Rapid AI-Driven Shifts in Work, but Benefits Are Uneven

Compared to the past five years of New Future of Work reports, this year's changes are far more urgent. Generative AI is not just making existing work faster—it is changing the very way people create, make decisions, collaborate, and learn.

The benefits are not spreading evenly. High-income countries still lead in usage rates, but the fastest growth is occurring in low- and middle-income regions, and when language support is weak, people switch to English to get better results. Usage gaps by gender and occupation are also large, and adoption gaps can turn into productivity, learning, and career gaps.

Inside organizations, culture determines adoption more than strategy documents. When employees trust their employer and feel safe experimenting, they adopt AI faster, and only the tools that actually help end up sticking around. Conversely, if AI looks like a mechanism for replacing people, resistance grows, and peer-to-peer sharing and grassroots experimentation produce the most substantial improvements.

Usage patterns are also becoming clearer.

• In Anthropic's analysis, 37% of Claude usage was tied to software and math occupations.
• Analysis of Microsoft Copilot conversations showed high utility across sales, media, tech, and administrative roles.
• Several studies show that AI users can be perceived as less competent even when they produce the same results.

Productivity effects are large but inconsistent. Enterprise users report saving 40-60 minutes a day, but that time can vanish again due to workslop—output that looks polished but is actually useless. At the macro level, no clear change has yet appeared in unemployment rates, total working hours, or job postings, but employment among young, less experienced workers aged 22-25 in high-exposure occupations has fallen by 16%, and a slowdown in junior hiring following AI adoption has also been confirmed.

The key point is that AI is not fully replacing people but is instead making human judgment an even more important resource. People are shifting away from the role of directly finishing tasks and toward guiding, critiquing, and improving the output that AI produces.

On the collaboration front, common ground becomes essential. Today's AI often skips the process of checking in as a conversation continues, making interactions prone to breaking down, but when systems are designed to prompt clarifying questions and collaborate across multiple turns—like CollabLLM—outcomes improve. Without trust and goal alignment, AI can actually lead to worse decisions, and selective delegation, where tasks are handed off selectively, helps people make better usage decisions.

Team-level use is also a new challenge. Because existing AI is designed for individual use, performance tends to drop when used in teams, but both a process-focused approach that helps share information and an outcome-focused approach that directly optimizes team performance are being studied simultaneously. Furthermore, a perspective is emerging that views LLMs not as personal tools but as collective intelligence that calls upon the combined intelligence of countless people.

The remaining challenge lies not in the technology itself but in the design and institutions surrounding it. For AI to create broad opportunity, accessibility, multilingual support, observability, humans' right to intervene, and interfaces for team collaboration must all advance together, and the future of work will ultimately depend on the choices we make.