How AI Transforms Work: Anthropic Research 2025

Introduction

Anthropic has published groundbreaking research examining how artificial intelligence is transforming work within their own organization. In August 2025, the company surveyed 132 engineers and researchers, conducted 53 in-depth qualitative interviews, and analyzed 200,000 internal Claude Code usage transcripts to understand how AI use is changing the nature of software development work.

The research reveals a workplace experiencing significant transformations: engineers are achieving substantial productivity gains, becoming more "full-stack" (able to work across domains beyond their normal expertise), accelerating learning and iteration speed, and tackling previously-neglected tasks. However, these changes also raise important questions about maintaining technical expertise, preserving meaningful collaboration, and preparing for an uncertain future.

This study represents one of the most comprehensive examinations of AI's impact on software engineering work, conducted by a company that builds AI systems and has early access to cutting-edge tools. While Anthropic recognizes this represents a privileged position, the findings may serve as an instructive harbinger of broader societal transformation as AI capabilities continue to advance.

Research Methodology

Survey and Interview Approach

Anthropic's research combined multiple methodologies to capture both quantitative metrics and qualitative insights:

Survey Data:

132 engineers and researchers surveyed across diverse teams
Questions about Claude usage patterns, productivity impacts, and work delegation
Self-reported metrics on time spent and output volume across task categories

Qualitative Interviews:

53 in-depth interviews with engineers and researchers
Explored day-to-day experiences, concerns, and strategic thinking about AI use
Captured nuanced perspectives on skill development, collaboration, and career evolution

Usage Data Analysis:

Analysis of 200,000 Claude Code transcripts from February and August 2025
Privacy-preserving analysis tool to study actual usage patterns
Classification of tasks, complexity levels, and autonomy metrics

Research Limitations

The study acknowledges several important limitations:

Selection Bias: Respondents may have been more engaged with Claude or had stronger opinions
Social Desirability Bias: Responses may have been influenced by being Anthropic employees
Self-Reported Productivity: Productivity is difficult to measure precisely, and self-reports should be interpreted with caution
Privileged Position: Anthropic engineers have early access to cutting-edge tools and work in a relatively stable field
Rapid Evolution: Research was conducted in August 2025 with Claude Sonnet 4 and Claude Opus 4; patterns may have already shifted

Despite these limitations, the research provides valuable insights into how AI is transforming software engineering work at scale.

Key Findings: Productivity and Usage

Dramatic Productivity Increases

The survey reveals substantial productivity gains from AI use:

Usage and Productivity Metrics:

Current Usage: Engineers use Claude in 59% of their daily work (up from 28% 12 months ago)
Productivity Boost: Average productivity increase of 50% (up from 20% a year ago)
Power Users: 14% of respondents report productivity increases of more than 100%
Year-over-Year Growth: More than 2x increase in both usage and productivity metrics in one year

These self-reported productivity gains roughly correlate with a 67% increase in merged pull requests per engineer per day when Anthropic adopted Claude Code across their Engineering organization.

Time Savings vs. Output Volume

An interesting pattern emerges when examining how Claude affects time spent versus output volume:

Time Spent:

Most engineers report slightly less time per task category
However, some engineers spend significantly more time on Claude-assisted tasks
Reasons include debugging AI-generated code, understanding code they didn't write, and doing more thorough testing

Output Volume:

Larger net increase in output volume across all task categories
Engineers can spend slightly less time on debugging overall while producing much more debugging output
Productivity gains appear to come primarily through greater output volume rather than just time savings

Most Common Use Cases

Engineers use Claude most frequently for specific types of tasks:

Daily Usage Patterns:

Debugging: 55% of engineers use Claude daily for fixing code errors
Code Understanding: 42% use Claude daily to understand existing codebases
Implementing New Features: 37% use Claude daily for feature development
Refactoring: Regular use for restructuring existing code
Less Common: High-level design/planning, data science, and front-end development

This distribution aligns with the Claude Code usage data, showing that engineers primarily use AI for tasks that are easily verifiable and well-defined.

