Most Claude Users Expect AI to Do Most of Their Work Within a Year

More than a third of Claude users expect AI to do most of their work within a year, according to Anthropic’s June 2026 Economic Index. The survey of 9,700 Claude users is the first time a major AI lab has asked its users directly how AI is reshaping their jobs, and the results split in an instructive way: the people who delegate the most to Claude are the most upbeat about their careers, while entry-level workers carry the most anxiety about the same technology.

Anthropic released its third Economic Index on June 26, framed around when people use Claude, what they produce with it, and how they perceive AI’s impact on their work. The Cadences report introduces hourly sampling and a new output classifier, and adds the first wave of the Economic Index Survey methodology page. For the first time, stated beliefs about AI capability are linked to actual Claude session data, not just inferred from behavior.

The First Survey From a Major AI Lab

Anthropic has been publishing the Economic Index since 2025, drawing on privacy-preserving data analysis to track how Claude is being used across the economy. The Cadences edition, published as the June 2026 Cadences report, marks the first time Anthropic has gone beyond behavioral telemetry. The new survey launched in April 2026 and ran through mid-May to early June, with Anthropic Interviewer as the tool.

Each respondent’s answers were cross-checked against up to 20 of their real sessions, sampled across Claude.ai, Cowork, and Claude Code. That linkage, run through Anthropic’s privacy-preserving CLIO system, lets researchers compare stated beliefs against observed delegation patterns, task complexity, and how much judgment Claude was given to exercise independently. Respondents with fewer than five sessions were excluded to cut sampling noise.

The shift matters because prior Economic Index reports inferred AI’s economic footprint from behavior alone. The June 2026 edition adds the qualitative layer that surveys can capture, from workplace changes users have noticed to expectations about the next twelve months. The data also carries a known limitation: every respondent is, by definition, a Claude user. Workers who have not adopted frontier AI tools are not in the sample at all.

The Optimism Splits by Who You Are

Anthropic measured optimism across six dimensions of work: pay, job security, ability to find a new job, meaning, autonomy, and human interaction. More than a third of respondents expected AI to take on a larger share of their work within a year. Users whose sessions leaned toward the most automated mode of interaction reported the most positive expected impacts on every one of those dimensions, a relationship that holds after controlling for how long someone has used Claude.

• 9,700 users surveyed, each linked to their real session data
• 93% of Claude conversations produce a recognizable output
• Roughly half of respondents say AI already handles 50% or more of their work
• 4% say AI could perform their entire job today
• 26% expect AI to take on the majority of their tasks within 12 months

The same data cuts the other way on early-career workers. Among respondents who could be classified by career stage, those in their first year of work reported both the highest share of AI-capable tasks and the most concern about displacement. Workers with at least 15 years of experience estimated that AI could perform about 10 percentage points fewer of their tasks than first-year workers did. The pattern lines up with the kinds of tasks AI handles most thoroughly: discrete, deliverable-producing work, the kind that fills entry-level roles.

Anxious respondents did not keep the worry to themselves. More than a third said it was likely or very likely that job responsibilities would change significantly over the next year. About 10% believed losing their own job was likely or very likely. Respondents were more concerned about job losses among their junior colleagues than for themselves, and over a third rated the probability of a junior colleague losing their job in the next year at over 60%.

Productivity gains were the bright spot in the survey. Around 86% of respondents said AI improved the speed of their work, 82% reported improvements in the scope of their work, and 69% said it improved quality. Some 68% said they were learning more with AI, and 57% felt AI had increased the value of their skills. Heavy delegators reported learning at the same rate as everyone else, undercutting the concern that offloading thinking causes skill atrophy. The internal data is detailed on the live Economic Index data dashboard.

Tasks People Hand to Claude

Anthropic introduced a new artifact classifier this round, labeling the primary output of each conversation across more than 30 categories. The classifier found that 93% of Claude conversations produce a recognizable artifact: a document, an explanation, a script, an app. The most common outputs were explanations at 17% of conversations, documents and reports at 15%, and guidance at 11%. Conversational outputs and written deliverables each accounted for about a third of conversations, with code and technical work making up about a sixth. Building an app uses about three times the tokens of the median conversation; a typical explanation uses about a fifth.

What an output is used for varies sharply by category. More than 80% of conversations producing creative writing, guidance, and recipes were classified as personal. The work-heavy categories sit at the other end: 82% of database query writing, 81% of blog or article writing, and 80% of marketing content creation were work-related. Some outputs split almost evenly between personal and work, including plans or strategies at 44% work and 49% personal, and translation at 42% work and 44% personal.

The cost of each conversation tracks the value of the work it replaces. Conversations mapped to higher-wage occupations consume more compute. Marketing managers earn roughly twice as much as editors, $80 per hour versus $37, and their Claude conversations use about 2.5 times as many tokens. The pattern holds beyond that comparison: pharmacists earn nearly three times what statistical assistants do, $68 versus $24, yet pharmacist conversations use only about a twentieth as many tokens. Compute is not a clean proxy for wage, but the direction is consistent across roles, and the same direction shows up in real-world deployments like the recent Janus Henderson Claude PRISM rollout for investment research.

