The Code-Volume Delusion: Rethinking Engineering Velocity in the AI Era

# The problem in plain terms AI coding assistants make generating code and tests fast. That shows up as tickets moving to "In Review" and lots of lines of code created. Yet deployment frequency can stay flat or decline. The cause is systems-level: speeding up one step that wasn't the main bottleneck simply shifts friction elsewhere.

# Where the bottleneck moves AI shifts the hardest work downstream into three places: pull-request review, CI/CD pipelines, and long-term architecture. Each introduces new costs that reduce throughput even as visible output (code volume) increases.

# Three metrics to watch and act on

• CI/CD stability and P95 build times. AI often generates many tests. If tests are low-value, flaky, or overly implementation-coupled, the pipeline will fail frequently or run slowly. Track build failure rate on main and P95 build duration. If P95 is ballooning (the article cites examples like pipelines hitting 45 minutes), your effective velocity is zero.

• Code churn rate and architectural decay. AI behaves like an eager junior developer: it fixes the immediate issue without system-level design. High churn (large percentages of code rewritten within weeks) signals trial-and-error development and growing technical debt. Compare tech-debt ratio against feature delivery to see if fixes outpace sustainable design.

# Practical controls to restore throughput

• Enforce PR size limits and review discipline. Require smaller, focused diffs. Rotate senior reviewers and limit approvals under five minutes with escalation policies.

• Create a Build Gardener role each sprint. Assign one engineer to prune useless tests, fix flaky assertions, and optimize caching. Remove low-value end-to-end tests and prefer stable unit or integration checks.

• Mandate architectural review checkpoints. For changes that touch core modules, require design notes and a lightweight architecture review before merging larger PRs.

• Track the right metrics, and connect them to incentives. Replace lines-of-code KPIs with time-to-first-review, build failure rate, P95 build duration, code churn, and tech-debt ratio.

# If you don't change

# Bottom line