Atlas · Methodology contrast

The Anthropic Index measures a behavior. The Atlas measures a system.

A 55.2 percentage-point gap between what chatbot-usage telemetry can observe and what economy-wide system-architecture observation actually finds — across 22 SOC major occupation groups.

55.2 pp
National blind-spot gap (Anthropic observed vs. Atlas reality, 22 SOC major groups, March 2026)

What each methodology actually measures

The Anthropic Economic Index publishes three useful series: which occupations show up most in Claude.ai conversations, which tasks within those occupations get help, and how those patterns shift over time. It is the cleanest signal we have on how people consult AI through a chatbot interface.

That is not the same as how AI affects production. A radiologist who pastes scan summaries into Claude and asks for second opinions is in the dataset. A radiology workflow that runs contract-gated multi-model ensembles, ingests DICOM directly through agent orchestration, escalates ambiguous cases to a human reviewer, and produces a final report without a human ever opening a chat window — is not.

Massenkoff & McCrory (March 2026) formalize this gap. Below is their core decomposition extended with DaedArch's per-NAICS transformation corpus.

The four observation channels

ChannelWhat it capturesWhat it misses
Chatbot conversation (Anthropic Index)User-initiated AI consultation patternsAnything not initiated as a chat turn
Contract-gated executionTool calls under formal preconditionsConversational coaching
Agent orchestrationMulti-step automated workflowsSingle-shot human judgment
Visual scanning & structured pipelinesContinuous machine-led inspectionAnything that's still a person reading a screen

The Anthropic Index sees channel 1. The Atlas adds channels 2–4. The gap between them is the 55.2 pp number above — and it is not evenly distributed.

Where the gap is largest

Across 22 SOC major occupation groups, blind-spot magnitude varies sharply. The table below is the headline contrast (full per-SOC table available via daedarch.business.atlas_intelligence_v1 on any NAICS code):

SOC major groupAnthropic observedAtlas realityBlind-spot gap
15 — Computer & Mathematical62.4%71.2%8.8 pp
13 — Business & Financial Operations19.1%63.7%44.6 pp
11 — Management12.8%58.4%45.6 pp
43 — Office & Administrative Support8.2%76.9%68.7 pp
49 — Installation, Maintenance, Repair3.1%29.4%26.3 pp
53 — Transportation & Material Moving1.8%71.6%69.8 pp

Source: Massenkoff & McCrory (March 2026), Atlas SOC × NAICS corpus (atlas_transformations, n=23,797).

The pattern is consistent: occupations whose AI exposure routes through chatbots (computing, knowledge work) show small gaps. Occupations whose AI exposure routes through automated structured pipelines — back-office processing, transportation, logistics, maintenance — show very large gaps because the work happens without anyone typing a question.

If a methodology only measures conversations, it cannot see the categories of work that are being automated away from conversations. The Atlas exists to fix that omission.

Why this is a research question, not an advocacy one

Trellison rates methodology, not conclusions. Two consequences:

  1. The Atlas does not claim AI economic impact is "really" 55.2 points larger than Anthropic reports. It claims that two methodologies measuring different surfaces produce different numbers — which is the methodologically uncontroversial part — and provides a public, auditable channel-by-channel decomposition so anyone can replicate or refute.
  2. Anthropic's methodology is good for what it measures. The Atlas does not replace it; it extends it. The combined view is more useful than either alone.

Reproducing this analysis

The Atlas is queryable at the per-NAICS level via daedarch.business.atlas_intelligence_v1 (see the operational tool on DaedArch). Inputs: NAICS code; outputs: SOC × NAICS transformations, scenario contrast (chatbot view vs. system-architecture reality), impact metrics, tools contrast, industry packages.

Source corpus: atlas_transformations (23,797 rows), atlas_oews (82,522 rows), Massenkoff & McCrory (March 2026). Methodology disclosure and reproduction instructions: /atlas/methodology.