Center estimates as of September 24, 2026 · Rebuilt from activity inputs
One worker-equivalent = 1,800 human task-hours of useful output per year. Each point annualizes the work happening at that date. These are modeled task equivalents, not a census of agents, jobs eliminated, or measured net time saved. Human prompting, checking and integration still take time.
Observed usage and deployments constrain the inputs; task duration, acceptance, coverage and utilization require judgment. The arithmetic is reproducible, but the totals are not empirically established. A precise plotted point does not make the estimate precise.
| Category | Worker-equivalents |
|---|
Coders started before coding agents. Copilot was available in June 2022. The November 2022 estimate is 600,000 active users × 15 minutes of accepted human-equivalent coding per week × 48 weeks ÷ 1,800 = 4,000 worker-equivalents. The active count and time credit are assumptions, not an observed worker count. Later agent use is modeled separately; a registration, paid seat, token or line of code never counts as a whole developer.
“Robots” now means commercial mobile/service robots. It includes warehouse transport, cleaning, delivery and humanoids, while excluding conventional fixed factory arms and household devices. This is a scope expansion from the old humanoid-only line. Amazon alone reported a million deployed robots in June 2025. With conservative utilization and task-rate assumptions, this model estimates about 201,000 robot worker-equivalents at the start of 2025 and 275,000 at its end. Humanoids alone contribute only about 170 at the end of 2025 and 390 in September 2026.
The robot line includes established automation that predates generative AI. It cannot be read as labor created by reinforcement learning since 2022. Disclosed hours from Figure and Agility cover particular deployments, not all robots worldwide.
| Robot group | Deployed units (est.) | Productive hours / unit / year (est.) | Human hours / robot-hour (est.) | Worker-equivalents |
|---|
No token-to-hour or dollar-to-hour conversion. The models work from task activity. Paid seats and robot shipments constrain scale; they are not counted as full-time labor. Repeated prompts, tool calls and retries are grouped into tasks. Credit is discounted for rejected or unusable output.
The weakest inputs: global standalone API task volume, deduplicated active coding-agent users, and human-equivalent task durations. Public sources do not establish these quantities. For example, halving the 12 hours credited per frequent coding-agent user reduces Coders from 4.32M to 2.40M. Halving robot task-rate assumptions halves the robot total. These are assumption checks, not confidence intervals; the chart shows only the center.
Avoid double counting: consumer chat/research reserve 10% of work messages for coding, which is counted only in Coders. Enterprise/API knowledge work excludes consumer-channel tasks and all coding. Twenty percent of enterprise-seat output and 10% of standalone API output are allocated to Researchers; these tasks are removed from Knowledge Workers. API task counts exclude requests already represented by paid seats. Products can overlap in users; tasks should not.
Time: the first point is November 30, 2022, ChatGPT’s launch day. The endpoint is September 24, 2026. Monthly points linearly interpolate the activity inputs and then apply the formulas; they are not monthly observations. Consumer work-message anchors for June 2024 and June 2025 are measured total traffic times estimated work shares; the intervening and later activity levels are modeled. The logarithmic chart gives equal vertical space to each tenfold increase.
Comparisons: 4.32M coding task-equivalents / 40M human developers ≈ 11%. As a fraction of a hypothetical combined 44.32M workforce, that would be ≈10%. Both compare equivalent task output with human headcount, not identical productivity measures. The 40M global developers and 62M U.S. knowledge workers are the article’s rough reference benchmarks; the U.S. line is not a global denominator. Multiplying this total by an average salary would not measure GDP or revenue.
All conversion parameters and global expansions below are judgmental assumptions. The accompanying source notes distinguish disclosed observations from inferred activity. The numerical inputs are published so another reader can replace any assumption.