New Types of Work Enabled by AI

Work That Wouldn't Have Been Done

One of the most striking findings is that 27% of Claude-assisted work wouldn't have been done without AI. Engineers cited several categories of new work:

New Work Categories:

Scaling Projects: Expanding projects that wouldn't have been cost-effective manually
Nice-to-Haves: Building interactive data dashboards and quality-of-life tools
Documentation and Testing: Useful but tedious work that was previously deprioritized
Exploratory Work: Testing multiple approaches simultaneously that wouldn't be cost-effective manually
Papercut Fixes: 8.6% of Claude Code tasks involve fixing minor issues that improve quality of life

One researcher explained the value of parallel exploration:

"People tend to think about super capable models as a single instance, like getting a faster car. But having a million horses… allows you to test a bunch of different ideas… It's exciting and more creative when you have that extra breadth to explore."

Becoming More "Full-Stack"

Engineers report expanding their capabilities into areas outside their core expertise:

Capability Expansion:

Backend engineers building UIs with Claude's help
Researchers creating visualizations they couldn't have built before
Security engineers analyzing unfamiliar code
Alignment & Safety teams building front-end visualizations

One backend engineer described building a complex UI: "It did a way better job than I ever would've. I would not have been able to do it, definitely not on time... [The designers] were like 'wait, you did this?' I said 'No, Claude did this - I just prompted it.'"

This capability expansion enables tighter feedback loops and faster learning, with processes that previously took weeks becoming "a couple hour working session" with colleagues present for live feedback.

AI Delegation Strategies

What Engineers Delegate to Claude

Engineers have developed clear strategies for what to delegate to AI:

Tasks Typically Delegated:

Outside Context and Low Complexity: Tasks where engineers have low context but overall complexity is low
Easily Verifiable: Tasks where validation effort isn't large compared to creation effort
Well-Defined or Self-Contained: Subcomponents sufficiently decoupled from the rest of the project
Code Quality Not Critical: Throwaway debugging or research code
Repetitive or Boring: Tasks engineers wouldn't enjoy doing themselves (44% of Claude-assisted work on average)
Faster to Prompt Than Execute: Tasks that would take less than 10 minutes might not be worth the "cold start" problem

Tasks Kept for Humans:

High-level or strategic thinking
Design decisions requiring organizational context or "taste"
Tasks where engineers have deep expertise and can verify outputs effectively
Tasks that are exciting or enjoyable to do manually

Trust Progression

Many engineers described a progression in their Claude usage, similar to adopting other technologies:

Trust Development:

Initial Stage: Using Claude for basic questions about unfamiliar technologies
Intermediate Stage: Using Claude for tasks they mostly know, but need help with specific parts
Advanced Stage: Using Claude for daily work, trusting it to consider options and make good decisions

One engineer compared it to Google Maps: "In the beginning I would use [Google Maps] only for routes I didn't know... Today I use Google Maps all the time, even for my daily commute. If it says to take a different way I do, and just trust that it considered all options... I use Claude Code in a similar way today."

Full Delegation: Still Limited

Despite frequent use, more than half of engineers said they can "fully delegate" only 0-20% of their work to Claude. Engineers described working actively and iteratively with Claude, validating its outputs—particularly for complex tasks or high-stakes areas where code quality standards are critical.

This suggests that engineers tend to collaborate closely with Claude and check its work rather than handing off tasks without verification, setting a high bar for what counts as "fully delegated."

Skill Transformations and Concerns

New Capabilities and Skill Expansion

Engineers report significant capability expansion:

Positive Developments:

Becoming more "full-stack" and able to work across domains
Faster prototyping and iteration
Reduced "activation energy" for starting new projects
Ability to parallelize work and explore multiple approaches
Junior engineers becoming more productive and bold with project types

One engineer noted: "The tools are definitely making junior engineers more productive and more bold with the types of projects they will take on."

