Generative AI at work: what the field experiments show

Most of what is said about generative AI and productivity is anecdote. A few studies are different. They compared workers who were given access to a model with colleagues who went without; access was assigned at random in two of the studies and rolled out in stages in the third. Three of them, published in leading journals and covering very different jobs, are worth reading side by side.

What the studies measured

• Writing. In a preregistered experiment with 453 college-educated professionals, those given ChatGPT took 40% less time on occupation-specific writing tasks, and their output was rated 18% higher for quality (Noy and Zhang, Science, 2023).
• Customer support. Among 5,179 support agents, access to an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and lower-skilled agents. For the most experienced, the effect was minimal (Brynjolfsson, Li and Raymond, The Quarterly Journal of Economics, 2025).
• Consulting. In a field experiment with 758 consultants, those using GPT-4 on tasks within the model's capabilities completed 12.2% more tasks, 25.1% faster and with more than 40% higher quality. On a task chosen to fall outside those capabilities, they were 19 percentage points less likely to reach a correct solution (Dell'Acqua et al., Organization Science, 2026).

What the three have in common

The gains are real but uneven. They peak where the task is well defined and the model is strong, and they shrink or turn negative once the task falls outside what the model does reliably. Dell'Acqua and colleagues call this boundary a jagged technological frontier. It does not follow intuitive lines of difficulty, so two tasks that look alike can sit on opposite sides of it.

The least experienced gain the most. In the writing and support studies, the gap between weaker and stronger performers narrowed. For an organisation that is a design question, because the tool spreads good practice only as far as the model has learned it.

Speed is easy to measure; quality is what decides. Each of these studies measured quality separately from time, and the consulting result shows why. Inside the frontier the tool raised quality. Outside it, the same tool lowered correctness.

What this means for AI in operations

Design the workflow before you choose the model. Decide which tasks sit inside the frontier, set a quality bar for each before rollout, and keep a human review step wherever a confident wrong answer is expensive. Measure quality task by task, alongside the time saved.

The product-content system I worked on at FLO followed the same pattern. Content production time fell by about 70%, yet the rollout went category by category, each behind a quality bar agreed in advance, and manual work moved from producing content to reviewing samples. That discipline is exactly what the research suggests matters.