This study by Anthropic examines how AI, and more specifically Claude Code, is transforming the work of its own software engineers. It draws on a survey of 132 engineers and researchers, 53 qualitative interviews, and analysis of internal usage data (August 2025, Claude Sonnet 4 and Opus 4 models). Its position is distinctive: observing "early adopters" within the very company developing the AI, likely representative of transformations to come elsewhere.
The figures are striking. Employees report using Claude for 60% of their work and estimate a productivity gain of about 50%, 2 to 3 times higher than the previous year. Above all, 27% of Claude-assisted work simply would not have been done otherwise: more ambitious projects, "nice-to-have" dashboards, exploratory work that is not profitable to do manually. Delegation nonetheless remains supervised: only 0 to 20% of work can be "fully delegated," with human validation still required for critical work.
The interviews reveal deep qualitative transformations. Engineers develop an intuition for delegation: first handing off verifiable, low-stakes, or tedious tasks, then progressively expanding — design and "taste" remaining human, for now. Developers become "full-stack," working competently on domains they would not previously have dared to touch. But a paradoxical concern emerges: the atrophy of the deep skills needed to write and critique code — "when producing results is so easy and fast, it becomes difficult to really take the time to learn."
Social dynamics are also evolving: Claude replaces colleagues as the first port of call for technical questions, reducing opportunities for mentorship. One senior engineer says with sadness that juniors no longer come to see him. On the career front, feelings are contradictory: short-term optimism (higher-level work, managing AI systems), long-term concern ("AI will eventually do everything and make me obsolete").
Usage data confirms growing autonomy: from 10 to 20 autonomous actions in six months, code design rising from 1% to 10% of uses, implementation of new features from 14% to 37%. 8.6% of tasks involve fixing "papercuts" previously deprioritized.
Anthropic draws internal initiatives from this — new mentorship models, maintaining deep skills — and stresses the importance of actively preparing the transition to an AI-augmented workplace, lessons transferable to other organizations.