The landscape of American knowledge work is undergoing a quiet but profound transformation as AI integration moves from experimentation to delegation. A new representative survey reveals that 20 percent of employed US adults have already handed off at least one task to AI that was previously managed by a human coworker or contractor.

Software Development and Data Analysis Lead the Shift

A joint study by Epoch AI and polling firm Ipsos, which interviewed 1,106 employed US adults, highlights a clear hierarchy in AI adoption across different professional functions. The survey utilized ten common work tasks based on US Department of Labor data to track how generative AI is being woven into the fabric of daily operations.

Software development and computer systems design emerged as the primary drivers of AI integration, with 57 percent of workers in this domain utilizing AI tools. Data analysis followed closely at 46 percent, while the task of reading work documents saw a 39 percent adoption rate. Conversely, more traditional administrative tasks like record-keeping lag behind, with only 25 percent of affected workers reporting AI usage.

Task Substitution vs. Partial Assistance

A critical distinction in the report is the difference between "AI-assisted" work and true "task substitution." Currently, most users view AI as a co-pilot rather than an autopilot. Full or near-full task completion by AI is still a rarity, hitting only 10 percent in software development and remaining below 7 percent for all other surveyed categories.

However, the phenomenon of task substitution—where AI takes over work previously handled by humans—is gaining measurable ground. The highest rates of direct substitution were found in:

  • Data Analysis: 7.1 percent of respondents.
  • Reading Work Documents: 5.7 percent of respondents.
  • Record-keeping: 5.3 percent of respondents.

Epoch AI emphasizes that this shift does not necessarily signal mass job displacement, but rather a redistribution of labor between human intelligence and machine processing.

The Paradox of Productivity and Output Quality

The data reveals a complex relationship between AI usage and time efficiency. When AI handles most or all of a task, 53 percent of workers report saving time, compared to 37 percent when the AI is used for only partial assistance. Interestingly, the "efficiency gain" is not universal; approximately one in six AI-assisted tasks actually takes longer than the traditional manual method. This suggests that the time required to prompt, verify, and interact with AI may offset some speed advantages.

Furthermore, the survey sheds light on the "trust gap" in AI outputs. A staggering 66 percent of AI-generated content is used with little to no editing, while only 5 percent of output is heavily reworked. While this suggests high perceived utility, researchers caution that low editing effort is not a definitive metric for the actual quality of the AI's output.

Key Takeaways

  • High-Tech Adoption: Software development and data analysis are the frontier sectors for AI, with adoption rates as high as 57% and 46% respectively.
  • Delegation Over Replacement: While 20% of workers delegate tasks to AI, full task automation remains low (under 10%), suggesting AI is currently a "versatile tool" rather than a replacement for human roles.
  • The Editing Trend: Two-thirds of workers accept AI output with minimal changes, indicating a growing reliance on the immediate utility of LLM-generated results.

Bottom Line

AI is no longer a fringe experiment for a handful of innovators; it is now a tool that one in five American workers relies on to offload tasks traditionally done by colleagues. The technology’s strongest foothold is in software development and data analysis, where adoption is 57 % and 46 % respectively, and full task substitution, though still modest, is growing. While productivity benefits are evident for many, the mixed efficiency results and the low level of editing suggest that the human review loop remains essential.