The Retirement Machine: AI, high earners, and the hidden risk to global stock markets
The monthly retirement savings of tens of millions of Americans provide a prop to the equity market. Could it be vulnerable?
Authors
Every month, tens of millions of Americans have money quietly swept from their paycheques and invested in stocks. Most of them barely notice. But in aggregate, this mechanism — the 401(k) system and its sister retirement vehicles — has become one of the most powerful structural forces in financial markets. It has become a large, systematic source of equity demand, running in the background regardless of the headlines.
What could put this US equity “prop” at risk? Shifting demography could play a part, as savers retire and the balance shifts from contributions to withdrawals, but other factors pose a more immediate risk. One is AI and its potential impact on high earners.
In their AI-boom scenario, our economists model unemployment rising to 10% as workers are displaced by technology. Importantly, whereas past technological revolutions have displaced lower earners, this time higher-earning white-collar workers are in the firing line. And they are the ones who disproportionately contribute to the monthly pension flows into equity markets. This article puts numbers on just how material a risk this is. Spoiler alert: it’s big.
The US pension system is enormous
Americans have $49.1 trillion earmarked for retirement as of late 2025. To put that in context, the US retirement pot is nearly as big as the $59.6 trillion US equity market and is bigger than all non-US developed and emerging stock markets combined (based on MSCI US, World ex US and emerging market equity indices).
The biggest slice — $19.2 trillion — sits in individual retirement accounts (IRAs). The next largest is the employer-sponsored defined contribution (DC) world, of which 401(k) plans are the dominant form at $10.1 trillion, with another $4.1 trillion in other kinds of DC plan. Add it all together and the DC-plus-IRA complex controls around $33 trillion. It is broadly the same size as the non-US global equity market.
Figure 1: The US retirement system is huge: DC plans + IRAs are worth the same as all non-US developed and emerging stock markets combined
Market size, $ trillions
Equity market sizes are market capitalisations of MSCI indices. As at end 2025. Source: Investment Company Institute factbook, LSEG Datastream, MSCI, Schroders
Steady flow of equity demand
The stock of assets matters but the flow matters more: contributions arriving month after month, year after year, regardless of market conditions. In 2024 alone, DC plans and IRAs took in an estimated combined $980 billion of gross contributions (excluding rollovers from other kinds of retirement savings vehicle, to capture genuine contributions).
This money goes mostly straight into equities, either directly in equity funds, or via the equity portion of target-date funds. The (dollar-weighted) average allocation to equities across 401(k)s is 75%. Younger individuals have more, with those in their 20s having 90% on average but even those in their 60s still hold over 60%.
Assuming the 75% equity allocation also applies to new contributions, US DC plans and IRAs funnelled around $735 billion into global equity markets in 2024. Given the US weight in the global stock market and typical TDF allocations, 60-70% of that will have gone into US equities, translating into approximately $480 billion. DC and IRA pension accounts bought an estimated nearly half a trillion dollars of US equities in 2024.
To put this into context, net aggregate demand for US equities from other major investors plus nonfinancial companies themselves (via buybacks and M&A, net of share issuance) in 2024 was $1,373 billion. In terms of scale, estimated DC+IRA US equity buying was equivalent to roughly one-third of reported net demand from the other major investor groups and companies themselves in 2024 (Figure 2).
What constitutes a “normal” year?
2024 was also a year when other investor types were buying. 2022 provides a contrasting example. It ushered in the transition from a world of low inflation, near-zero rates and abundant liquidity to one of high inflation, rapidly rising interest rates and much tougher conditions for both stocks and bonds. Household demand for individual US equities fell 76%, ETF buying fell 43%, and overseas investors pulled money out of US equities. Aside from companies buying back their own shares or being acquired, the only meaningful bid for US equities came from DC and IRA pensions. The market still fell by 20%. Without pensions’ stabilising source of demand, it could easily have been worse.
In more serious economic crises such as 2000-02, 2008 and 2020, companies also flipped from buying back their shares to issuing lots, to provide liquidity and bolster their balance sheets. It is not hard to envisage DC and IRA pension funds being the last remaining buyer in a future crisis.
This is not an apples-to-apples accounting comparison and it’s important to be clear that Figure 2 does not show a decomposition of demand. DC and IRA contributions are estimated gross buying flows, while the other rows show net buying after sales.
However, the question this article seeks to answer is: how consequential would it be for equities if the US experienced a rise in AI-linked unemployment among high earners which led to a fall in pension contributions. For that question it is gross pension flows which we need to focus on.