Concerns About Skill Atrophy

However, some engineers express concerns about "skills atrophying as [they] delegate more":

Atrophy Concerns:

Loss of Incidental Learning: Missing the learning that happens during manual problem-solving
Reduced System Understanding: Less time reading docs and code that builds understanding of how systems work
Supervision Paradox: Effectively supervising Claude requires the very coding skills that may atrophy from AI overuse
Junior Engineer Risk: More experienced engineers worry less because they "know what the answer should be or should look like"

One engineer explained: "If you were to go out and debug a hard issue yourself, you're going to spend time reading docs and code that isn't directly useful for solving your problem—but this entire time you're building a model of how the system works. There's a lot less of that going on because Claude can just get you to the problem right away."

The Future of Coding Skills

Engineers are divided on whether skill atrophy matters:

Optimistic Perspective:

Software engineering has moved to higher levels of abstraction before (from assembly to high-level languages)
Perhaps we're moving to "English as a programming language"
Focus should be on higher-level concepts and patterns rather than low-level operations

Pragmatic Perspective:

Some skills are atrophying, but they could come back if needed
Only less-important skills are being lost (like making charts)
Critical coding skills are still maintained

Challenging the Premise:

One engineer noted: "The 'getting rusty' framing relies on an assumption that coding will someday go back to the way it was pre-Claude 3.5. And I don't think it will."

Changing Social Dynamics

Claude as First Stop for Questions

One prominent theme is that Claude has become the first stop for questions that once went to colleagues:

Changing Patterns:

Engineers ask way more questions now, but 80-90% go to Claude instead of colleagues
Claude handles routine inquiries, leaving colleagues for complex, strategic, or context-heavy issues
Some report reduced dependence on teams by 80%, but the remaining 20% is crucial

Impact on Collaboration:

About Half: Report unchanged team collaboration patterns
Others: Experience less interaction with colleagues ("I work way more with Claude than with any of my colleagues")
Mentorship Concerns: Senior engineers note that "more junior people don't come to me with questions as often"

One senior engineer expressed mixed feelings: "It's been sad that more junior people don't come to me with questions as often, though they definitely get their questions answered more effectively and learn faster."

Reduced Social Friction vs. Lost Connection

Engineers have different perspectives on these changes:

Positive Aspects:

Reduced social friction ("I don't feel bad about taking my colleague's time")
More efficient question-answering
Faster learning for junior engineers

Concerns:

Loss of mentorship opportunities
Reduced in-person collaboration
Missing the enjoyment of working with people

One engineer noted: "I like working with people and it is sad that I 'need' them less now."

Career Evolution and Uncertainty

Role Transformation: From Coding to Managing AI

Many engineers describe their role shifting from writing code to managing AIs:

Role Changes:

Engineers increasingly see themselves as "managers of AI agents"
Some "constantly have at least a few [Claude] instances running"
One person estimated their work has shifted "70%+ to being a code reviewer/reviser rather than a net-new code writer"
Another saw "taking accountability for the work of 1, 5, or 100 Claudes" as part of their future role

Career Uncertainty

In the longer term, career uncertainty is widespread:

Uncertainty Themes:

Engineers say it's "hard to say" what their careers might look like in a few years
Some express conflict between short-term optimism and long-term uncertainty
One engineer: "I feel optimistic in the short term but in the long term I think AI will end up doing everything and make me and many others irrelevant"
Another: "It kind of feels like I'm coming to work every day to put myself out of a job"

Adaptation Strategies:

Specialization: Developing skills to meaningfully review AI's work
Interpersonal Focus: Anticipating more time on consensus-building and strategic work
Career Development: Using Claude for feedback on work and leadership skills
Flexibility: Emphasizing adaptability as the key skill

One team lead said: "Nobody knows what's going to happen… the important thing is to just be really adaptable."

