Zscaler's AI Review System Saves Engineers Thousands of Hours

At Zscaler, the shift to a self-service data model was a victory for speed but created a new crisis: an unmanageable flood of pull requests. The cybersecurity firm's data team found itself drowning in up to 1,000 code reviews every quarter, each requiring deep scrutiny of complex data pipelines. Manual governance had become the bottleneck it was meant to prevent.

"We thought we had built a self-service paradise," said Rahan Raman, Head of Enterprise Data Platform at Zscaler. "Instead, we turned the data team into a help desk for peer reviews."

The solution, developed internally and named PRISM, is a multi-agent AI system built on a LangGraph orchestrator. Its effectiveness hinges on a rich feed of contextual information from dbt Cloud, including model lineage, test results, and performance metrics from CI jobs, combined with execution insights from Snowflake. This context allows the AI to move beyond generic suggestions.

"Without real context, AI is just noise," explained Senior Data Engineer Rishi Varahagiri. "When it understands lineage, performance signals, and validated compilation, it can give targeted, meaningful feedback."

The system deploys specialized agents to automatically check code structure, enforce documentation rules, analyze downstream impacts, and even propose optimized SQL. Critically, it validates any suggested changes before presenting them. Developers see the feedback as comments in their GitHub pull requests and can often accept fixes with a single reply.

The results are stark. Zscaler reports a 90% reduction in time spent by human reviewers. In one quarter, the system handled 956 PRs, projecting to annual savings of 2,100 engineering hours—the equivalent of a full-time employee. Reviews that once caused delays now provide immediate, consistent guidance, freeing the data team to focus on more strategic work. The AI, fed with structured context, has turned a governance bottleneck into a sustainable workflow.