Precision or Automation? The Value–Population Ratio and the Demand for AI across Industries

Some industries employ a great many people and create modest value per person. Others employ few and create a great deal. In the United States in 2025, food services employed 12.46 million people and securities firms 1.18 million. Yet each securities job produced 2.40 times the private economy's average value added, and each food-service job 0.34 times.

We wanted to know whether that gap shapes the kind of AI an industry needs. One intuition says that where each task is valuable and errors are expensive, as in law and finance, the most accurate model is worth its price. Another says that where tasks are many and each is worth little, as in retail, logistics and much of manufacturing, AI pays by doing a large volume of work cheaply and reliably, and the smartest model may be the wrong purchase.

To test this, we combined the Anthropic Economic Index with the US national accounts, the occupational make-up of every industry and a Census survey of firms, and built a map of 65 private industries covering 146.8 million US jobs.

One ratio for every industry

The ratio is simple: take an industry's share of the private economy's value added and divide it by its share of jobs. A value above one means each worker produces more than average; below one, the industry employs more people than its value added would suggest.

Two thirds of private-sector jobs sit in industries below the average: social assistance, restaurants, retail, health care, education, administrative services and construction. The knowledge-intensive industries, law, software, data and finance, hold 8.6% of jobs and 18.5% of value added.

Manufacturing is not where intuition puts it. Its ratio is 1.34 in the United States and between 1.08 and 1.23 in Germany, Viet Nam, China and India. What sets manufacturing apart is not low value per worker but its sheer headcount (155 million workers in China) and the physical nature of its work.

A high ratio also has two sources. Securities and software are high because their people are well paid; oil and gas, refining and real estate are high because they are capital-intensive. Splitting the ratio into these two parts turns out to matter for AI.

Where AI is used

We measure AI use per worker: an industry's share of AI use divided by its share of jobs. It ranges 149-fold, from 9.8 in publishing, including software, to 0.07 in restaurants.

Use grows in proportion to value added per worker: the elasticity is 1.16, statistically indistinguishable from one. With pay per worker it is steeper, 1.79, and once pay is accounted for, capital intensity shows no detectable association. AI goes where human work time is expensive, not where value added happens to be high.

Much of that link runs through what the work is. Holding fixed the share of an industry's tasks that language models could speed up, the elasticity with pay falls to 0.67. Well-paid industries use more AI largely because their occupations do the kind of work language models are good at.

Firm surveys point the same way. In the Census Bureau's surveys from July to September 2026, 45% of information-sector firms had used AI in the previous two weeks, against 9% in accommodation and food and 8% in mining. Across sectors, adoption goes with pay per worker, not with capital intensity.

What drives compute and failure

The Economic Index describes each task: how long it would take a person alone, whether the user hands it off or works through it with the model, and, for API traffic, how much compute it consumed and whether it succeeded. Pricing each task at its occupation's wage gives the value of the task.

Longer tasks consume more compute and fail more often. Cost rises with the human time a task takes (elasticity 0.48) but not with the wage (0.00), and success falls from 75% for the shortest fifth of tasks to 65% for the longest. Better-paid work does not buy more compute per task; it brings longer tasks. How people use AI also varies little with the value of a task once each occupation is weighted by how much it uses AI. Software is the exception: its tasks are valuable yet 64% automated, because code can be tested and cheap verification substitutes for model accuracy.

The price of accuracy

A simple model shows why value should matter. A task is worth $v$. A model completes it correctly with probability $q$ at cost $c$; an undetected error loses the task's value and a further share $\kappa$ of it. The expected net value of using the model is

The price worth paying for one more point of accuracy is $(1+\kappa )v$: proportional to the value of the task. Comparing a frontier model with an efficient one, the frontier model wins for every task worth more than

At September 2026 list prices this threshold is about $4.05 per task, below the value of every task we observe: even in restaurants the average task brought to AI is worth $46. For the long tasks people bring to AI today, accuracy dominates everywhere. The cheaper model wins only on short tasks: at restaurant wages, tasks shorter than 806 seconds at list prices, or about 4.1 minutes if cost scales with task length as observed. Falling prices shift both bounds down in proportion, so cheap models reach ever shorter tasks.

Two pools of demand

Labour-dense industries hold most of the jobs and most of the work that language models could speed up, but less of the use.

The precision pool is concentrated and visible in the data: few workers, long and valuable tasks, heavy use today. The automation pool is spread across two thirds of the workforce and is mostly not yet in the data: many short tasks, much of them physical or administrative, where price, reliability and integration into daily operations decide adoption. For manufacturing and other labour-dense industries, the opportunity is inexpensive models embedded in workflows, with verification built in, rather than the most capable model in a chat window. Whether that demand materialises is the open question; measuring short tasks by industry would answer it.

How we did it

We divided the US private economy into 65 industries using value added from the Bureau of Economic Analysis and the occupational make-up of each industry from the Bureau of Labor Statistics, including the self-employed. We mapped the Anthropic Economic Index's occupation-level use (April and May 2026) and task-level API data (one week of February 2026) to industries through that occupational mix, and checked the results against Census firm surveys, Eurostat data for 31 European economies and World Bank and ILO data for 174 economies. We wrote down our hypotheses before joining the data and report each one, with corrections for multiple testing. The data are cross-sectional and come from one AI provider, so the results are associations. The full paper sets out the model, the data construction and the robustness checks.

Yanming Guo, Haixin Wang, Yanjun Lu · AIDC Research

@techreport{guo2026industry,
  title       = {Precision or Automation? The Value–Population Ratio and the Demand for AI across Industries},
  author      = {Guo, Yanming and Wang, Haixin and Lu, Yanjun},
  institution = {AIDC Research},
  year        = {2026},
  type        = {Research paper},
  url         = {https://www.ai-dc.ai/research/industry-ai-demand}
}