A Harvard analysis of 718 firms using the Jellyfish analytics platform shows that AI coding assistants and agents raise the amount of code written, but this does not translate into more finished work because the extra code creates additional load during later stages, especially code review. Engineers and platform teams need to understand why higher output can stall delivery pipelines.
AI coding productivity and delivery rates
The data indicate a clear separation between raw code generation and actual task completion. While developers see more lines or files produced, the study reports no corresponding rise in completed work items. This suggests that the metric of code volume alone is insufficient for measuring engineering throughput when AI tools are involved.
Review stage becomes a bottleneck
One of the primary downstream effects identified is an increase in code‑review workload. The additional AI‑generated code must still pass the same quality gates, and reviewers report higher volume without a proportional rise in review capacity. This can elongate merge cycles, increase context‑switching, and raise the risk of missed defects.
Architectural and operational considerations
Teams should treat the review bottleneck as a capacity planning signal. Potential actions include:
- Instrumenting CI/CD pipelines to track review queue length and cycle time.
- Allocating dedicated review resources or rotating review duty to balance load.
- Evaluating automated static analysis or linting tools to offload routine checks.
- Revisiting branch‑merge strategies to limit the size of AI‑generated change sets.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Before scaling AI coding assistants, measure their impact on downstream stages, especially code review throughput. Adjust staffing, tooling, and pipeline metrics to ensure that the apparent productivity boost does not become a hidden cost in delivery speed or quality.
