← back

What We Learned Deploying AI within Bloomberg's Engineering Organization – Lei Zhang, Bloomberg

Original: What We Learned Deploying AI within Bloomberg’s Engineering Organization – Lei Zhang, Bloomberg

15.9K views · Dec 16, 2025 · 18:20 min · Watch on YouTube ↗
Takeaway

At enterprise scale the win is uplift/incident agents on a paved path, not just developer copilots — and reviewer bandwidth becomes the new bottleneck.

Summary

  • Lei Zhang (Bloomberg infrastructure lead) describes deploying AI tooling across ~9,000 engineers running one of the world's largest JavaScript codebases supporting the Bloomberg Terminal's thousands of 'functions'.
  • Survey-based ROI showed quick wins on POCs, tests, and one-off scripts but the productivity gains dropped fast on brownfield work in a hundreds-of-millions-line codebase.
  • Two ROI examples: 'uplift agents' that scan the monorepo and propose patches with reasoning (replacing prior regex-based refactor tooling), and 'incident response agents' that traverse codebase, telemetry, feature flags, and traces faster and less biased than humans.
  • Side-effect of broad AI rollout: open pull requests increased and time-to-merge grew because more code is generated than humans can review.
  • Bloomberg uses a 'paved path' platform model (golden path + enablement team) with a model gateway for evals, plus standardized MCP servers for metrics, logs, topology, alarms — to prevent 10 teams independently building incident bots.
bloombergenterpriseagents
Original description
When it comes to using AI for software engineering, much of the spotlight falls on how large language models (LLMs) can write code—sometimes entirely from scratch. Countless studies highlight productivity gains from turning requirements directly into runnable code. But the reality of applying AI at scale inside a mature engineering organization is far more complex and nuanced. Over the past year, we’ve been on that journey at Bloomberg—integrating AI into the workflows of 9,000+ software engineers—and we’ve learned a few important lessons worth sharing:

Where the real ROI lies once you move beyond toy examples
What it takes to actually enable AI across a large, established engineering org
The best practices, cultural shifts, and guardrails that are required to make it work in practice
If you’re wondering what happens after the first demo magic fades and the real work begins, this talk is for you.

Speaker: Lei Zhang  |  Head of Technology Infrastructure Engineering, Bloomberg

Timestamps

00:00 Introduction to Bloomberg's Scale & Infrastructure 
03:32 AI for Coding: Initial Adoption & The "Greenfield" Drop-off 
06:14 Uplift Agents: Automating Refactoring & Maintenance 
08:40 Incident Response Agents: Unbiased Troubleshooting & Speed 
09:37 The "Paved Path": Standardizing AI Tool Building (MCP) 
11:51 Platform Components: Gateway, Discovery Hub, and Deployment 
13:34 Leveraging Training & Communities for Adoption 
16:15 The Leadership Gap & The Changing Cost Function of Engineering