AI
When AI Makes Code Changes Easy, Bottlenecks Shift to Review and CI Systems
A real-world case study reveals how large-scale AI-assisted code remediation strains continuous integration pipelines and review workflows, shifting the constraint from writing changes to managing and validating them.
Researchers have documented how large-scale code remediation powered by AI assistants fundamentally reshapes the bottlenecks in software development workflows. Instead of freeing teams from tedious work, the ability to generate hundreds of changes rapidly exposed new friction points in build systems, code review, and team coordination.
Study setup and findings#
A team led by Andreas Bexell conducted a 15-day field study in an industrial C++ repository, where a developer used an AI coding assistant to fix widespread code issues at scale. Over that period, the system rapidly generated hundreds of commits affecting thousands of lines of code—far more than traditional manual refactoring could produce.
The sheer velocity of changes had unintended consequences. Continuous integration (CI) systems designed for conventional patch workflows became overwhelmed. Code reviewers faced saturation from the volume of changes. Team members struggled to coordinate around the influx of modifications, despite their mechanical nature.
The shift in constraints#
The core finding challenges a common assumption: that making code changes cheaper through AI would simply accelerate development. Instead, the researchers observed that CI capacity, review bandwidth, and orchestration of change sets became the primary constraints—not the generation of fixes themselves.
When naive approaches delivered one commit per file, the per-file granularity overloaded build-on-commit systems. Switching to directory-based batching and capping files per change helped restore throughput, but introducing these workarounds required:
- Explicit negotiation with teams over acceptable commit sizes
- Manual solicitation of reviewer attention
- Iterative troubleshooting to resolve build and static-analysis failures
What sustainable remediation requires#
The research suggests that large-scale AI-assisted remediation in major codebases demands deliberate planning around three dimensions:
- Commit granularity: batching related changes to prevent per-file saturation of CI systems
- Review strategies: treating semantic change sets (e.g., "fix all instances of warning X") as coherent units that can be sliced differently for different stakeholders
- : designing batch processing that matches the volume and velocity of AI-generated changes