The logic is simple. Retirees also draw down income but those outflows are relatively independent[1] of the economic cycle. It is changes in gross contributions that ultimately drive changes in the net equity flow.
The counterfactual can never be known, but the scale and persistence of DC and IRA flows suggest they have been an important source of valuation support for US equities, especially during periods of market stress.
[1] We say “relatively independently” because the dominant driver of when someone retires is usually their age. However, in an AI-boom scenario where unemployment rises sharply, it is reasonable to expect that some people may retire early as an alternative to being made redundant. Others may seek to access their accumulated savings early to pay the bills when out of work. Both of these would accelerate outflows, amplifying the negative impulse on equity markets from the DC+IRA channel.
Figure 2: Persistent DC and IRA inflows have been a major structural bid for US equities
Past performance is not a guide to the future and may not be repeated
2024 IRA contribution estimated. Calculations assume 75% of gross DC+IRA flows go into equities and 65% of that flow goes into US equities. Households refers to direct ownership only, avoiding any double counting with ETFs/mutual funds. The ETF/mutual fund categories largely offset each other and represent a shift in investor preferences. Source: BEA, Federal Reserve Economic Data (FRED), Federal Reserve Bank of St. Louis, IRS, Morningstar, Schroders
Who do these pension contributions come from?
The retirement system is funded disproportionately by high earners and those working in particular industries (Figure 3). DC plan participation runs at 80% in information industries and 87% in financial activities versus 47% in construction and just 14% in leisure and hospitality. 95% of financial services workers have access to a DC plan and 92% of those take that up; 38% of leisure and hospitality have access and only 37% take it up. The pattern is clear: white-collar, knowledge-intensive industries have near-universal coverage and high participation; lower-wage service industries participate far less, some barely at all.
Figure 3: People working in higher-earning industries are far more likely to contribute to a DC pension plan than lower earners
Past performance is not a guide to the future and may not be repeated
Data as of 2025. Note that, with the exception of employees in the educational services industry, very few individuals have access to a defined benefit (DB) plan. Only 1% of leisure & hospitality workers participate in a DB plan. Variation in DC participation rates is not explained by access to DB pensions as an alternative. Source: BLS, and Schroders
It is not only that higher earners are more likely to participate in a DC plan. They also contribute a higher percentage of their salary and are far more likely to max out, receive larger employer matches, and make additional after-tax contributions when given the opportunity. These additional contributions can allow individuals to save tens of thousands of dollars beyond the usual employee contribution limit.
Once pay is layered on top, the impact grows even wider. Contributions are typically a fixed percentage of pay, so a 10% contribution from someone earning $150,000 is three times the size of one from someone earning $50,000. Coverage, contribution behaviour and pay all push in the same direction.
The result is a highly concentrated contribution base. The Congressional Budget Office’s most recent comprehensive distributional analysis of retirement-related tax expenditures — the closest available proxy for the distribution of contribution dollars — found the highest-income quintile captured 63% of the tax benefits in 2019, against 1% for the bottom quintile. The tax-benefit share will overstate the contribution share to some extent because higher earners face higher marginal tax rates. Even so, this is broadly consistent with IRS data which shows that 62% of contributions to IRAs in 2023 came from people earning over $100k a year. Only about 15% came from people earning what was then the median income of around $50k or less.
There is also direct evidence from within company plans. As Figure 4 shows, the best paid 20% of employees at a given firm have, on average, taken in 44% of employer contributions.
Figure 4: Employer contributions are concentrated among higher earners
The chart shows the share of employer contributions accruing to workers, by within-plan income quintile. “Income” refers to a worker’s benefit eligible income that can qualify for employee or employer contributions (for example, subject to the benefit compensation limit, which was $330,000 in 2022). Results are participant-weighted among 8,479 plan-years between 2013 and 2022. Source: Are employers optimizing their 401(k) match? Vanguard, Yale, MIT
Figure 4 gives a conservative lower anchor: the highest-paid 20% of workers within firms receive 44% of employer contributions. The CBO tax-benefit data provide an upper anchor, with the top income quintile receiving 63% of retirement-related tax benefits. Because tax benefits overstate contribution shares, I use 60% as a rough upper bound. Combining these with estimated contribution-flows into US equities suggests higher earners invested $210-287 billion in US equities via their DC and IRA plans in 2024 (Figure 5). Given the uncertainty, $200–300 billion is a plausible working range.
