Microsoft Study: AI Coding Agents Raise Pull Requests 24%, But Review Queues Pile Up

A large-scale Microsoft study published on July 1, 2026 found that developers who regularly used command-line AI coding agents merged roughly 24% more pull requests per engineer per day over a four-month period. The tools studied were Anthropic's Claude Code and GitHub Copilot CLI, both rolled out internally in early 2026. The key finding is not the headline number but the conditions behind it: usage frequency drove almost all of the gain, and a parallel enterprise study shows that more code output is creating serious pressure on human code reviewers.

What happened

Data point	Detail
Study published	July 1, 2026
Tools studied	Claude Code and GitHub Copilot CLI
Increase in merged pull requests	+24.0% per engineer per day
Likely range	+14.5% to +33.7%
Study duration	Four months
Lift at 5+ days/week usage	More than 50%
Lift at ~3 days/week usage	Roughly 15%
Copilot CLI vs Claude Code lift ratio	Copilot CLI ~2.2x higher (Microsoft environment only)

Microsoft studied its own early-2026 rollout of two command-line AI coding agents. Unlike editor plugins, these tools run in the terminal and let developers hand off coding tasks with fewer step-by-step instructions. The researchers measured merged pull requests, the unit of work where a developer submits code changes for team review and approval, as their primary metric.

The 24% figure held for the full four-month window without fading. To check whether the jump was coincidental, the team ran a placebo test, treating an earlier date as if the rollout had started then. No similar spike appeared, which supported the conclusion that actual tool use drove the increase.

Why usage frequency is the real story

The gap between heavy and light users is striking. Developers who used the agents five or more days a week saw gains above 50%. Those who used them around three days a week saw about 15%. Simply having a license made little difference if the developer did not actually open the tool.

At Microsoft, adoption spread mainly through peer and manager networks. Engineers were more likely to start using the tools when colleagues or direct managers were already on board. Developers who already wrote a lot of code kept using the agents; lower-activity developers were more likely to drop off.

That pattern has a practical implication for IT and engineering managers: seat count is a poor proxy for value. A team with 50 licenses and 10 active daily users is not getting 50-license results.

What a separate enterprise study found

A second study, published July 2, tracked 802 developers and 196,212 pull requests at a mid-sized company from January 2024 through April 2026. The company’s CTO announced a formal “2x mandate” in mid-2025, using merged pull requests per engineer per month as the target metric.

By April 2026, throughput had reached 2.09 times the pre-mandate baseline, rising from 21.2 to 44.3 merged pull requests per active developer per month. But the gains were uneven.

Metric	Before mandate	After mandate (April 2026)
Merged PRs per developer/month	21.2	44.3
PRs receiving human review	89%	68%
Automated AI review coverage	~19%	~84%
Time to merge after first human review	Baseline	~20% longer
Time to merge overall	Baseline	~22% longer

Output grew most in newer repositories. Legacy codebases saw little lift, and the pattern held regardless of developer seniority. Individual contributors and principal engineers showed similar results, which points to the codebase itself as the dividing line rather than skill level.

Is review quality keeping up?

This is where the data gets uncomfortable. As AI-authored pull requests multiplied, human review coverage fell from 89% to 68%. Automated AI review coverage jumped from 19% to 84%, and workload per human reviewer roughly doubled. Merge rates stayed mostly flat and the revert rate actually declined, so work kept shipping. But each AI-authored PR took about 20% longer to merge after the first human review and 22% longer overall.

A third study, published in ACM Transactions on Software Engineering and Methodology, looked at 567 Claude Code pull requests across 157 open-source projects. The merge rate was 83.8%, but only 54.9% of those merged without additional changes. The remaining 45.1% required human revision, particularly for bug fixes, documentation, and project-specific standards.

Taken together, the three studies suggest AI agents can move code faster through a pipeline while quietly shifting work from writing to reviewing and cleaning up.

Our take

The 24% headline is real, but it comes with a footnote the size of the finding itself. You only get the full gain if developers use the tools almost every day, and you pay for it partly in reviewer time and slower final merges. For teams already stretched on senior engineering capacity, doubling code output without expanding review bandwidth is not a productivity win, it is a bottleneck waiting to appear.

The legacy codebase result is also worth sitting with. Organizations running older systems hoping AI agents will accelerate a rewrite or modernization project may find the tools work poorly in exactly the environments that need the most help. If you are evaluating these tools, test them on your actual repositories, not a greenfield demo project.

The Microsoft caveat on Copilot CLI versus Claude Code is correct: a 2.2x ratio measured inside Microsoft’s own infrastructure and culture tells you very little about which tool suits your team. Run your own comparison for at least four to six weeks before drawing conclusions.

For teams curious about weaving AI tooling into broader development workflows, our AI integration services cover practical setups beyond the IDE. And if the throughput numbers above have you reconsidering how your development pipeline is structured, GPT-5.6 Sol’s agentic coding efficiency gains offer a useful comparison point on where the model side of this equation is heading.

What to do about it

1. Measure actual daily active usage of any AI coding agent, not just license activations.
2. Audit your review capacity before expanding access: if reviewers are already a bottleneck, adding AI-authored PRs will make it worse.
3. Pilot the tools on a newer, well-documented repository first; legacy codebases are a harder environment to start in.
4. Track review time and revert rates alongside PR volume so you catch quality degradation early.
5. Let peer and manager networks lead adoption organically rather than mandating use, since the Microsoft data suggests forced adoption produces lower and shorter-lived gains.

The metric that matters is not how many PRs shipped, it is how many stayed in production without problems six months later.

Source: Bing News · Claude AI