Bottom line: In Anthropic's occupation data, observed AI use sits well below theoretical capability. In Computer & Math jobs, an LLM could theoretically speed up 94% of tasks, and work-related Claude use covered 33%. For most businesses, the practical question is which of their own tasks sit in that gap.
In the occupations with the most theoretical exposure to AI, actual use still covers a fraction of the tasks. That gap is the opportunity.
Every figure on this page comes from one source: Anthropic's research paper Labor market impacts of AI: A new measure and early evidence (Maxim Massenkoff and Peter McCrory, 5 March 2026) and the occupation data Anthropic released with it. The paper compares two measures. Theoretical coverage is the share of a job's tasks that an LLM could theoretically make at least twice as fast, using the task ratings from Eloundou et al. (2023). Observed coverage, which Anthropic calls observed exposure, is the time-weighted share of a job's tasks that Anthropic sees in work-related Claude use, with automated use counted in full and assisted use counted at half weight. Both measures use US occupation data, so they describe US jobs. Australian figures may differ.
Occupation Breakdown: Capability vs Observed Use
Anthropic groups its results by occupational category, and it states exact figures as numbers for two of those categories. In the chart below, the pale bar shows theoretical coverage and the solid bar shows observed coverage in Claude use. Figure 2 of the paper plots the other 20 categories, including Legal, Business & finance, Sales and Construction, but the paper does not give their values as numbers, so they are left out here.
Anthropic also publishes observed coverage for individual occupations. Figure 3 of the paper lists the ten most exposed, and Anthropic's occupation data file gives the figure for 756 occupations. The rows below are occupations common in Australian small and medium businesses, using Anthropic's own occupation titles. Anthropic does not publish a theoretical figure for each occupation, so only the observed bar is shown.
What the Numbers Actually Mean
The 90% theoretical figure for Office & Admin is a statement about tasks. It says that an LLM could in theory make 90% of the tasks in those jobs at least twice as fast, either on its own or with extra software built on top. It makes no prediction about how many office workers will lose their jobs.
Observed coverage shows how much of that is happening in practice. Anthropic reports that Claude covers 33% of all tasks in the Computer & Math category, against a theoretical figure of 94%. Individual jobs vary widely. Data entry keyers, whose main task is reading source documents and entering data, are 67% covered.
The two measures point the same way. Anthropic found that 97% of the tasks it observed in Claude use fall into categories rated as theoretically feasible. The gap between them comes from model limitations and from slower diffusion, which the paper puts down to legal constraints, specific software requirements, human verification steps and other hurdles.
The gap between theoretical coverage (94%) and observed coverage (33%) in Computer & Math occupations. This is the only category where Anthropic's paper states both figures as numbers.
Occupation-by-Occupation Breakdown
Office & Admin: 90% theoretical coverage
Anthropic puts theoretical coverage for Office & Admin occupations at 90%. Observed coverage varies by job: 70.1% for customer service representatives, 67.1% for data entry keyers, 45.0% for general office clerks and 31.0% for bookkeeping, accounting and auditing clerks. Anthropic says it increasingly sees the main tasks of customer service representatives in first-party API traffic.
Legal: under a third of tasks observed
The paper does not state a theoretical figure for the Legal category as a number. Observed coverage is 29.3% for paralegals and legal assistants and 16.7% for lawyers. The study names representing clients in court as a legal task that remains beyond AI's reach. For drafting and research, a sound pattern is for AI to draft and a solicitor to review, on infrastructure that protects privilege and confidentiality.
Business & Finance: analysts lead
Two of the ten most exposed occupations in the study are business and finance roles: market research analysts and marketing specialists (64.8%) and financial and investment analysts (57.2%). Accountants and auditors sit at 34.8%. In the paper's appendix, Business & Financial is one of the four broad categories in the top quartile of observed coverage, along with Computer & Mathematical, Office & Administrative Support and Sales.
