When the board asks about the Anthropic Economic Index
Your CEO forwards a headline about Claude usage data and asks, in the subject line, "does this affect us?" You have three weeks until the board meeting, a PE sponsor who wants a workforce risk slide, and a research report in your inbox that was built from millions of anonymized AI conversations, not from your org chart. The Anthropic Economic Index is real, well-documented, and genuinely useful context. It is also not, by itself, an answer to the question your board is actually asking. Here is what the Index measures, what its numbers mean, and where the translation from macro usage data to a company-specific role list has to happen separately.
What the Anthropic Economic Index actually measures
The Anthropic Economic Index is built from aggregated, anonymized data on how people use Claude, mapped against standardized occupational task frameworks. It is a usage signal, not a prediction engine: it tells you how often a given occupation's tasks show up in real conversations with the model, not whether those tasks get automated, reduced, or eliminated inside any specific employer.
That distinction matters more than it sounds. A task appearing frequently in AI usage data tells you people are trying AI on that kind of work. It does not tell you whether your company has adopted that workflow, whether your process allows it, or whether the task in question is core to the role or a small fraction of it. The Index is descriptive of aggregate behavior across the economy. Your organization is one data point inside that aggregate, and possibly an unusual one.
The headline numbers, and what they don't say
The most-cited figures from the Anthropic Economic Index, 2025, show that 36% of occupations used Claude for at least a quarter of their tasks in early 2025, rising to 49% on a pooled basis as adoption spread. That is a meaningful jump in a short window, and it is the kind of number that belongs on a board slide as context.
Anthropic Economic Index (2025): the share of occupations using Claude for at least a quarter of their tasks rose from 36% to 49% on a pooled basis.
What it does not say is which of those occupations exist in your company, at what headcount, doing what mix of tasks. "Occupation" in this dataset is a broad, standardized category — the same category might cover a role at your company that is 90% routine data entry and a role at another company that is 90% client judgment calls, and the Index cannot see that difference. It is built for economy-wide pattern detection, not for organizational decision-making.
Task feasibility versus actual usage: the 68/3 split
A second Anthropic dataset, published in 2026, breaks usage down by feasibility rather than frequency: of the tasks people brought to Claude, 68% were tasks the model could complete fully, and 3% were tasks it could not complete at all, with the remainder falling somewhere in between.
This is a more interesting number than it first appears, and also more frequently misread. "Fully feasible" describes what a model can technically complete in a single interaction. It says nothing about whether an organization has redesigned a workflow around that capability, whether a human reviewer still checks the output, or whether the task is one-tenth of a role or the whole job. Capability and adoption are different variables, and adoption inside any one company is shaped by process, compliance requirements, change management capacity, and dozens of factors the Index was never built to capture.
This is the gap where the phrase "AI can do this task" quietly turns into "this job will be automated" in a board conversation, without anyone deciding to make that leap. It is worth naming out loud so it does not happen by accident. For a longer walkthrough of how occupational exposure indices translate technical capability into a usable score, see our explainer on the AI occupational exposure index.
Where the Index sits among other macro research
The Anthropic Economic Index is one entry in a growing field of AI-and-labor research. McKinsey has estimated that 60–70% of work hours could theoretically be automated with current and emerging technology, up from roughly 50% in earlier estimates (McKinsey, 2023). Goldman Sachs has estimated the equivalent of 300 million full-time jobs globally are exposed to generative AI in some form (Goldman Sachs, 2023). The World Economic Forum's Future of Jobs Report 2025 projects 170 million jobs created and 92 million displaced by 2030, a net churn figure, not a net-loss one (WEF, 2025).
BCG and PwC publish comparable macro barometers on AI's labor effects, and both are worth reading alongside the Anthropic Index for a fuller picture — we cover what those reports add in our BCG and PwC AI jobs reports roundup. PwC's own Global CEO Survey work offers a useful adjacent data point: in 2024, 25% of CEOs expected a headcount reduction of at least 5% attributable to AI, against 39% who expected an increase of at least 5%; by the 2025 survey, 13% of CEOs reported an actual headcount reduction tied to generative AI (16% in the most AI-exposed sectors), while 17% reported an increase (PwC, 2024; PwC, 2025). Expectation and outcome, notably, did not move in lockstep.
None of these sources — Anthropic included — produce a company-specific, role-level output. They are an alternative frame for understanding the macro trend, not a substitute for organizational analysis. For a broader synthesis of what all of this macro research adds up to for HR strategy specifically, see how AI is changing the future of work.
Why usage data can't answer your board's actual question
Your board is not asking "what share of occupations nationally use AI for a quarter of their tasks." It is asking "which of our roles carry exposure, and what should we do about the ones that do." That question requires two things the Anthropic Index was never designed to provide: a mapping from your actual role list to standardized occupational data, and a scoring method that breaks a role down into the kinds of work that vary in how automatable they are.
This is the translation layer. ONET, the U.S. Department of Labor's occupational database, provides the standardized bridge: 1,016 occupational titles covering 55,000+ jobs, each described by roughly 277 descriptors updated on a regular cycle (ONET Resource Center, 2019; U.S. DOL, 2025). A role-level exposure methodology maps each of your organization's roles to its nearest ONET occupation, then scores the underlying tasks — not the job title as a whole — across four dimensions: cognitive routine, physical routine, social/judgment, and creative, each on a 0–100 scale, weighted by each task's ONET-documented importance to the occupation.
As a worked example: if a role's tasks are 60% weighted toward cognitive-routine work scoring 80, 25% toward social/judgment work scoring 20, and 15% toward creative work scoring 15, the importance-weighted composite comes out closer to 55 than to any single dimension in isolation — a very different number than eyeballing the job title would suggest. That composite score, and where it falls relative to a company's own thresholds, is what determines whether a role lands in Monitor, Review, or Redeployment Candidate — categories describing where a role sits for human attention, never a verdict on whether a person keeps their job. For a deeper look at which occupations tend to land where and why, see which jobs are most exposed to AI.
This site incorporates information from ONET. Used under the CC BY 4.0 license. ONET is a trademark of USDOL/ETA.
Turning macro signal into a company-specific plan
The right way to use the Anthropic Economic Index is as framing, not as a finding. It tells your board that AI usage is broad and accelerating across the economy — useful for urgency, not for specificity. The specificity has to come from mapping your own roles to O*NET occupations and scoring the underlying tasks, which is a different exercise entirely from reading usage statistics, however well-sourced.
Manual versions of that mapping exercise — a consultant or an internal analyst working through O*NET by hand — are customized but slow and hard to repeat every time the org chart changes. A self-serve, task-level tool built for this exact translation gives HR teams a repeatable, mid-market-priced way to move from "here's what's happening in the economy" to "here's our department-level heatmap," without an enterprise consulting engagement. You can see how the tiers are structured on our pricing page.
If you want the Anthropic Index and reports like it delivered as ongoing context rather than a one-off read, subscribe to our newsletter — we track new macro research as it publishes and flag what's actually decision-useful for HR strategy teams, versus what's interesting but not yet actionable.