Which Roles Should Use the Most AI?

The roles that should use the most AI are the ones where most of the work is made of tasks that suit AI. Anthropic’s March 2026 labor market paper measures how much of each job’s work shows up in work use of Claude, and ranks computer programmers first at 74.5%, followed by customer service representatives at 70.1% and data entry keyers at 67.1%. At the other end, the paper finds 30% of workers in occupations with zero coverage, and hands-on work such as production and repair sits close to zero, so little AI use there is usually the right amount.

The general method is set out in Is your AI spend in the right roles?. This article applies it to ten roles from different functions.

What does public data show about AI use by role?

Anthropic’s Economic Index maps samples of conversations with Claude, Anthropic’s AI model, and of business traffic through its API to the tasks listed for each occupation in O*NET, the US Department of Labor’s catalog of occupations. The labor market paper of March 5, 2026 turns that into one figure per occupation, observed exposure. It counts the tasks a language model could speed up in theory that have seen enough work-related use of Claude, gives automated use full weight and assisted use half weight, and weights each task by the time spent on it. It measures task coverage, not how many tokens a role uses, and draws on usage from August and November 2025.

The paper’s Figure 3 lists the ten most exposed occupations. After the top three come medical record specialists (66.7%), market research analysts and marketing specialists (64.8%), wholesale and manufacturing sales representatives (62.8%) and financial and investment analysts (57.2%). At the bottom, 30% of workers have zero coverage, a group that includes cooks and motorcycle mechanics, and the paper’s chart by occupational category shows observed coverage close to zero for production, installation and repair, construction and transportation.

The January 2026 Economic Index report found data entry workers among the highest once each task is weighted by Claude’s success rate and its share of the workday: only two of their nine tasks are covered, but the largest one, reading source documents and entering data, has high success rates. The March edition reported that automated customer service workflows contributed to that role’s exposure, and that automated sales and outreach work, such as lead qualification research and cold-email drafting, at least doubled its share of API traffic in three months. It also describes Claude’s use as “not broadly representative of the US economy,” so treat all of this as an outside reference.

Cost per conversation also differs by role. A token is the unit AI vendors bill by, roughly a piece of a word. In the June 2026 edition, conversations tied to marketing managers’ tasks used about 2.5 times as many tokens as those tied to editors’ tasks, and the two jobs pay about $80 and $37 an hour.

Which roles should use the most AI? Ten roles side by side

The table sets Anthropic’s figures beside JobRoute’s published score for each role. Each O*NET task is scored from 0 to 100 for how much of it AI can do, from use recorded in the Economic Index’s June 2026 release where the task appears there and from a documented rule where it does not, and the role score weights each task by its importance. The recommended tier is the lowest that covers 80% of the work AI can address. Figures were copied from the role pages on November 13, 2026 (data version 2026-11-03); the methodology explains the scoring.

The two measures differ in method and scale, so read them side by side and do not subtract one from the other.

Where do the two measures disagree, and why?

Customer service shows how the methods differ. Anthropic’s labor market paper puts customer service representatives at 70.1%, and the role page gives 44 out of 100. The page scores the main task, conferring with customers to give information, take orders or cancel accounts, at 85 from observed use, and scores follow-up work such as resolving billing complaints and keeping records of each interaction at 30. Anthropic weights tasks by time spent and counts automated use in full, and its March report says automated customer service workflows contributed to the role’s figure. One large task that is automated at scale can put a role near the top for use while much of its other work scores lower.

Sales differs in a similar way. The paper puts wholesale and manufacturing sales representatives at 62.8%, and the role page gives 41. On the page, recommending products scores 100 and answering customers’ product questions scores 86, while negotiating, stock and contract tasks score 30 to 35. A company’s own usage reports show which of these tasks its sales team actually brings to AI.

Accounting and legal support run the other way. The accountants and auditors page shows 23 of 30 tasks scoring 60 or more, 2 of them observed in AI use, and the paralegals page shows 9 of 12, also with 2 observed. The labor market paper names reasons usage can trail what a model could do in theory, among them “legal constraints, specific software requirements, human verification steps.” Most of the high scores for these two roles come from the documented rule rather than observed use, so check them against the team’s own work before moving budget.

Which roles should use little AI?

Roles built on physical work score low on both measures. The general maintenance and repair page gives 21 out of 100, with none of 27 tasks at 60 or more, and the team assemblers page also gives 21, with 72% of the work hands-on. The parts of these jobs that suit AI are the paperwork: planning repair work from diagrams and manuals scores 50 for maintenance workers, and reviewing work orders and blueprints scores 65 for assemblers. Before setting seats for such a role, find out who does that paperwork and with what tools.

What does high or low use mean for each kind of role?

Set each role’s AI use per person, from the AI tools’ usage reports, against the share of its work that suits AI. Each role lands in one of four positions.

Many tasks suit AI and use is high. This is where programmers, data entry keyers and financial analysts should sit if the company’s reports show heavy use. Then ask whether the role’s output measure moved and whether the model is larger than the tasks need. The programmers page puts 21% of the task weight at the efficient tier and 36% at standard, so work such as documentation may not need the largest model. Anthropic’s March report found users already shift model size with the work: 34% of software developer tasks on Claude.ai used Opus, then its most intelligent model class, against 12% of tutor tasks.

Many tasks suit AI and use is low. If the company’s reports show light use by accountants or paralegals, ask those teams what stands in the way, such as training or a rule about client data, and fix that before buying more licenses.

Few tasks suit AI and use is high. If a maintenance or assembly team shows heavy use, ask what the tokens are going to before cutting anything. The answer may be the planners’ paperwork, or a use the public data has not yet recorded.

Few tasks suit AI and use is low. This is the expected position for maintenance and production roles, and there is no reason to push use higher.

For customer service and sales, where the two sources disagree on how much of the work suits AI, read the role’s output measure beside cost and quality before deciding which position applies: tickets resolved per hour for support, or quote turnaround for sales. Measuring AI value outside engineering covers how to choose one.

How do you apply this to your own roles?

Start with the roles that hold the most people or the most AI spend. For each, take AI use per person over a quarter from the usage reports of the tools the company pays for, and set it beside the role’s task score. Check where your own job descriptions differ from the O*NET description, because a paralegal team that only drafts contracts is not the average paralegal. Then attach the output measure each team already tracks for the same period.

JobRoute reads the usage reports of a company’s AI tools, joins them to each role’s task scores and output measure, and shows spend by role, the roles where spend deserves a question, and paid seats nobody is using: see AI spend by role.