The consumer spending risk that AI is quietly building

Most commentary on AI and financial markets focuses on the productivity upside. Corporate earnings, margin expansion, efficiency gains flowing through to profit. That story is probably correct in the short term.

But it is only half the picture. And the second half does not look as clean.

What the research is starting to show

Anthropic published research this week mapping observed AI usage against theoretical capability across occupational groups. The demographic profile of who is most exposed is the detail worth sitting with.

The most AI-exposed workers earn 47% more on average than the least exposed group, and are nearly four times as likely to hold a graduate degree. The paper specifically names the scenario everyone in the knowledge economy should be thinking about: a "Great Recession for white-collar workers," drawing a parallel to the 2007 to 2009 financial crisis in which US unemployment doubled from 5% to 10%.

The researchers note that a comparable doubling in the top quartile of AI-exposed occupations, from 3% to 6%, would be clearly detectable. It has not happened yet. But the conditions for it are being assembled.

The mechanism worth watching

When businesses automate roles, they capture the benefit in one of two ways: fewer people doing the same work, or the same number of people doing significantly more. Either way, corporate earnings go up in the near term.

The problem is that the workers whose roles are eliminated or whose wages are compressed are also consumers.

This is not a novel observation. It is the same question economists asked about factory automation in the twentieth century. What makes the current moment different is the speed of the shift, and the demographic profile of who is most exposed.

Why the developer cohort matters

Software developers are not a random sample of the workforce. They are disproportionately well paid, concentrated in cities with high property prices, and significant contributors to discretionary consumer spending.

A developer on 70,000 pounds a year is not just a line item on a payroll. They are a mortgage holder, a regular restaurant customer, a business traveller, a regular spender on premium services. That spending supports businesses across hospitality, retail, and professional services whose revenues have no direct relationship to the technology sector.

The Anthropic research identifies computer programmers as among the most exposed occupations. For that group specifically, large language models are theoretically capable of handling 94% of tasks. Current observed usage covers 33%. That gap will close. When it does, the headcount required to maintain current software output drops materially.

If AI compresses developer wages significantly over the next few years, or eliminates a meaningful number of roles, the downstream effect on consumer spending extends well beyond technology.

The lag that makes this hard to price

Markets are reasonably good at pricing near-term earnings changes. They are considerably less good at pricing lagged second-order effects, particularly when the causal chain runs over two or three years.

The sequence runs something like this: AI investment increases, productivity gains appear in corporate earnings, employment starts to fall or wages compress across knowledge worker roles, consumer spending weakens, and businesses in sectors with no direct AI exposure start to miss revenue targets.

By the time that last step becomes visible in earnings reports, the underlying cause is several years upstream. The market is unlikely to see it coming clearly until it is already arriving.

Developers are not the only cohort

The developer example is the most immediate and the most legible, because the productivity impact of AI coding tools is already visible and measurable.

But the same dynamic applies to other well-compensated knowledge worker roles where AI productivity tools reduce the headcount required: legal document review, financial analysis, content production, customer service management at the more complex end.

The Anthropic research names business and finance, management, legal, and office administration as areas where AI is theoretically capable of covering most tasks. These are not low-wage roles. The people in them are meaningful contributors to discretionary consumer spending. Their employment and income levels affect a broader range of businesses than is usually factored into analysis of AI's economic impact.

What the fragmentation does not fully solve

There is a counter-argument worth acknowledging. Some of the employment lost in large technology companies will be absorbed by smaller operators: former SaaS developers building specialist tools, consultants operating independently, small firms serving niche markets. That fragmentation is real and it will create employment.

But those roles are not equivalent in income or spending profile to the roles they replace. A former developer at a mid-size SaaS company, now operating as an independent consultant with a smaller client base, is likely spending less, not more. The aggregate demand effect is still negative even if the employment displacement is partially cushioned.

What investors might consider

I am not making a prediction about timing. Structural shifts of this kind can take longer to show up in market data than the underlying economics suggest they should.

What I would suggest is that investment theses heavily dependent on sustained growth in consumer spending deserve scrutiny about their assumptions. Specifically: how much of that spending is driven by knowledge workers in roles where AI productivity tools are now materially reducing the headcount required to maintain current output?

The businesses most exposed are those in sectors where demand is disproportionately driven by the spending habits of well-paid professionals: premium hospitality, executive travel, high-end retail, discretionary professional services. These are not AI-adjacent sectors. But their revenues are more connected to AI's labour market impact than most current analysis acknowledges.

The risk is real. It is not immediate. That combination is precisely what tends to catch markets off guard.