Is Your AI Spend in the Right Roles?

Tracking the AI bill has become common. In a Deloitte survey of 3,500 senior leaders in 25 countries, published October 29, 2026, 71% said they track their total spend on running AI models and 51% said they track token volume. A token is the unit AI vendors bill by, roughly a piece of a word.

A total shows how much was spent, and a breakdown by team shows who spent it. What neither shows is whether the teams using the most AI are the ones whose work suits it. That question needs a second piece of information beside the usage figure: what the work in each role consists of, task by task.

Why does a usage count need a second number beside it?

Usage differs between roles for good reasons. A team that writes software and a team that inspects equipment on site will not use the same amount of AI, and the gap between them says nothing about either team’s effort. Amazon’s engineering leadership made the point in a memo published October 27, 2026. It called input measures such as token use “deeply flawed” and told teams that when their work does not benefit from current AI tools, “that is a finding, not a failing.”

That sentence describes half of the comparison this article sets out. A low count in a role where little of the work suits AI is the expected result. The same low count in a role where most of the work suits AI means something else, and a usage dashboard alone cannot tell the two apart.

Where does data on the work come from?

Three public sources cover most of the need. The first is O*NET, the US Department of Labor’s catalogue of occupations. It lists the tasks that make up each occupation and rates how important each task is to the job.

The second is usage research that maps real AI conversations to those tasks. The Anthropic Economic Index does this for Claude, Anthropic’s AI model. We covered what its three 2026 editions measure in a separate article, so this piece uses only the figures that bear on where use concentrates.

The third joins use to capability. Anthropic’s paper Labor market impacts of AI, published March 5, 2026, defines a measure called observed exposure. It combines O*NET tasks, earlier academic estimates of which tasks a language model could speed up, and Claude usage data. The paper puts computer programmers at the top with 75% of their tasks covered. It also reports that 30% of workers are in occupations with zero coverage, because their tasks appeared too rarely in the data to count.

JobRoute scores every O*NET task in a role from 0 to 100 for AI exposure, meaning how far the task suits AI, using observed use from the Economic Index where it exists and a documented rule where it does not, and records which size of model each task needs. The methodology page sets out the scoring.

Where does AI use concentrate today?

Use is concentrated in a narrow band of work, and more so in business software than in chat. In the January 2026 edition of the Economic Index, tasks tied to computer and mathematical occupations made up 46% of traffic to Anthropic’s API, the route software uses to send work to the model. The March edition found that the ten most common tasks accounted for 33% of API traffic and 19% of conversations in the consumer product.

The June edition added a survey of about 9,700 users, which Anthropic says is “not representative of the general population”. About 30% of respondents worked in computer and mathematical occupations, which hold 4% of US jobs. People in management occupations were 23% of respondents, while management work accounted for only 4% of sessions. The job a person holds and the tasks that person brings to AI are two separate things, which is why the comparison has to be made at the level of the task.

The labor market paper shows how far use still sits below capability. For computer and math occupations, a language model could in theory speed up 94% of tasks, and 33% of them show up in actual use of Claude. The paper’s own summary is that “AI is far from reaching its theoretical capability.”

All of these figures describe one vendor’s product. They are an outside reference, and a company may or may not find the same pattern in its own data: heavy use in technical roles, and other roles with suitable tasks where use has barely started.

What does the map look like?

Take each role and answer two questions. How much AI does the role use per person over a set period, such as a quarter, compared with the middle of the range for the company? What share of the role’s work can AI do, according to the task data? Each answer is high or low, which gives four positions. Dividing by headcount matters, because a raw total mostly reflects how large the team is.

The position that draws attention first is high use where few tasks fit, because it looks like waste. Sometimes it is waste, and sometimes it is a team that has found a use the public data does not yet record, so the first step is to ask the team what the tokens are going to before any budget is cut.

The position that usage dashboards hide is low use where many tasks fit. A dashboard built on activity shows who is using the tools. Activity alone cannot show that a quiet team has work the tools would suit, because that judgment needs the task data. Where the team confirms the fit, the next step is training and rollout.

How do you tell whether the right place is paying off?

A role with high use and many suitable tasks still has two open questions. The first is whether the work improved. Most roles have a measure their own managers already track, such as tickets resolved or days to close the books. Read that measure from the system where the work happens, for the same team and the same period as the usage figure. A rise in output alongside a rise in use is an association. Other things change at the same time, including staffing and the mix of work, so treat it as evidence that still has to be tested.

The second is whether the model is larger than the task needs. Vendors charge more per token for their more capable models. Anthropic’s model documentation lists Claude Fable 5.2 at $8 per million input tokens, Opus 5.5 at $4 and Sonnet 5.5 at $2. Choice of model also varies by product. In the June Economic Index report, 54% of Claude Code sessions were served by Opus, against 10% of chat and Cowork conversations. A role whose suitable tasks are mostly summaries and first drafts may be well served by a smaller, cheaper model. Test that on the role’s own work and quality bar before switching. Where it holds, the saving comes without asking anyone to use AI less.

What do you need to build it?

Start with the usage reports from the AI tools the company already pays for. They differ in what they expose. Anthropic’s Enterprise Analytics API returns token usage and cost per user on usage-based plans, while Microsoft’s Copilot usage report returns the date each user was last active in each app and no token count. Where a tool reports activity only, keep it as a separate, labelled measure and do not add it to token figures.

Then map each user to a role through the HR system, and add the figures up by role or team, divided by headcount, before anyone looks at them. Match each role to its tasks and score those tasks. Last, attach the output measure each team already tracks.

JobRoute joins the three, token use, the role’s tasks and the role’s output, for each role: see AI spend by role.

Who else measures AI spend?

Several products measure AI spend or usage, and a buyer should compare them. Larridin brings token usage, seat licences and cloud model costs into one spend view and reports adoption by team and role. Rippling’s AI Spend Console, announced August 6, 2026, ties AI spend to employee attributes and breaks costs down by teams, roles or departments. Microsoft’s Copilot Business Impact report sets Copilot usage beside outcome data that the customer uploads. All three pages were read on November 11, 2026. None of them describes scoring the tasks in each role as the starting point, which is the comparison this article is about.

What should you be careful about?

The task data has limits too. Scores built on public usage reflect the tools that were measured and the period they were measured in, so a role can score low today because nobody has yet recorded the use. Your own job descriptions may also differ from the national description of the occupation.

Using data about AI use to evaluate individual workers can bring employment and privacy rules into play. The EU AI Act, for example, lists AI systems used to monitor and evaluate the performance of workers as high-risk in Annex III. This is general information, not legal advice.

The map does not settle any decision by itself. It tells a finance leader and an HR leader which roles to ask about first, and it gives both of them the same picture to ask from.