPO is a bet that enough customers will pay enough for AI to cover the cost of developing and running it, with a margin left over. That can be a good business even if some of the sectors under the most demographic pressure take years to see a material benefit. The number we should be watching is what employers can deliver per hour once AI moves into routine production, especially where there are already too few people to do the work. That is where the promise of offsetting ageing will be tested. [14]

References

1. Anthropic (18 and 20 November 2026), IPO pricing (https://www.anthropic.com/news/anthropic-prices-initial-public-offering) and closing announcement (https://www.anthropic.com/news/ipo-close). Reuters (18 November 2026), Anthropic’s IPO valuation (https://www.reuters.com/legal/transactional/anthropic-prices-blockbuster-ipo-83-share-source-says-2026-11-18/). The Wall Street Journal (19 November 2026), “Anthropic’s IPO Success Shifts Spotlight to OpenAI” (https://www.wsj.com/livecoverage/anthropic-ipo-stock-market-today-11-19-2026/card/anthropic-s-ipo-success-shifts-spotlight-to-openai-d57VPQzNOFyIRKVnkiXJ). Reuters (20 November 2026), Nasdaq weekly decline and Anthropic’s second session (https://www.reuters.com/business/view-nasdaq-plunges-5-day-after-anthropic-debut-investors-still-believe-ai-2026-11-20/).

2. US Census Bureau, Business Trends and Outlook Survey data (https://www.census.gov/hfp/btos/data) and November 2025 AI question wording update (https://www.census.gov/hfp/btos/downloads/AI%20Question%20Wording%20Updates.pdf). November 2026 figure from a public transcription of the Census series (https://programs.com/resources/business-ai-adoption/).

3. Kathryn Bonney et al. (2026), “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks” (https://www2.census.gov/ces/wp/2026/CES-WP-26-25R.pdf).

4. Ivan Yotzov et al. (2026, October revision), “Firm Data on AI” (https://doi.org/10.3386/w34836), NBER Working Paper 34836.

5. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond (2025), “Generative AI at Work” (https://doi.org/10.1093/qje/qjae044), Quarterly Journal of Economics. Salomé Baslandze et al. (2026), “Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives” (https://www.frbsf.org/research-and-insights/publications/system-research-atlanta-fed/2026/04/artificial-intelligence-productivity-workforce-evidence-from-corporate-executives/).

6. Anders Humlum and Emilie Vestergaard (2026 revision), “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI” (https://www.nber.org/system/files/working_papers/w33777/w33777.pdf), NBER Working Paper 33777.

7. Paul A. David (1990), “The Dynamo and the Computer” (https://ideas.repec.org/a/aea/aecrev/v80y1990i2p355-61.html), American Economic Review. Erik Brynjolfsson, Daniel Rock and Chad Syverson (2021), “The Productivity J-Curve” (https://www.nber.org/papers/w25148), American Economic Journal: Macroeconomics.

8. Tania Babina, Alex X. He and Renhao Jiang (2026), “Canaries in the Gold Mine: Early Productivity Gains from Artificial Intelligence Creating Organization Capital” (https://www.nber.org/system/files/working_papers/w35684/w35684.pdf), NBER Working Paper 35684.

9. Daron Acemoglu (2025), “The Simple Macroeconomics of AI” (https://dspace.mit.edu/entities/publication/0fcf7575-a092-4728-a0c5-077beda38427), Economic Policy. Francesco Filippucci, Peter Gal, Katharina Laengle and Matthias Schief (2025), “Macroeconomic Productivity Gains from Artificial Intelligence in G7 Economies” (https://www.oecd.org/en/publications/macroeconomic-productivity-gains-from-artificial-intelligence-in-g7-economies_a5319ab5-en.html), OECD.

10. Daron Acemoglu and Pascual Restrepo (2022), “Demographics and Automation” (https://doi.org/10.1093/restud/rdab031), The Review of Economic Studies. Acemoglu and Restrepo (2017), “Secular Stagnation? The Effect of Aging on Economic Growth in the Age of Automation” (https://www.aeaweb.org/articles?id=10.1257%2Faer.p20171101), American Economic Review. Kamil Kovar, “Will automation offset ageing?” (https://kamilkovar.substack.com/p/will-automation-offset-ageing) and empirical follow-up (https://kamilkovar.substack.com/p/some-empirics-showing-that-automation), European Macro Perspectives. Both essays were inaccessible during research. The discussion uses his publicly reposted summary (https://teapawt.org/benbawan). The empirical follow-up’s results remain unverified.

11. Charles Goodhart and Manoj Pradhan (2017), “Demographics Will Reverse Three Multi-Decade Global Trends” (https://www.bis.org/publications/working-paper-656-demographics-will-reverse-three-multi-decade-global-trends), BIS Working Paper 656.

12. Iñaki Aldasoro, Sebastian Doerr and Daniel Rees (2026), “Old Workers, Young Machines: Can AI and Automation Offset Population Ageing?” (https://www.bis.org/publications/bulletin-135-old-workers-young-machines-can-ai-and-automation-offset-population-ageing), BIS Bulletin 135.

13. Masao Fukui, Emi Nakamura and Jón Steinsson (2026), “The Commoditization of Labor” (https://www.nber.org/papers/w35815), NBER Working Paper 35815.

14. Anthropic (2026), Final IPO prospectus (https://www.sec.gov/Archives/edgar/data/2130357/000162828026074944/anthropicpbc-424b4.htm).

Archive essays consulted include “AI Is Not One Technology” (1 October 2026), “AI, Productivity and Demography,” “Ageing and Automation,” “Technology Diffusion” and “Intangible Investment.” AI Blogs archive (https://drive.google.com/drive/folders/1y0-xC4rm0LLu_OjjBWq2sIFkqHnC9sF5)