Education & Library: varies by teaching level
The paper does not state a theoretical figure for Education & Library as a number. Observed coverage is 29.0% for secondary school teachers, 10.3% for elementary school teachers and 0% for preschool teachers. The paper's own example of the task-based view is that AI can grade homework but cannot manage a classroom.
Healthcare Practitioners: records work leads
Medical records specialists are among the ten most exposed occupations in the study, at 66.7%, with compiling, abstracting and coding patient data as the leading automated task. Registered nurses sit at 6.0%. Anthropic notes that some healthcare tasks rated as theoretically feasible, such as authorising drug refills, have not been seen in Claude use.
Computer & Math: 94% theoretical, 33% observed
Computer programmers top Anthropic's list at 74.5%, with writing, updating and maintaining software as the leading automated task. Software quality assurance analysts and testers (51.9%), information security analysts (48.6%) and computer user support specialists (46.8%) are also in the top ten. Across the whole category, Claude covers 33% of tasks against a theoretical figure of 94%.
Why the Gap Persists
Anthropic gives two broad reasons why observed use falls short of theoretical capability. Some tasks are held back by model limitations. Others spread slowly because of legal constraints, specific software requirements, human verification steps or other hurdles. Three practical blockers sit alongside those:
1. Workflow integration cost. Knowing that AI can draft a contract is different from having a workflow where contracts flow from a template, through AI drafting, to solicitor review, to CRM update. Designing and building that workflow takes deliberate effort, and many businesses skip it.
2. Data sovereignty uncertainty. For regulated industries, the question "where does my data go?" remains unanswered for most generic AI tools. This slows adoption in regulated fields such as healthcare, legal and financial services.
3. The wrong first use case. Many businesses try AI on a low-value task, get mediocre results, and conclude AI is not ready. A better start is the single biggest time bottleneck in the business.
What This Means for a Queensland Business
If most of your team's work sits in office, admin, finance or customer service tasks, Anthropic's data suggests that a large share of those tasks could be done faster with AI, and that real use still covers a smaller share. The figures are US averages, so treat them as a guide to where to look first. They do not measure your own business.
Businesses that close part of that gap early build up workflow knowledge, trained teams and refined systems, and those advantages compound over time.
The useful question is "which specific task, in which specific workflow, with what data infrastructure?" That is the conversation worth having.
Frequently Asked Questions
Which occupations have the most tasks AI could speed up?
Anthropic's March 2026 labour market study reports its figures by US occupational category and does not break them down by industry. It states that an LLM could theoretically speed up 94% of tasks in Computer & Math occupations and 90% of tasks in Office & Admin occupations. Measured by actual work-related Claude use, the most exposed single occupations were computer programmers (74.5%), customer service representatives (70.1%) and data entry keyers (67.1%).
What is the capability-versus-adoption gap?
It is the difference between the share of a job's tasks that an LLM could theoretically speed up and the share Anthropic actually sees in work-related Claude use. In Computer & Math occupations the theoretical figure is 94% and the observed figure is 33%. Anthropic's usage figures come from Claude alone, so they describe one platform's traffic in the US occupation data.
Why is AI adoption lower than its capability?
Three reasons dominate. First, businesses skip the workflow audit, try AI on the wrong task and conclude AI is not ready. Second, change management gets too little investment. The technology works, but the team does not change how they work. Third, procurement and compliance cycles move slower than tool capability, especially in regulated sectors. The common thread is that the blocker is rarely the model.
Which jobs show the least AI use?
In Anthropic's data, 30% of US workers are in jobs with zero observed coverage, because their tasks appeared too rarely in Claude usage to meet the minimum threshold. Examples the study names include cooks, motorcycle mechanics, lifeguards, bartenders and dishwashers. The study also names physical agricultural work, such as pruning trees and operating farm machinery, as beyond AI's reach.
How should an Australian business pick its first AI use case?
Start with the single biggest non-billable time sink in the business: the task that consumes the most staff hours without producing revenue. Most of the time this is document drafting, data entry, client intake, or reporting. AI deployed against the largest time bottleneck pays back in weeks; AI deployed against a generic experiment rarely does. The order matters more than the tool.