Claude Code Usage Trends

Increasing Autonomy and Complexity

Analysis of 200,000 Claude Code transcripts reveals significant changes over six months:

Autonomy Metrics:

Consecutive Tool Calls: Increased by 116% (from 9.8 to 21.2 actions before needing human input)
Human Turns: Decreased by 33% (from 6.2 to 4.1 per transcript)
Task Complexity: Increased from 3.2 to 3.8 on a 5-point scale (where 5 is expert-level tasks requiring weeks/months)

Task Distribution Changes:

Implementing New Features: Increased from 14.3% to 36.9% of usage
Code Design/Planning: Increased from 1.0% to 9.9% of usage
Papercut Fixes: 8.6% of current tasks involve quality-of-life improvements

These trends suggest that Claude is becoming better at handling complex tasks autonomously, requiring less human oversight while tackling more sophisticated work.

Team-Specific Patterns

Different teams use Claude in different ways:

Notable Patterns:

Pre-training Team: 54.6% of usage for building new features (running extra experiments)
Alignment & Safety / Post-training: 7.5% and 7.4% for front-end development (creating data visualizations)
Security Team: 48.9% for code understanding (analyzing security implications)
Non-technical Employees: 51.5% for debugging, 12.7% for data science

These patterns demonstrate capability expansion: teams using Claude for tasks outside their core expertise, enabling everyone to become more "full-stack" in their work.

Looking Forward: Anthropic's Response

Internal Steps

Anthropic is taking several steps to address the opportunities and challenges raised:

Current Initiatives:

Talking to engineers, researchers, and leadership about collaboration, professional development, and best practices
Examining how teams come together and collaborate
Establishing best practices for AI-augmented work (guided by their AI fluency framework)
Expanding research beyond engineers to understand AI transformation across roles

Future Plans:

Supporting external organizations like CodePath as they adapt computer science curricula
Considering structural approaches like new pathways for role evolution or reskilling
Sharing more concrete plans in 2026 as thinking matures

Anthropic as a Laboratory

Anthropic positions itself as "a laboratory for responsible workplace transition," wanting to "not just study how AI transforms work, but also experiment with how to navigate that transformation thoughtfully, starting with ourselves first."

Conclusion

Anthropic's research provides one of the most comprehensive views of how AI is transforming software engineering work. The findings reveal a workplace experiencing dramatic productivity gains, capability expansion, and role transformation, alongside genuine uncertainty about the future.

Key Takeaways:

Productivity Gains: Engineers report 50% productivity increases on average, with usage in 59% of daily work
New Work Enabled: 27% of Claude-assisted work wouldn't have been done otherwise, enabling scaling, exploration, and quality-of-life improvements
Capability Expansion: Engineers are becoming more "full-stack," working across domains beyond their core expertise
Limited Full Delegation: Despite frequent use, most work involves active collaboration and validation
Skill Concerns: Engineers worry about skill atrophy while also recognizing that software engineering may be moving to higher levels of abstraction
Social Dynamics: Claude has become the first stop for questions, potentially reducing mentorship opportunities while increasing efficiency
Career Uncertainty: Widespread uncertainty about the long-term trajectory of software engineering, with emphasis on adaptability

What This Means:

The research suggests that AI is fundamentally changing the nature of software engineering work, not just accelerating existing patterns. Engineers are shifting from writing code to managing AI systems, expanding their capabilities while potentially losing some hands-on skills, and navigating new social dynamics in the workplace.

While Anthropic's position as an AI company with early access to cutting-edge tools represents a privileged case study, these findings may serve as a harbinger of broader transformation across industries as AI capabilities continue to advance.

The research raises important questions about how organizations can thoughtfully navigate these changes, support professional development, maintain technical expertise, and preserve meaningful collaboration in an AI-augmented workplace.

Sources

Interested in learning more about AI's impact on work and productivity? Explore our AI fundamentals courses, check out our glossary of AI terms, or discover the latest AI models and AI tools in our comprehensive catalog.