Microsoft AI coding agents boost pull request merges by 24%

Microsoft’s internal study of its early‑2026 rollout of Anthropic’s Claude Code and GitHub Copilot CLI found that engineers who adopted the agents merged roughly 24 % more pull requests than they would have otherwise, a lift that persisted across a four‑month window【1】. The boost matters for large firms because merged PRs are a common proxy for engineering output, and the finding suggests that visible peer use, not demographics, drives sustained adoption.

Adoption driven by social networks

The study tracked usage among tens of thousands of Microsoft engineers. Initial exposure spread primarily through existing social networks rather than formal training programs, indicating that peer visibility is a key catalyst for trying the agents. Retention correlated more with an engineer’s overall coding activity than with age, gender, or tenure, suggesting that the tools become part of the workflow for active developers rather than serving as a novelty for occasional users【1】.

Impact on engineering velocity

Using merged pull requests as a proxy for output, the researchers measured a 24 % increase in merged PRs for adopters versus a counterfactual baseline. This lift held steady over the four‑month observation period, implying that the productivity gain is not a short‑term spike but a durable effect. The authors caution that a merged PR does not equate to business value, but the metric is widely used to gauge engineering throughput at scale【1】.

Implications for enterprise rollouts

The findings challenge the assumption that AI coding assistants will be uniformly adopted across an organization. Instead, rollout strategies should prioritize making peer usage visible—such as dashboards or internal showcases—to stimulate organic diffusion. Companies must also weigh the token costs, which can reach millions of dollars annually at scale, against the measured output gains to determine ROI【1】.

What to watch

• Internal adoption metrics – Microsoft’s next internal report may reveal whether the 24 % lift extends beyond the initial four‑month window.
• Token spend trends – Monitoring token consumption will indicate if cost efficiencies emerge as usage matures.
• Competitor rollouts – Watch for announcements from Google’s Gemini CLI or other enterprise AI tools that could alter the adoption dynamics.

The study underscores that AI‑driven command‑line agents can meaningfully augment engineering output, but the effect hinges on peer influence and sustained usage rather than broad, demographic‑based adoption. Whether similar lifts will appear in other enterprises remains an open question.