AI & The Labour Market: The capability mirage

 

Artificial intelligence and the labour market

The capability mirage

AI can theoretically do far more than it is actually doing. The gap between those two facts is where the real story of workplace disruption lives

Nagesh Bhushan

The headlines write themselves. Robots are coming for your job. The white-collar workforce faces an existential reckoning. Artificial intelligence will hollow out the professional class within a decade. These predictions are not entirely wrong — but they are, at present, substantially premature. The more interesting story is not what large language models can theoretically do. It is the yawning chasm between that potential and what they are actually doing in workplaces today.

Data from the Anthropic Economic Index — which tracks real-world API usage patterns in professional settings — provides the clearest picture yet of this gap. Among computer and mathematics occupations, large language models are theoretically capable of performing 94% of tasks at meaningful speed. The proportion actually being automated today: 33%. Office and administrative roles show an almost identical pattern — a 90% theoretical feasibility rate, with actual automated usage running at a fraction of that. The gap is not marginal. It is the dominant feature of the current AI landscape.

Theoretical vs. observed AI exposure, selected occupations

Percentage of tasks covered, 2024. Gray = theoretical capability; Red = observed (actual) exposure

Sources: Eloundou et al. (2023); Anthropic Economic Index

The last mile problem

Why does the gap exist? Not because the technology is oversold — the 97% correlation between theoretical and actual usage categories confirms that researchers are broadly correct about what LLMs can do. The bottleneck lies elsewhere: in what economists call the “last mile” of implementation. A pharmacist’s system may be theoretically automatable, but a licensed human signature is still legally required for drug refills. A financial model may be buildable by an AI, but the compliance team insists a human check the output. Software interoperability, institutional inertia and legal liability combine to slow adoption far below what technical benchmarks would predict.

This distinction between theoretical capability and observed exposure matters enormously for anyone trying to predict economic disruption. Capability tells you what might happen in a laboratory. Observed exposure — what is actually being deployed today — is the leading indicator of structural change in the labour market. It identifies which roles are genuinely at risk before that risk shows up in unemployment statistics.

“Technical feasibility does not equal immediate utility. The bottleneck is institutional, not computational.”

Front lines and physical guards

The distribution of actual exposure is highly uneven. At one extreme, computer programmers face the highest observed exposure of any occupation — 75% of their tasks are being actively automated via API implementations. Financial analysts, customer service representatives and data entry keyers are close behind, the latter seeing 67% coverage as AI now autonomously reads source documents and enters data. These are roles that involve high volumes of digital-first tasks: the kind of work that can be packaged into an API call.

At the other extreme, roughly 30% of the workforce faces zero observed exposure at present. Motorcycle mechanics. Cooks. Bartenders. Lifeguards. Dishwashers. These roles share a common feature: they require physical presence and manual dexterity that no large language model can replicate. They are shielded not by complexity or creativity but by embodiment — the simple fact that the work happens in the physical world rather than on a screen.

 

The exposure divide — at a glance
Computer programmers (observed)	75%
Data entry keyers (observed)	67%
Computer & math (theoretical)	94%
Office & admin (theoretical)	90%
Share of workforce: zero exposure	~30%
Hiring drop, exposed roles (ages 22–25)	−14%

 

An uncomfortable demographic

Who, precisely, is most exposed? The answer inverts almost everything history suggests about technological displacement. Previous waves of automation — from the loom to the assembly line — targeted manual workers. The current wave is aimed squarely at the educated professional class. Workers in the top quartile of AI exposure are nearly four times more likely to hold a graduate degree than those in the bottom quartile (17.4% versus 4.5%) and earn, on average, 47% more.

The demographic profile adds further nuance. Highly exposed workers are 16 percentage points more likely to be female, 11 points more likely to be white, and nearly twice as likely to be Asian compared with workers who face no exposure. This is the “white-collar displacement risk” — a disruption concentrated not among the economically marginal but among precisely those workers who have historically benefited most from technological change.

AI coverage vs. projected job growth, 2024–2034

Each 10% rise in AI task coverage is associated with a 0.6pp drop in projected job growth

Sources: Bureau of Labor Statistics; Anthropic Economic Index

Transition, not rupture

The aggregate unemployment statistics have not yet registered the shift. But more granular data offers an early warning. For workers aged 22 to 25 in highly exposed fields, the job-finding rate has dropped by approximately 0.5 percentage points per month — translating to a 14% decline in hiring since late 2022. The labour market is not breaking; it is bending, quietly, at the entry level.

This is, on one reading, encouraging. Unlike the sudden shock of a pandemic or a financial crisis, AI-driven restructuring appears to be operating on an “internet-style” timescale — gradual enough that policy has time to respond. The window for intervention is open. Whether governments will use it is a separate question.

The policy agenda writes itself more easily than it is implemented. Workers need formal mechanisms to negotiate with management over how AI tools are deployed, ensuring the technology improves job quality rather than simply intensifying workload. Safety nets need to become adaptive — triggered automatically by measurable displacement thresholds, offering time-limited wage insurance or retraining vouchers rather than permanent dependency. And benefits — healthcare, retirement — need to be decoupled from specific employers, so that workers moving between jobs, or into freelance arrangements, do not fall through institutional gaps designed for a more static labour market.

“The goal is not to slow the technology. It is to ensure the social contract scales alongside it.”

None of this is unprecedented. Earlier technological transitions reshaped entire categories of work without producing permanent mass unemployment — not because displacement did not occur, but because new roles emerged and institutions adapted. The question for the intelligence age is not whether AI will change the labour market. It will, substantially. The question is whether the gap between what the technology can do and what it is doing will close slowly enough — and with enough political attention — for workers to land somewhere decent on the other side.

The capability mirage may be the most important economic fact of the current moment. It buys time. What happens to that time is a choice.

 

 

Sources

Eloundou et al. (2023); Anthropic Economic Index; Bureau of Labor Statistics Occupational Projections 2024–2034; Current Population Survey.