Figure 5: Higher earners invested an estimated $200-300 billion in US equities via their DC and IRA pensions in 2024
Source: BEA, CBO, Federal Reserve Economic Data (FRED), Federal Reserve Bank of St. Louis, IRS, MIT, Morningstar, Schroders, Vanguard, Yale.
The AI question
Historically, the jobs exposed to automation have been routine, physical or semi-skilled: assembly lines, data entry, call centres. The narrative was that knowledge workers — lawyers, analysts, engineers, doctors — were safe. This is where AI has the potential to change things.
A growing body of research reaches the same broad conclusion: higher-income, white collar occupations requiring postsecondary education show the highest exposure to AI capabilities (Brynjolfsson and others (2018); Webb (2020); Muro and others (2019); Felten and others (2023); Eloundou and others (2024); Kinder and others (2024); and Hampole and others (2025)).
There is a big difference, however, between tasks AI can perform and jobs AI will eliminate. Some tasks may be substituted, leaving fewer people needed. Others may be augmented, making workers more productive rather than redundant. A financial analyst might save 30% of report writing time but that does not automatically imply 30% fewer analysts. It might lead to more output, new services, lower costs, different task allocations etc. It could also lead to fewer analysts but the pass-through is not a given. Impacted workers may also be able to retrain and get new jobs, although not immediately and not necessarily on what they were earning before.
So far, analysis of employment trends by the Stanford Digital Laboratory finds little evidence that AI has caused broad-based job losses across the economy. However, younger workers appear more vulnerable. Employment among 22-25 year olds in AI-exposed occupations has fallen significantly behind that of their peers in less-exposed roles, while more experienced workers have seen little comparable impact (Figure 6). The gap has widened over time and, importantly, so far appears to reflect weaker hiring rather than increased layoffs.
Figure 6: Employment index by age group and AI exposure
Source: Stanford Digital Economy Lab, ADP Research, Eloundou et al. (2024). Results use a five-year balanced sample of firms using ADP payroll services and include all workers matched to an occupation code.
You see similar trends within industries. Software development is a good example, given it is one of the most-exposed occupations (Figure 7).
Figure 7: Software developer employment index by age group
Source: Stanford Digital Economy Lab, ADP Research, Eloundou et al. (2024). Results use a five-year balanced sample of firms using ADP payroll services and include all workers matched to an occupation code.
We are still in the first innings though when it comes to the potential disruptive effects of AI. In their AI-boom scenario, Schroders economists model unemployment getting to 10% by 2030, similar to a 12% economy-wide unemployment projection from Anthropic. Anthropic also projects that 22% of knowledge workers could lose their jobs by 2030 (one of many plausible scenarios, not a prediction). Some would find alternative work – “At the individual level, it means coders and call service center agents may have to switch to jobs like electrician and nurse, which are less exposed to AI” – although affected workers may be unemployed for long periods of time. In this scenario, Anthropic says “Most knowledge workers face either lower wages or unemployment“.
A research memo from Citrini Research earlier this year, written as if from 2028, went viral by painting a dystopian vision of how these trends could evolve. It is only a scenario, not evidence, but it captures the uncomfortable version of the risk.
“AI is now a general intelligence that improves at the very tasks humans would redeploy to. Displaced coders cannot simply move to “AI management” because AI is already capable of that…
They write essentially all code. The highest performing of them are substantially smarter than almost all humans at almost all things. And they keep getting cheaper.
AI has created new jobs. Prompt engineers. AI safety researchers. Infrastructure technicians. Humans are still in the loop, coordinating at the highest level or directing for taste. For every new role AI created, though, it rendered dozens obsolete. The new roles paid a fraction of what the old ones did…
In this cycle, the job losses have been concentrated in the upper deciles of the income distribution… When these workers lost their jobs, or took 50% pay cuts to move into available roles, the consumption hit was enormous relative to the number of jobs lost…a Claude agent can do the work of a $180,000 product manager for $200/month.”
You could take the junior-first pattern that the data currently shows as reassurance for the pension channel since junior workers contribute less than managing directors. But that would be naive. Junior workers in high-paying industries are still high earners by economy-wide standards. A 25-year-old software engineer on $140,000, or a first-year banking analyst on $120,000, sits comfortably in the top quintile of all US earners. The displacement happening right now is concentrated in industries where even the juniors out-earn the median American worker by a factor of two. These are exactly the contributors identified earlier. In addition, participants in their 20s hold 90% equity on average, the highest allocation of any age cohort. So even a junior-first shock would hit a group whose pension contributions are unusually equity-heavy.
