When Making Productivity Assessments, Output Matters

Since productivity measurements involve a comparison of output compared to input, we cannot ascertain much by looking only at the hours people work. We have to compare those inputs to the outputs created during those hours. 

And that is the sense in which we must evaluate studies that suggest workers using AI sometimes wind up working more hours than they used to. 

To be sure, some of that might simply reflect time spent learning how to use AI. Some increased time might be consumed checking AI outputs for accuracy. 

But it is at least conceivable that outputs might be changing as well. 

AI automates routine tasks, freeing capacity and making more complex or additional work feasible and intrinsically rewarding. This can lead workers to voluntarily expand their scope, take on new roles, multitask more, or extend hours, rather than reducing total hours worked, one study found.  

Study / Source	Key Findings on Productivity & Hours/Complexity	Link
Ye & Ranganathan (UC Berkeley Haas, 2026) – Ethnographic field study	AI intensified work: faster pace, broader/more complex tasks (voluntary role expansion), extended hours without being asked. Felt rewarding short-term but risked unsustainability.	HBR Article; Haas News
ActivTrak (Observational, ~164k-443M work hours)	AI users showed intensified activity (e.g., +104% email, +145% chat, more weekend work); focused/deep work down ~9%; productivity via denser output rather than fewer hours.	Reported in WSJ/HBR coverage
Noy & Zhang (2023) – RCT with 453 college-educated professionals on writing tasks	ChatGPT: ~40% less time, +18% quality. Workers enjoyed tasks more; weaker writers benefited most (could tackle higher-quality/complex output).	Science
Dell'Acqua et al. (BCG/HBS/MIT/Wharton, 2023) – Field experiment with 758 consultants	GPT-4: +12% tasks completed, 25% faster, significantly higher quality (esp. "inside the frontier"). Enabled broader capabilities.	HBS Working Paper; MIT Sloan coverage
St. Louis Fed / Bick et al. (survey data, 2024-2025)	GenAI users: ~5.4% work hours saved (more for frequent users); implies ~33% higher productivity per AI-assisted hour. Aggregate ~1.1% U.S. productivity boost. Time savings could enable more complex work.	St. Louis Fed
Other experiments (e.g., coding/consulting)	Mixed: Gains often larger for juniors (enabling complex work); some show output increases without proportional time reduction due to scope expansion.	Various (e.g., Microsoft/Accenture Copilot trials)

In other words, it is possible that AI can result in workers doing more, rather than less; spending more input hours rather than fewer. 

That might happen because AI makes complex work more accessible and engaging. 

So working more, rather than less, is rational for career growth, intrinsic motivation, firm expectations or simply because the new work is interesting.

What we do not know yet is AI impact on productivity. We cannot only measure inputs. We have to know whether outputs have increased, and if so, by how much.