Anthropic’s AI Labor Report Shows a Gap Between Capability and Real Use

AI is increasingly used within individual tasks, but most workflows have not yet integrated it fully. Image credit: KorishTech (AI-generated).

The Anthropic AI labor report measures how AI is actually used in real work—and finds that most tasks AI can perform are not being used in practice.

The report combines two layers. First, it uses task-level estimates of what large language models can theoretically perform. Second, it overlays this with real usage data from Claude, focusing specifically on work-related and automated use rather than casual interaction. This produces what the authors call observed exposure—a measure of where AI is already active in real tasks.

This shift moves the focus from what AI can do to what is actually being used.

Across major occupational categories, the report finds that theoretical exposure is consistently higher than observed exposure.

In Computer and Mathematical roles, prior estimates suggest AI could affect roughly 94% of tasks. Office and Administrative roles reach around 90%. These figures represent what AI systems are capable of doing at the task level.

However, Anthropic’s usage data shows that only a fraction of those tasks are actually used in real work environments, based on Claude platform activity. The report visualises this as a clear gap between the “theoretical” curve and the “observed” curve across occupations.

A second data point reinforces this gap. This is also reflected in the Anthropic AI labor report, which finds no measurable increase in unemployment in highly exposed occupations so far.

A third signal appears in hiring data. For workers aged 22–25, entry into highly exposed roles has declined by about 0.5 percentage points, roughly a 14% drop relative to 2022, while less-exposed roles remain stable at around 2% per month. This also reflects the limited use of AI within tasks rather than full automation.

The report’s central mechanism is straightforward: AI affects tasks, not whole jobs—and only a subset of those tasks are actually being used in practice.

Jobs are made up of multiple tasks. Many of these tasks—especially those that are text-based, codified, and individually completable—are technically feasible for AI systems. This is what theoretical exposure captures.

But for AI to have real impact, those tasks must also be:

• integrated into workflows
• used repeatedly in professional settings
• executed in automated or semi-automated ways

Anthropic’s data shows that this second step is incomplete. Even when tasks are technically feasible, they are often not operationalised.

For example, a programming task that AI can complete in seconds may still be done manually if it is not integrated into a company’s workflow.

This creates a structural gap:

• Capability layer → what AI can do
• Usage layer → what organisations actually implement

The difference between these layers defines current impact.

Occupational Category	Theoretical Task Exposure	Observed Exposure	Task Characteristics	Growth Outlook
Computer & Math	~94%	Partial	Code, structured text, digital tasks	Lower
Office & Administrative	~90%	Partial	Routine documentation, messaging	Lower
Education & Training	Moderate–High	Limited	Mix of prep + in-person interaction	Moderate
Healthcare Support	Low	Minimal	Physical, in-person care	Stronger

This pattern is consistent: high theoretical exposure does not translate into equivalent real-world use.

The gap is most visible in occupations built around digital, text-based work.

Roles in programming, finance, and administrative functions show high theoretical exposure because their tasks are modular and compatible with AI systems. These same roles also show the highest levels of observed usage—but still far below what is technically possible.

In contrast, roles requiring physical presence, supervision, or real-time human judgment—such as healthcare support or classroom teaching—show both low theoretical and low observed exposure. Even where AI can assist (e.g., lesson planning or documentation), it does not replace core tasks.

The difference is not about job titles, but about task structure.

This gap explains why labour market signals remain muted despite rapid advances in AI capability, as the Anthropic AI labor report shows.

If most theoretically automatable tasks are not yet implemented in real workflows, then immediate displacement should not be expected. Instead, the first effects appear in slower adjustments:

• hiring into exposed roles becomes more selective
• growth projections weaken for those roles
• adoption accumulates gradually at the task level

These changes reflect the gap between capability and actual usage rather than full task replacement.

The report’s findings align with this pattern. High-exposure occupations are projected to grow more slowly through 2034, even without current unemployment effects.

This suggests that AI impact may scale through accumulation of task-level adoption, not through sudden job replacement.

Anthropic positions observed exposure as a metric that can be tracked over time. As AI tools become embedded in workflows, the gap between capability and usage is expected to narrow.

The direction of change is clear: if more tasks move from theoretical feasibility into routine use, then labour market effects should become more visible.

However, the timing depends less on model capability and more on:

• organisational adoption
• workflow integration
• trust, regulation, and risk tolerance

The report does not assume these factors resolve quickly.

This report provides a useful baseline because it measures AI use through real workflow data rather than assumptions about capability.

At the same time, the findings depend heavily on Anthropic’s own usage data, which represents only one part of a much broader AI ecosystem. Usage patterns may differ across industries, regions, and other AI platforms, so the results should be interpreted as an early snapshot rather than a complete picture.

What the report makes clear is that automated workflows have not yet been widely implemented, even in roles where AI could theoretically perform most tasks. This reinforces the idea that current impact is constrained not by capability, but by how AI is integrated into real work environments.

However, this gap also represents latent capacity. Many large technology companies continue to invest in more advanced models and enterprise integrations, which suggests that the conditions limiting adoption today may not remain stable. If workflow integration accelerates, the same tasks that are currently unused could be rapidly operationalised.

In that sense, the report does not indicate stability, but timing. The current gap delays visible labour market impact, but it does not remove the underlying potential.

For companies, this creates a practical signal. The question is no longer whether AI can perform certain tasks, but how quickly those tasks can be embedded into workflows. This may shift focus toward reducing inefficiencies, restructuring processes, and identifying where AI can replace repetitive task layers rather than entire roles.

Anthropic — Labor market impacts of AI: A new measure and early evidence
https://www.anthropic.com/research/labor-market-impacts

Anthropic — Estimating AI productivity gains from Claude conversations
https://www.anthropic.com/research/estimating-productivity-gains

Forbes — Anthropic’s Study Does Not Measure AI’s Labor-Market Impacts
https://www.forbes.com/sites/hamiltonmann/2026/03/08/anthropics-study-does-not-measure-ais-labor-market-impacts/