How a Day With Claude Looks

Anthropic upgraded its data pipeline to sample hourly instead of pulling weekly snapshots, a shift that exposes how Claude usage tracks the rhythms of ordinary life. The new view finds that people most often ask for sleep advice around 5 a.m. and for recipes around 6 p.m. local time. Business correspondence peaks at 10 to 11 a.m., and people ask for the news around 7 a.m.

• Recipe requests spike at 6 p.m. and run 2.3x the daily average
• Sleep advice clusters in the few hours before dawn, peaking around 5 a.m.
• News requests rise in the morning, peaking at 7 a.m. local time
• Business email drafting traces the workday arc, peaking 10 to 11 a.m.
• Media recommendations cluster in the evening hours

Daily and weekly patterns both show up. Personal use spikes from around 35% on weekdays to just under 50% on weekends, the largest swing in high-income countries. Tax-related conversations spiked eight times higher on April 14 than the average May day and stayed elevated through the April 15 filing deadline, then dropped sharply on April 16. Tasks tied to higher-wage occupations increased their share of total conversations on nights and weekends; tasks in the bottom two wage quartiles fell. The gap persists even when computer and math jobs are removed from the analysis.

Delegation Looks Different Across Claude Products

How much judgment users hand to Claude depends heavily on which product they are using. Anthropic measures autonomy on a 1-to-5 scale, with zero for tasks where the answer is essentially determined by the input (math, translation) and maximum for sustained sequential decisions (building apps, websites, games). Claude Code and Cowork sessions register higher autonomy than chat sessions across most output types. About two-thirds of that gap comes from the same tasks being executed with more delegation on Claude Code.

The difference is not just about model choice. Even when the underlying model is held constant, Claude Code sessions show about 0.26 points more autonomy than chat on the same task. The product itself drives how much control people hand over. Reading level moves the other direction: Claude consistently responds at about one year higher than the prompt, a gap that widens for image and graphics prompts (+2.6 years) and games (+1.9) and shrinks for audience-facing writing.

For higher-wage knowledge work, the data looks more like augmentation than substitution. Conversations tied to higher-wage jobs saw Claude produce more output per turn, 1.34x, with users engaging more, 1.53x as many turns, and extended thinking used in 34% of conversations versus 31%. The user does not step back as Claude does more; the user engages more as the stakes rise. The contrast between worker experience and broader anxiety is real, and the workers most exposed to AI’s capabilities may not be the ones whose optimism the survey measures.

Who the Survey Misses

Every respondent in the 9,700-person survey is, by definition, an active Claude user, a population that has already adopted frontier AI tools in daily workflows. Reported task coverage among this group exceeds what usage telemetry shows across the broader economy. The most exposed workers may be the ones not generating any Claude sessions at all: entry-level and administrative roles, the kind of work that does not produce the artifact trail a survey can sample. Federal Reserve Bank of Dallas research from February 2026, reported by TechTimes, found employment declines in AI-exposed sectors falling disproportionately on workers under 25, with a collapse in job-finding rates for new graduates rather than layoffs. The workers the survey cannot reach are the ones the broader labor market data flags as most exposed, a gap that long-running studies on attention spans and AI also pick up.

The survey also exposes a gender skew inside the sample itself. Women make up only 12% of the linked survey, and their automation share is 0.33 standard deviations lower than men’s after accounting for occupational differences. Women log more active time on chat, suggesting more iterative, collaborative engagement rather than single-prompt delegation. The asymmetry is a feature of the user base Anthropic has, not the technology, but it shapes whose optimism the survey can report.

Frequently Asked Questions

How many Claude users did Anthropic survey?

Anthropic’s June 2026 Economic Index draws on about 9,700 Claude users, each linked to their real session data through the privacy-preserving CLIO system. The survey launched in April 2026 and ran through mid-May to early June 2026.

What do users think AI can already do?

Roughly half of the 9,700 respondents said AI can already handle 50% or more of their work tasks, and 4% said AI could perform their entire job today. Looking ahead 12 months, 26% expected AI to take over the majority of their tasks, a figure Anthropic describes as broadly consistent across experience levels, locations, and professions.

Who is most optimistic about AI’s impact on their work?

Users whose sessions lean most toward automated, directive delegation reported the most positive expected impacts on pay, job security, ability to find work, meaning, autonomy, and human interaction. Early-career workers reported the highest share of AI-capable tasks but also the most concern about displacement.

What are people actually using Claude for?

A new artifact classifier found that 93% of Claude conversations produce a recognizable output. The most common are explanations at 17%, documents and reports at 15%, and guidance at 11%. More than 80% of creative writing, guidance, and recipe conversations are personal; 80% or more of database query writing, blogs, and marketing content are work-related.

Are entry-level workers most at risk from AI?

The survey shows early-career workers reporting both the highest share of AI-capable tasks and the greatest concern about job loss. Outside data lines up: research from the Federal Reserve Bank of Dallas found AI-exposed employment declines falling disproportionately on workers under 25, with new graduates facing a collapsed job-finding rate.