Evidence note

The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers

Controlled Experiment2026Management Science

What was examined?

Three randomised field experiments involving professional developers at Microsoft, Accenture and an anonymous Fortune 100 company.

Randomised access to an AI coding assistant, with development task completion used as a primary productivity outcome.

4,867 developers across three experiments.

Key findings

01Across the combined experiments, developers with AI assistance completed approximately 26.08% more development tasks.
02Less-experienced developers showed stronger adoption and larger measured gains.
03Individual experimental estimates were noisier than the combined result.

Why it matters. Lornets interpretation.

There is credible causal evidence that AI assistance can materially increase development output in some professional settings. That result does not independently establish faster releases, higher production quality or equivalent increases in business value.

This is the Lornets reading of the source, not a finding of the source itself.

What it does not establish

01Completed development tasks are not equivalent to production releases.
02The result does not establish a universal productivity effect across tasks, tools or engineering environments.
03The measured productivity effect should not be translated directly into a percentage reduction in software cost or delivery time.