Just as importantly, junior-first describes where current AI tools are biting first. We don’t know where it will end. The substitution frontier has moved up the capability curve every year since 2022, and there is no obvious reason to expect it to stop at the entry level.
A gradual risk, layered on top of other pressures
This gives a more practical way to think about the risk. DC and IRA pension accounts bought an estimated $478 billion of US equities in 2024. If higher earners account for roughly 45–60% of those contribution dollars — a range consistent with within-firm data at the low end and the CBO tax-benefit evidence at the high end — then around $200–300 billion of annual US equity demand is linked to the cohort most exposed to AI disruption.
The Global Financial Crisis provides a historical precedent, but only a limited one. It’s limited because, although the unemployment rate hit 10%, lower-skilled, lower-earning workers bore the brunt. This was true both of the level of their unemployment rate and its increase. As outlined earlier, these individuals are less likely to contribute to a DC pension and, when they do, their contributions are lower in dollar terms. The result is that DC contributions fell but only by a relatively modest 3.5% in 2009 (IRA contributions fell by slightly more from their peak, -8%). The average period spent unemployed also reached 35 weeks. While this was a record, it is not on the same scale as the potentially permanent disruption AI could cause across entire industries. If an AI boom led to a material and persistent rise in unemployment among higher earners, the impact on pension contributions would be bigger and longer lasting than during the GFC.
That is not a forecast that this entire flow disappears. Task exposure is not the same as job loss, and displaced workers may retrain or move into new roles. But even a partial impairment matters. A 10% hit to that exposed flow would remove roughly $20–30 billion of annual equity buying; a 25% hit would remove $50–75 billion; a 50% hit would remove $100–150 billion. In a normal year that is material. In a downturn, when buybacks, foreign demand and retail risk appetite are likely to weaken at the same time, it could matter a lot more.
This is still only a partial measure of the risk. Higher earners matter most because they account for a disproportionate share of contribution dollars, but they are not the only contributors. Lower- and middle-income workers would also be affected in a broad AI-driven labour-market downturn. Their individual contributions are smaller, and their equity allocations may differ, but the aggregate flow is not zero. The higher-earner estimate should therefore be read as the most clearly identifiable part of the exposed flow, not the full amount that could be at risk. The wider impact would depend on how far displacement extended beyond high-income knowledge workers, how quickly people found new jobs, and how much earnings recovered.
A global risk, not just a US one
The analysis has focused on the US because the data are deepest and the DC system is largest. But the mechanism is not uniquely American. DC pension systems are growing across advanced economies, and the same high-income industries — finance, technology and professional services — are central to contribution flows elsewhere.
The risk is also correlated. AI capabilities are global, the affected jobs are similar across countries, and the equity markets receiving the flows increasingly overlap. US target-date funds allocate materially overseas; Australian superannuation funds hold large international equity positions, with the US a major destination. Slower contributions in one country therefore reduce demand elsewhere too.
That is what makes this different from a local labour-market shock. A global AI displacement event would hit the same kind of pension contributor in multiple countries at the same time.
Verdict
A rapid displacement of high-paying jobs by AI would likely have far-reaching social and economic consequences. In terms of equity markets, the AI risk looks more serious when framed as a flow risk rather than only a labour-market risk.
US DC plans and IRAs together generate nearly a trillion dollars of annual contributions, with roughly half a trillion dollars estimated to flow into US equities. A plausible $200–300 billion of that annual equity flow comes from higher earners — a group which is highly exposed to AI disruption.
This does not mean those flows vanish. Task exposure is not job loss, and any labour-market adjustment would be gradual. Nor is AI the only pressure on the retirement machine. Demography will also matter as ageing populations shift more money from saving into drawdown. But the retirement channel is not just another buyer of equities. It is one of the most consistent: money arrives every pay period, largely irrespective of sentiment, headlines or geopolitics.
That consistency is what makes it powerful. It is also what makes a weakening of the flow worth monitoring. The world’s equity markets have benefited from automatic retirement buying for decades. AI raises the uncomfortable question of whether part of that structural bid is more fragile than investors assume.
Subscribe to our Insights
Visit our preference centre, where you can choose which Schroders Insights you would like to receive.
Authors
Topics