Anthropic Employees Use Claude for 60% of Work, Report 50% Productivity Gain — The Metric That Actually Matters
Anthropic's internal research: 60% of work uses Claude, 50% self-reported productivity gain, 2-3x increase year-over-year. But the most significant finding isn't the speed number — it's that 27% of AI-assisted work is work that wouldn't have been done at all without AI.
Anthropic published internal research this week on how its own employees use Claude for work. The headline numbers: 60% of work tasks involve Claude, 50% self-reported productivity improvement, and a 2-3x increase in AI tool usage compared to the previous year.
These numbers deserve careful reading. Some of what they reveal about AI-assisted productivity is counterintuitive.
The 50% Number and Why It’s Slippery
Self-reported productivity gains are notoriously unreliable. Workers consistently overestimate their own productivity improvements, especially with tools they’re enthusiastic about. The 50% figure reflects how employees feel about their productivity, not an objective measurement of throughput or output quality.
Anthropic acknowledges this explicitly in the research. They note that AI tool trustworthiness is rated at 29-46% by users — meaning more than half of employees don’t fully trust the AI-generated outputs without verification. When workers are verifying AI outputs before using them, the “productivity gain” includes verification time that doesn’t appear in the self-assessment.
METR’s independent research provides a useful counter-anchor. Their 2025 study found that developers using AI coding tools completed tasks 19% slower on average than unassisted developers. Their 2026 update is more optimistic but still shows that measured productivity gains are significantly smaller than self-reported ones. The gap between perception and measurement is large and consistent across studies.
The 50% figure is signal about employee experience and satisfaction, which matters for retention and morale. It’s not a validated measure of actual output increase.
The Finding That Actually Matters: 27% New Work
Buried beneath the headline numbers is the most significant data point in the research: 27% of AI-assisted work described by employees is work that “wouldn’t have been done without AI.”
This is the metric that changes the productivity conversation.
The standard frame for AI productivity is efficiency — the same work done faster. If you write documentation in 2 hours that used to take 4, that’s 2 hours saved. The efficiency frame is real but has a ceiling: you can save at most 100% of the time on any given task.
The 27% figure represents something different: expansion of what’s possible. Projects explored because AI made them feasible. Experiments run because AI lowered the marginal cost of trying. Questions investigated because AI made research accessible without specialist knowledge. Automations built because AI made the implementation tractable for non-specialists.
This is what economists call capability expansion rather than efficiency gain. It’s the difference between doing existing work faster and doing work that previously wasn’t economically viable. Historically, capability expansion from new technologies has had larger economic effects than efficiency gains, precisely because it expands the frontier of what’s attempted rather than just optimizing existing activities.
For individual developers and knowledge workers, the practical implication is direct: the teams and individuals capturing the most value from AI tools right now are not primarily the ones who moved fastest on existing tasks. They’re the ones who started new projects, conducted more experiments, and took on work previously considered out of scope.
What 60% AI Usage Actually Means at Anthropic
The 60% figure — 60% of work tasks involving Claude — reflects both the distribution of task types at an AI company and the degree to which employees have built AI into their core workflows.
At Anthropic specifically, the employee base skews heavily toward research, engineering, and knowledge-intensive roles where AI tool utility is high. This isn’t representative of most workplaces. But it’s a data point on what deep AI integration looks like in practice at a company where the tools are developed and constantly improved.
The 2-3x year-over-year increase in usage is more broadly meaningful than the absolute level. It suggests that AI tool integration in knowledge work isn’t plateauing — it’s still in a phase of rapid adoption expansion even among already-sophisticated users.
The 84% Developer Adoption Number
Separately, the research cites industry data showing 84% of developers now use AI tools, with AI writing 41% of all new code. These figures (likely sourced from GitHub Copilot and similar surveys) are consistent with public data from Stack Overflow and developer surveys in early 2026.
41% of new code being AI-written is a number worth sitting with. In 2023, that figure was in single digits. The shift happened fast enough that many development teams haven’t updated their code review practices, quality standards, or security auditing procedures to account for the changed risk profile of AI-generated code.
AI-generated code has different failure characteristics than human-written code. It’s often syntactically clean and superficially correct but can contain subtle logic errors, security vulnerabilities from training data patterns, or integration issues that aren’t visible in isolation. Review processes calibrated for human-authored code may not catch these failure modes reliably.
The Trust Gap
The 29-46% trustworthiness rating from employees is perhaps the most important operational finding in the research. Employees use AI for 60% of work but don’t fully trust the output in at least half of cases. The workflow isn’t “AI replaces human judgment” — it’s “human reviews AI-generated starting point.”
This is actually the correct workflow for most knowledge work in 2026. But it means that productivity calculations need to include the verification layer. A task that AI drafts in 5 minutes but requires 15 minutes of careful human verification isn’t a 5-minute task — it’s a 20-minute task, which may or may not be faster than the unassisted alternative depending on what the unassisted baseline was.
The companies and individuals who will see sustainable productivity gains from AI tools are the ones who invest in developing accurate mental models of where AI is reliable versus where it requires heavy verification. Treating AI output as uniformly trustworthy or uniformly suspect are both optimized-wrong strategies.
Source: Anthropic — How AI Is Transforming Work at Anthropic