When one step of a system gets faster, the constraint moves to the next slowest step. AI coding is a textbook case.
Symptom 1: bigger changes, slower review
DX reported in June 2026 that median pull request size nearly doubled over the year to June 2026, from 44 to 72 lines. Larger PRs take longer to review and hide more defects. If review time climbs while PRs merged per engineer stays flat, review is the bottleneck.
What helps: PR size limits, AI-assisted first-pass review, review SLAs, spreading review beyond a few senior engineers, and stacked or smaller changes.
Symptom 2: more output, more rework
GitClear analysed 211 million changed lines from 2020 to 2024 and found copy-pasted lines rising from 8.3% to 12.3% of changes, while lines classified as refactoring fell from 25% of changed lines in 2021 to under 10% in 2024. More duplicated code means more future maintenance. If change-failure rate rises alongside the AI-authored share, your quality gates are the bottleneck.
What helps: required tests for AI-written changes, security and quality scanning in CI, duplication checks, and clear rules for which changes need senior review.
Symptom 3: fast code, slow pipeline
If cycle time is measured in weeks or deploys are rare, faster coding barely matters. Delivery foundations — CI speed, test automation, trunk-based development, deployment automation — are the constraint.
What helps: fund CI and deployment work before more AI tooling.
Symptom 4: low or shallow adoption
If few engineers use AI, or only for completion, the bottleneck is enablement: access, training, approved tools and use cases that fit your codebase.
Diagnose, then act
The productivity scorecard applies these patterns as transparent rules to your numbers. Treat its output as a hypothesis to discuss with each team, not a verdict.