When review time plateaued, reviewers had stopped reading

Salesforce's Scaling Code Reviews report reveals a troubling trend: as AI-generated code volume surged, code review latency increased and plateaued on the largest pull requests, even declining in some cases. This suggests reviewers were disengaging rather than becoming more efficient. The plateau indicated reviewers were no longer meaningfully engaging with changes, viewing the effort required for a thorough review as not worth the perceived value.

The authors argue this is a systemic failure of the workflow, not an individual reviewee's lack of effort.

The problem lies in the review model itself, which no longer supports the reasoning required for AI-generated code. Salesforce's response is to stop treating code reviews as simply reading diffs. Their internal system, Prizm, reconstructs intent behind the code rather than just presenting changes file-by-file. By grouping diffs by conceptual structure rather than file order, related changes like backend and UI components are reviewed together.

Context is drawn from work items, prior pull requests, and historical defects, surfacing the most relevant information progressively so reviewers can focus their attention where it matters most.

However, the review process should still center on human judgment. The machine can only provide a second pair of eyes insofar as it accurately reconstructs the underlying intent of the code. Salesforce emphasizes monitoring review time as a critical metric - if it plateaus or drops despite rising submitted volume, it signals the system is approving code that hasn't been fully understood.

This is a more telling sign of an efficiency failure than raw throughput numbers alone. Reviewers must watch this metric closely to ensure the review process continues supporting meaningful reasoning rather than becoming a surface-level validation exercise.