Understanding the Productivity Impact of Generative AI in the Workplace
Early workplace studies of generative AI did not identify a single, universal productivity multiplier. Instead, they revealed a nuanced pattern—a productivity slope influenced by employee experience and skill level. At the lower end of the experience curve, AI assistants functioned like on-demand coaches, providing phrases, procedures, and links precisely when workers needed them. Conversely, at the higher end, these assistants often repeated knowledge already well known by the most experienced employees. This dynamic not only increased overall output but also narrowed the performance gap between novices and veterans.
One of the most prominent findings emerged from a large-scale study involving over 5,000 customer-support agents. The widely cited 2023 working paper reported a 34 percent productivity increase among novice and lower-skilled workers, while experienced agents showed minimal gains. However, this figure requires context: subsequent peer-reviewed revisions adjusted the estimate slightly downward but confirmed that the largest benefits consistently accrued to less experienced employees.
The Origin and Evolution of the 34 Percent Productivity Figure
In a 2023 NBER working paper, Erik Brynjolfsson (Stanford), Danielle Li, and Lindsey Raymond (MIT) analyzed the staggered rollout of a generative AI assistant at a Fortune 500 business-process software company. Covering 5,179 customer-support agents, their study found that AI access increased issues resolved per hour by 14 percent on average. Remarkably, novice and lower-skilled agents improved their productivity by 34 percent, while the most experienced workers saw little change.
The peer-reviewed version, published in 2025 in the Quarterly Journal of Economics, refined the analysis with data from 5,172 agents. It estimated a 15 percent average increase in successful resolutions per hour and roughly a 30 percent improvement among less-skilled workers. This adjustment reflects typical academic refinement rather than contradiction, underscoring the core message: AI’s productivity effects vary significantly based on initial skill and experience.
How the AI Assistant Supported Agents Without Replacing Them
The AI tool operated by monitoring live text chats between agents and customers, generating suggested responses. It could recommend diagnostic questions, suggest phrasing, or link to technical documentation. Importantly, agents retained full control—they could accept, modify, or ignore AI suggestions. Productivity was objectively measured by the number of successfully resolved customer issues per hour, integrating chat duration, volume of handled chats, and resolution rates. This metric was independent of any vendor ratings.
The rollout was staggered, with managerial input determining when teams and agents gained AI access. Researchers employed difference-in-differences models controlling for agent-specific factors, calendar time, and job tenure. They also conducted robustness checks using alternative estimators and instrumental-variable analyses based on team rollout timing. These methodological safeguards enhance the credibility of findings while acknowledging the limits of generalizability: the study reflects one AI tool deployed in one company within a relatively stable product context.
Accelerating Learning Curves Through AI Assistance
The most striking evidence came from comparing agents’ experience curves. Without AI, agents started at approximately 1.8 successful resolutions per hour, reaching about 2.5 after eight to ten months. Agents using AI from their first month achieved the same performance level in roughly two months and continued to improve thereafter. In other words, two months of AI-assisted tenure matched the output of over six months without AI.
This pattern suggests that AI steepened the initial learning curve rather than replacing human learning. The mechanism behind this acceleration is likely knowledge transfer: the AI was trained on past customer interactions, including strategies from top-performing agents, enabling it to deliver effective responses in real-time. Gains were most pronounced on moderately uncommon problems, where AI had sufficient training data and human agents lacked experience. For very common issues, novices were already proficient, and for rare issues, the AI had limited data to assist.
Evidence That Learning Persisted Beyond AI Assistance
A key concern with AI productivity tools is whether workers become dependent or truly learn. The study found suggestive evidence that learning survived AI use. During rare system outages, agents with longer prior AI exposure maintained faster resolution rates compared to their pre-AI baseline, with the effect strengthening alongside AI tenure.
While this evidence is not definitive—outages were infrequent and unevenly distributed—it implies that some AI-generated knowledge became internalized rather than vanishing when suggestions ceased. Additionally, customers interacting with AI-assisted agents became more polite and less likely to escalate issues. The study also linked AI access with reduced employee turnover, particularly among newer workers, though authors cautioned about causal interpretation given nonrandom access and the one-time nature of attrition events.
Supporting Studies Confirm the Experience-Based Productivity Gradient
The call-center findings align with other early research. A preregistered experiment involving 453 college-educated professionals randomly assigned ChatGPT access for occupation-specific writing tasks. Results showed a 40 percent reduction in completion time and an 18 percent increase in quality ratings. Importantly, participants with weaker initial skills benefited more, narrowing the productivity spread.
Similarly, an experiment with 758 Boston Consulting Group consultants found that GPT-4 assistance improved performance by 43 percent for those below median baseline skill, compared to 17 percent for higher performers. Notably, on tasks outside GPT-4’s competencies, AI users performed 19 percent worse, illustrating the risk of misplaced reliance—a phenomenon dubbed the “jagged technological frontier.”
Further, a controlled study of GitHub Copilot revealed developers completed JavaScript programming tasks 55.8 percent faster with AI assistance, with larger gains for less-experienced coders. While promising, such short coding exercises differ markedly from the complexities of maintaining production systems.
Implications: Compressing Skill Gaps Without Eliminating Expertise
Generative AI excels at redistributing codified knowledge—if a pattern appears frequently in training data and is expressible in text, AI can deliver it directly to novices, raising the baseline performance floor. However, true expertise remains vital at the margins. Experienced workers excel at recognizing when AI-generated patterns misapply, detecting novel customer problems, identifying incorrect confident answers, and balancing speed with quality.
Interestingly, the most skilled agents in the support study sometimes experienced slight declines in conversation quality metrics when using AI, hinting at complexities in human-AI collaboration. Another concern is circularity: high performers create the examples that train AI. If they increasingly rely on AI’s adequate but unoriginal suggestions, future knowledge bases may lose innovation, potentially eroding the foundation for future improvements.
For management, this shift suggests a redefinition rather than elimination of senior roles. Coaching, quality assurance, exception handling, and innovation cultivation may become more crucial as routine expertise spreads more widely via AI tools.
Productivity Gains Do Not Guarantee Wage or Employment Outcomes
The customer-support study did not assess wages, employment levels, or hiring practices. Firms might respond to faster novice ramp-up by hiring more entry-level staff, reducing training costs, redesigning roles, or even downsizing. Productivity improvements alone cannot predict which path a company will take.
Previous analysis by Silicon Canals highlighted how this AI assistant narrowed performance gaps between new and veteran agents. The broader early evidence suggests that when tasks align with AI strengths, weaker workers gain more. Yet, AI compresses advantages derived from repetition and exposure to common cases without replacing critical judgment for atypical situations. Thus, the emerging skill hierarchy may flatten in routine execution but steepen in strategic decision-making about when to trust AI.
For further details, see the original source Here.
