Anthropic's economic index reveals AI already does half the work… but its data has a huge blind spot

By Momentum IA · June 28, 2026.

On June 26, Anthropic published the latest edition of its Economic Index with a methodological twist that sets it apart from nearly everything published so far on AI's impact on employment. The 9,700 Claude users who answered the survey did not merely state their perceptions: each response was cross-referenced with up to 20 real sessions of their activity on Claude.ai, Cowork and Claude Code, extracted via the CLIO privacy system for the period from mid-May to early June 2026. For the first time, what someone says they delegate can be compared with what they actually delegate.

The main finding has already made the rounds of headlines: roughly half of respondents claim that AI can do 50% or more of their work tasks, and 4% say Claude could perform their entire job today. Some 26% expect AI to take on the majority of their tasks within the next twelve months, a figure Anthropic describes as «surprisingly uniform» regardless of experience, geography or sector. The tasks with the highest real-work coverage are discrete and deliverable-oriented: writing database queries (82% of conversations classified as work-related), articles and blog posts (81%) and marketing content (80%).

So far, the headline. The analytically interesting part begins when one examines whom this methodology captures —and whom it does not.

**The survey's structural blind spot**

The sample is made up of workers who are already active Claude users. That is not an avoidable flaw: it is the population Anthropic can measure. But it has a brutal consequence for interpreting the data. The workers most exposed to AI displacement —those in entry-level positions that automation replaces rather than augments— do not generate Claude sessions precisely because those are the jobs that are no longer being offered.

Two external studies published in 2026 put figures to what Anthropic's survey cannot see. The Federal Reserve Bank of Dallas documented in February that employment declines in AI-exposed sectors are falling disproportionately on those under 25 —not through layoffs, but through a collapse in the job-placement rate of new graduates. Goldman Sachs, in April, estimated that AI is eliminating around 16,000 net jobs a month, concentrated in entry-level and administrative roles. These are exactly the profiles that do not answer Claude surveys because they do not use Claude in a job that no longer exists.

This matters in order to avoid falling into a trap of statistical optimism: Anthropic's data faithfully reflect the experience of active users of frontier AI tools, who tend to be skilled workers in knowledge roles. They are a sample of the short-term winners of the adjustment, not of the whole.

**Augmentation vs. substitution: the line that matters most**

Anthropic's autonomy classifier —a scale from 1 to 5 that measures how much independent judgment Claude exercises— adds valuable texture. Claude Code sessions recorded higher autonomy scores than chat sessions in 26 of the 31 output types analyzed. And there is a striking correlation: conversations related to higher-paying occupations consumed significantly more compute; a marketing director's sessions used roughly 2.5 times more tokens than an editor's, in proportion roughly to the salary differential. More relevant still: the greater the autonomy delegated, the greater the human involvement, not the lesser. The model does more and the professional gets more involved, not less.

This fits the augmentation thesis for high-value knowledge work: AI amplifies the expert's capacity rather than replacing it. But the other side of that distinction is precisely the problem: AI does replace the codified, repeatable work that defines entry-level positions. And those positions are the historical rung by which one climbs toward higher-value work. If that rung disappears, the traditional mechanism of professional development is called into question.

**The optimism of heavy users: real protection or selection bias?**

The survey reveals that the heaviest Claude users —those who delegate tasks most autonomously— are the most optimistic about their professional prospects. Many believe their skills are becoming more valuable thanks to AI. It is a counterintuitive finding at first glance: the more exposure to AI, the more confidence, not less.

With the available data, Anthropic cannot disentangle whether this reflects a real protective effect of AI fluency —those who truly learn to work with Claude become more productive and harder to replace— or a selection bias: the most self-confident and adaptable workers are the ones who adopt AI tools most aggressively, regardless of what that AI is doing to the broader labor market. The distinction is not minor: it is one thing for AI adoption to protect the worker, and quite another for the workers who were already well positioned to be the ones who adopt AI.

**Young workers, at the center of the storm**

Early-career respondents reported both the highest estimate of AI task coverage and the greatest anxiety about displacement. There is direct logic to this: entry-level roles concentrate exactly the kind of discrete, concrete-deliverable tasks that AI handles best. It is in that segment where substitution is most immediate and where the anguish has the most empirical basis.

What the study also cannot resolve is whether that anxiety is calibrated or whether it too is contaminated by the same optimism bias that appears in the rest of the data —but in the opposite direction, overestimating the threat. The uncertainty in both directions is honest; the problem is that while it is being resolved, the market is already giving clear signals in the world outside the survey.

**Our read**

Anthropic's Economic Index is the methodologically most rigorous study of this kind published to date. Cross-referencing stated perceptions with real behavioral data is exactly what this field needed to climb out of the swamp of self-perception surveys. And the findings on augmentation in high-value work, the pattern of greater human involvement as the model's autonomy rises, and the uniformity of expectations of progress are genuinely useful data.

But it must be read with clear awareness of its structural limit: it measures the experience of those who have already crossed the threshold of adoption and, in many cases, have benefited from it. The 2026 labor market is simultaneously showing two realities that do not contradict each other: AI augments experts and replaces beginners. The first is more visible in AI-platform data; the second, in employment statistics.

That asymmetry is the core of the transition challenge. In the long term, the scenario emerging from increasingly capable tools —systems that amplify human intelligence rather than compete with it— points toward radically greater collective productivity, toward the real possibility of tackling problems that have been out of reach for lack of talent or time: diseases, scientific discovery, personalized education. But that horizon offers no comfort to the 2026 graduate who cannot find a first job because the entry-level position that should exist has been filled before being posted.

The question the next Economic Index should try to answer is not how much AI already does, but how the entry rung into skilled work is rebuilt once AI has eliminated it. That is the political and economic question that Claude's behavioral data cannot answer on their own.