How AI Is Transforming Work at Anthropic

Anthropic has conducted an internal study on its own engineers and researchers to determine how AI tools, specifically Claude and Claude Code, are reshaping software development. The findings indicate that AI is radically changing the nature of work by increasing output volume and broadening individual technical capabilities, though it introduces new tensions regarding technical expertise and professional mentorship.

Productivity Gains and Output Volume

AI integration has led to significant increases in both the frequency of use and self-reported productivity.

Increased Usage: Employees reported using Claude in 59% of their work, up from 28% a year prior.
Productivity Boost: On average, employees reported a 50% productivity gain, a 2-3x increase from the previous year. This is corroborated by a 67% increase in merged pull requests per engineer per day.
Output vs. Time: Productivity gains are primarily driven by a substantial increase in output volume across almost all task categories, rather than a simple decrease in time spent per task.
New Work: Approximately 27% of Claude-assisted work consists of tasks that would not have been performed otherwise, such as scaling projects, creating interactive data dashboards, and fixing "papercuts" (minor quality-of-life improvements).

AI Delegation Strategies

Engineers are developing specific intuitions for which tasks to delegate to AI. Most employees report they can "fully delegate" only 0-20% of their work, indicating that AI is used as a collaborator requiring active supervision rather than a total hand-off.

Criteria for Delegation

Users tend to delegate tasks that meet the following criteria:

Low Complexity/Low Context: Tasks where the user has low familiarity but the overall complexity is low (e.g., basic Linux or Git operations).
Easily Verifiable: Tasks where the effort to validate the output is small compared to the effort to create it.
Well-Defined: Decoupled subcomponents of a project.
Low Stakes: "Throwaway" research or debug code where high quality is not critical.
Repetitive: Boring tasks that create psychological resistance to starting.

The Trust Progression

Many engineers describe a "trust progression" similar to adopting GPS technology: starting with unfamiliar routes (basic SQL) and eventually trusting the tool for daily commutes (complex Python code), while still maintaining the ability to verify outputs.

Skill Transformation and the "Paradox of Supervision"

AI is enabling engineers to become "full-stack" by allowing them to work in domains outside their core expertise. For example, backend engineers are building UIs, and researchers are creating front-end visualizations.

The Risk of Skill Atrophy

Despite the breadth of new capabilities, there is significant concern regarding the atrophy of deep technical skills.

Loss of Collateral Learning: Manual debugging often forces engineers to read documentation and unrelated code, building a mental model of the system. AI-driven solutions can bypass this process.
The Paradox of Supervision: Effective AI use requires high-level supervision, but the skills needed to supervise AI are the same manual coding skills that may atrophy through AI overuse.

Shifting Abstractions

Some engineers view this as a natural evolution of software engineering, comparing it to the move from assembly language to high-level languages. In this view, "vibe coding" (using English as a programming language) allows engineers to focus on end-products and high-level patterns rather than low-level implementation details.

Changes in Workplace Social Dynamics

Claude has become the primary point of contact for technical questions, which has altered traditional mentorship and collaboration.

Reduced Peer Dependence: Some employees report that 80-90% of their questions now go to Claude instead of colleagues.
Mentorship Impact: Senior engineers have noted that junior staff come to them with questions less frequently, as Claude provides immediate coaching and answers.
Social Friction: While some appreciate the reduced social friction of not "taking a colleague's time," others miss the interpersonal connection and collaborative problem-solving of the pre-AI era.

Claude Code Usage Trends

Analysis of 200,000 internal transcripts from February to August 2025 shows that AI is handling increasingly complex tasks with greater autonomy.

Increased Autonomy: The maximum number of consecutive tool calls Claude Code makes without human intervention increased by 116% (from 9.8 to 21.2).
Reduced Human Input: The average number of human turns per transcript decreased by 33% (from 6.2 to 4.1).
Task Complexity: Average task complexity rose from 3.2 to 3.8 on a 5-point scale.
Shift in Task Type: There was a marked increase in using Claude for implementing new features (14.3% to 36.9%) and code design/planning (1.0% to 9.9%).

Career Evolution and Uncertainty

Engineers are increasingly viewing their roles as "managers of AI agents" rather than net-new code writers. This shift has led to widespread career uncertainty, with some engineers expressing a conflict between short-term productivity gains and long-term fears of professional irrelevance. Adaptation strategies include further specialization in AI review and a shift toward more strategic, interpersonal, and consensus-building work.