A healthcare-adjacent organization based in the United States relies on software to serve customers and manage internal operations. Stack Builders works alongside its engineering team to expand delivery capacity through AI-assisted development.
The team needed to increase delivery capacity without lowering engineering standards. The question was not whether AI could write more code, but whether it could help a smaller team ship more useful work without creating more defects or review risk.
We introduced Claude Code into the delivery workflow, then paired AI-assisted development with peer review, adversarial AI review, a written Definition of Done, and recorded browser and API smoke tests.
A closer look at how we worked together on this project, including the engineering decisions, delivery practices, and collaboration model that helped us accomplish the goal.
We compared two four-week delivery periods using GitHub, git history, and Jira data to understand how team output changed after Claude Code entered the repository.
AI supported implementation and review, but was validated by peer review, engineering judgment, and the team's established delivery process.
An adversarial AI reviewer assumes code may be wrong and actively looks for issues against strict quality, correctness, and security criteria.
As AI helped the team take on more and larger changes, the review process became more rigorous too. Additional testing and AI-assisted quality gates helped ensure higher throughput didn’t come at the expense of engineering standards.
AI was introduced as an accelerator within an established engineering process. Developers used Claude Code for implementation, while peer review and AI-driven quality gates remained in place to increase output without lowering production standards.
About 94% of merged changes became AI co-authored after adoption.
Every pull request faced an AI reviewer designed to challenge correctness, quality, and security before merge.
A written Definition of Done plus recorded browser and API smoke tests added stronger checks around each change.
Peer review remained part of every pull request, preserving team judgment, feedback, and shared ownership.
After AI-assisted development became part of the workflow, the team delivered substantially more work with fewer active developers. Pull request and ticket throughput rose, while test code continued to keep pace with product code, and bug-fix PRs became less common. The gain was in total output, not faster turnaround for each change alone.
More pull requests merged over four weeks, with half the engineer count.
More Jira tickets delivered across the same four-week comparison.
Pull requests per developer per week increased from 1.9 to 4.8.
The workflow combined Claude Code, GitHub, and Jira with peer code review, adversarial AI review, automated test suites, and recorded browser and API smoke tests.
