Bloomberg

Bloomberg employs 9,000 engineers on a codebase whose failure modes are financial-market failure modes, and that constraint shapes every AI decision Lei Zhang describes. In the note published 2025-11-23, Zhang, who leads technology infrastructure at Bloomberg, separates AI for coding (writing new features) from AI for software engineering (maintenance and operation), and puts the organization's weight on the second.

Two agent families carry that bet. Uplift Agents handle mass migrations and security patches, proposing explained patches across the codebase. Incident Response Agents scan logs, telemetry and configuration during an outage; Zhang's argument for them is that they are fast and unbiased, where humans arrive with priors about the likely cause.

The delivery mechanism is the Paved Path: make the right method easy and the wrong one hard. Bloomberg built a central platform with an «AI Gateway», simplified PaaS deployment for internal tools, and an «MCP Hub» so connectors are shared rather than rebuilt team by team. Zhang also notes that individual contributors adopt faster than managers, which pushed Bloomberg to put AI into onboarding training and let new joiners challenge seniors.

The other reason Bloomberg recurs is «BloombergGPT». Olivier Rafal cited it in February 2024 as the exception that proves his rule: building a foundation model from scratch costs months and millions, and only makes sense when you own monetizable proprietary data. Zhang's conclusion is narrower and sharper: AI changes engineering's cost function, so the old trade-offs need rewriting.