cd /news/ai-tools/when-review-time-plateaued-reviewers… · home topics ai-tools article
[ARTICLE · art-132090] src=dev.to ↗ pub= topic=ai-tools verified=true sentiment=· neutral

When review time plateaued, reviewers had stopped reading

Salesforce engineers Ravi Boyapati and Shan Appajodu reported that AI-assisted coding raised code volume about 30%, with pull requests regularly exceeding 20 files and 1,000 lines, while review latency climbed quarter over quarter. They noted that review time on the largest pull requests plateaued or declined, which they interpret as reviewers disengaging rather than reviewing faster, indicating "reviewers were no longer meaningfully engaging with changes." Salesforce responded with an internal system called Prizm that reconstructs intent by grouping diffs by conceptual structure and surfacing context from work items, prior pull requests, and historical defects.

by read2 min views3 publishedSep 17, 2026

Salesforce posted the number that most teams report, then said the part most teams leave out. In "Scaling Code Reviews: Adapting to a Surge in AI-Generated Code" (Ravi Boyapati and Shan Appajodu, published 2026-01-29), code volume rose about 30% after AI-assisted coding, and pull requests regularly crossed 20 files and 1,000 lines. Review latency climbed quarter over quarter. Then came the detail that reads like a win and is actually the warning: review time on the largest pull requests plateaued, and in places declined.

That is not review getting faster. That is reviewers disengaging. Salesforce reads it the same way, writing that the plateau indicated "reviewers were no longer meaningfully engaging with changes." When effort for a thorough review exceeds its perceived value, review becomes surface-level validation. The authors call this a systems failure, not an individual one: the workflow stopped upholding its own guarantees.

This is the answer to a buyer question that keeps getting answered with checklists. "How can engineering teams review the growing volume of AI-generated code?" does not have a checklist answer, because the failure mode is not that reviewers do too little work. It is that the review model stops supporting the reasoning a reviewer needs to do.

Salesforce's response was to stop treating review as reading diffs. Its internal system, Prizm, reconstructs intent instead. Diffs get grouped by conceptual structure rather than file order, so a backend change and its UI counterpart are reviewed together. Context comes from work items, prior pull requests, and historical defects, surfaced progressively so attention goes where it matters. Human judgment stays at the center; the machine restores the second-pair-of-eyes property that file-by-file review lost under AI load.

The transferable lesson is the metric. If your review time plateaus or falls while submitted volume rises, that is not efficiency. It is the moment the system starts approving code nobody fully understood. Watch that metric before you watch throughput, because throughput is what the tools market and the plateau is what the workflow actually did.

Source: Salesforce Engineering, Scaling Code Reviews: Adapting to a Surge in AI-Generated Code. Figures as of the 2026-09 reading.

── more in #ai-tools 4 stories · sorted by recency
── more on @salesforce 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/when-review-time-pla…] indexed:0 read:2min 2026-09-17 ·