Anthropic released findings from its Economic Index report showing that AI adoption varies dramatically across occupations, with certain manual, creative, and trades-based jobs registering almost no AI tool usage. The data, drawn from a privacy-preserving analysis of real-world conversations with the company’s Claude model, offers one of the first large-scale looks at where AI is actually being used and, just as telling, where it is not. For workers in fields that rely on physical skill, spatial reasoning, or hands-on craft, the AI revolution has barely registered.
How Anthropic Measured Real-World AI Use
The data behind these findings comes from a system Anthropic built specifically to study how people interact with its AI models at scale. Described in a methods paper called Clio, the system allows researchers to analyze conversations without exposing raw transcript content. That distinction matters because it lets Anthropic publish occupational usage statistics while maintaining user confidentiality, a balance that most other AI companies have not attempted publicly.
The Clio system works by clustering and categorizing conversation topics at an aggregate level rather than reading individual exchanges. This approach means the resulting data reflects broad patterns of use across professions without requiring anyone at Anthropic to review what specific users typed. It is a methodological choice that shapes both the strengths and the limits of the findings: the data can show which job categories generate AI queries and which do not, but it cannot easily explain the reasons behind those gaps.
Anthropic’s Economic Index builds on this infrastructure by mapping usage patterns to what the company describes as underlying economic primitives of work. Instead of just counting prompts, the framework looks at the kinds of tasks people bring to Claude, such as planning, drafting, or coding, and how those tasks cluster in different occupations. That lens helps clarify why some kinds of jobs show up heavily in the dataset while others barely appear at all.
Desk Jobs Dominate AI Queries
The clearest pattern in the Anthropic data is a sharp divide between knowledge work performed at a computer and work performed with tools, materials, or physical presence. Software development, business analysis, marketing, and similar desk-bound roles account for the bulk of AI interactions. Workers in these fields use Claude for tasks like drafting text, generating code, summarizing documents, and answering technical questions, all activities that translate naturally into a text-based AI interface.
This concentration is not surprising on its own. AI chatbots are, by design, text-in and text-out tools. They excel at language manipulation, pattern recognition in code, and information retrieval. But the degree of the gap is striking. The Anthropic Economic Index report, which traces its analytical framework to research on the structure of tasks rather than job titles alone, found that entire occupational categories barely appear in the usage data at all.
Within office settings, the most intensive users tend to be people whose output is already highly modular and document-based. A financial analyst can ask Claude to explore scenarios, draft memos, or check formulas. A marketer can generate campaign concepts, rewrite copy for different audiences, or adapt a presentation for a new client. Even within the same company, these kinds of roles show far more AI activity than, say, facilities staff or on-site technicians.
Which Jobs Show Almost No AI Activity
Construction workers, electricians, plumbers, and other skilled tradespeople are among the occupations with the lowest AI engagement in the Anthropic dataset. The same holds for roles in food preparation, janitorial services, and many forms of manual manufacturing. These are jobs where the core task involves physical manipulation of materials, spatial problem-solving in real environments, or repetitive motor skills that no chatbot can perform.
Creative and artistic roles also appear far less frequently than many AI enthusiasts might expect. Graphic designers, fine artists, and performers show limited interaction with Claude compared to their knowledge-work counterparts. This is a notable finding given the intense public debate over AI-generated art and the assumption that creative professionals would be early and heavy adopters, or at least frequent experimenters.
The low usage among creative workers suggests something more complicated than simple unfamiliarity with the technology. Many artists and designers are well aware of AI tools but may actively resist integrating them into workflows that depend on personal style, tactile feedback, and iterative physical processes like sketching or sculpting. A painter working with oils and a carpenter fitting joints share something in common: their expertise lives in their hands and eyes, not in a text prompt.
Even in hybrid creative roles that do involve computers, such as video editing or 3D modeling, the way work is structured can limit the appeal of a general-purpose chatbot. Specialists often rely on tightly integrated software suites, custom scripts, and visual interfaces. For them, a separate text box in a browser may feel like an interruption rather than an enhancement, especially if it cannot directly manipulate project files or timelines.
Why the Gap Is Not Just About Access
One easy explanation for uneven AI adoption is that workers in lower-usage fields simply lack access to the tools. That explanation does not hold up well under scrutiny. Smartphones and laptops are widespread across income levels in the United States, and free tiers of AI chatbots are available to anyone with an internet connection. The barrier is not hardware or cost.
A more convincing explanation involves task fit. AI chatbots are useful when work can be broken into discrete informational or linguistic subtasks. Writing an email, debugging a function, or summarizing a legal brief all have clear text-based inputs and outputs. Replacing a section of drywall, adjusting a plumbing joint to account for an unexpected pipe angle, or shaping clay on a wheel do not. The physical world resists digitization in ways that office work does not, and current AI tools have no mechanism for bridging that gap.
Cultural factors likely play a role as well. Trades and craft professions often place high value on apprenticeship, hands-on learning, and personal judgment developed through years of practice. Introducing an AI assistant into that workflow can feel irrelevant at best and insulting at worst. A master electrician diagnosing a wiring problem in a 1920s house is drawing on pattern recognition that no language model can replicate, because the relevant information is visual, spatial, and tactile rather than textual.
There is also the question of time pressure and context. Many manual and service jobs unfold in noisy, fast-moving environments, where pulling out a phone to consult an AI system is impractical. By contrast, office workers already sit in front of screens, making it easy to add another tab to their existing digital toolkit.
What Low Adoption Signals for the Labor Market
The uneven distribution of AI use has real consequences for how the technology reshapes work. If AI tools primarily accelerate productivity in white-collar knowledge jobs, the economic benefits and disruptions will concentrate in those sectors first. Office workers may find their output per hour increasing while their employers need fewer of them. Meanwhile, tradespeople and manual workers could see relatively little change to their daily routines or job security in the near term.
This creates an unusual dynamic. Much of the public conversation about AI and employment focuses on fears of mass displacement. But the Anthropic data suggests the displacement risk is highly uneven. A paralegal faces a very different AI exposure profile than a welder. Companies racing to integrate AI into their operations will find that some departments transform quickly while others remain largely untouched, creating internal friction over resource allocation and workforce planning.
There is also a wage dimension to consider. Many of the low-AI-usage occupations, particularly in the skilled trades, already face labor shortages. If AI fails to automate or augment these roles, demand for human workers in plumbing, electrical work, and construction could remain strong or even grow as other sectors shed jobs. The irony would be significant. The jobs least touched by AI could become relative safe havens, while mid-level office roles face compression as software eats away at routine tasks.
At the same time, low current usage does not guarantee permanent insulation. As AI systems become more tightly integrated with sensors, robotics, and augmented reality, new forms of assistance could emerge for hands-on work, from overlaying wiring diagrams onto walls to guiding repairs in real time. For now, though, Anthropic’s Economic Index suggests that the frontier of AI adoption runs straight through the cubicle farm, not the construction site, and that any serious discussion of the technology’s labor impact has to grapple with that divide.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
