The productivity conversation surrounding artificial intelligence (AI) in businesses has been narrowed down to one question: how quickly can employees complete existing tasks? Anthropic’s internal research suggests that framing leaves something out. The company found that 27% of AI-assisted work within Anthropic was due to tasks that employees could not attempt without AI. The work wasn’t unrealistic just because it wasn’t worth it. That was not practical due to the time cost.
What the data shows
Anthropic surveyed engineers and researchers across the organization, conducted 53 in-depth interviews, and analyzed 200,000 internal code records. Employees reported using Claude for 60% of their work and estimated average productivity gains of about 50% (up from 20% a year ago). Over the same period, usage increased from 28% to 60% of daily work.
The output data will be more specific. Across nearly every task category, employees reported spending slightly less time per task, but producing significantly more. The use of Claude code has moved on to more complex tasks. The average number of consecutive tool calls completed by the model without human input doubled from approximately 10 to 21, and the percentage of tasks involved in implementing new features increased from 14% to 37%.
Engineers talked about using AI to build interactive dashboards, scale low-priority projects, fix long-neglected code quality issues, and perform exploratory research where the time cost isn’t justified manually. One researcher explained that they run multiple Claude instances in parallel to test different approaches simultaneously, treating the model more like a fleet than a faster car.
OpenAI’s enterprise research found a similar pattern, with 75% of employees surveyed reporting that they were now able to complete new tasks that they were previously unable to perform. EY’s U.S. AI Pulse study found that 39% of organizations are reinvesting AI productivity gains into research and development, suggesting that the scaling effect extends beyond the completion of individual tasks.
Where companies are struggling
The picture across companies is not so homogeneous. PYMNTS Intelligence found that 71% of executives at companies with annual revenue of $1 billion or more identify organizational readiness as the primary limitation to AI performance. Only 11% said technology itself was a barrier. PYMNTS Intelligence reports that 58% of CFOs cited talent shortages as their main challenge, rising to 71% for service companies.
Advertisement: SCROLL TO CONTINUE
Cost control is also under pressure. PYMNTS reports that Uber’s AI budget has skyrocketed more than expected as internal use of Claude Code expands, with about 11% of live updates to backend systems being written by AI agents and R&D spending increasing 9% to $3.4 billion in 2025.
economics of execution
As AI reduces the cost of analysis, documentation, coding, and research, tasks that were once below the threshold of feasibility will now cross the threshold of feasibility. Deloitte found that only 34% of companies are using AI to significantly transform their core processes and products, while the remaining two-thirds are achieving efficiency gains without redesigning their underlying operations.
Internal investigation results have limitations. Human engineers have early access to frontier models, work in stable fields, and build technology themselves. The company acknowledged that the findings are not directly applicable to other organizations. According to a PYMNTS Intelligence study, 34% of CFOs at large companies cite productivity as the top reason for implementing AI. Anthropic plans to expand its research beyond engineers to understand how AI impacts roles across organizations, with further discoveries expected in 2026.
