The 2026-published Copilot RCTs studied 2022–23 autocomplete

Source. Cui et al., Management Science (published February 27, 2026); full author-hosted February 2025 draft. The underlying Microsoft experiment ended May 3, 2023; Accenture began July 2023; anonymous Fortune 100 rollout was September–October 2023.

Result. Three randomized workplace experiments across 4,867 developers found a pooled +26.08% weekly completed tasks (standard error 10.3%) among tool users by an instrumental-variable estimator, with code suggestions/autocomplete Copilot. Weekly pull requests operationalize completed tasks; fewer gains for senior/long-tenured than recent/junior Microsoft engineers. This is evidence of gains in real work, but not a trial of command-line coding agents. Study quality proxies include build success, not a full measurement of maintenance or user value. Company experiments were ad hoc and differed.

Reading hazard. Publication year is not intervention year. An attractive '2026 randomized trial shows agent boost' headline would misdate the technology by ~3 years and conflate workflow mechanisms. Compare 2026 Microsoft CLI telemetry and METR's later agentic-era experiment.