AI coding tools helped Bloomberg with prototypes, tests, and one-off scripts, but their measured impact fell quickly on existing codebases.
2
Bloomberg found stronger returns in agents that handle maintenance, refactoring, and incident response across large software systems.
3
A shared platform, training program, engineering community, and leadership workshops are needed to deploy AI across 9,000-plus engineers.
Summary
Lei Zhang describes Bloomberg's attempt to introduce AI into the workflows of more than 9,000 engineers. Early coding tools made proofs of concept and tests faster, but their impact dropped when teams moved beyond greenfield work. Bloomberg then focused on maintenance tasks, such as software patches and migrations, and on incident response agents that can inspect code, telemetry, topology, alarms, and service dependencies. At this scale, separate teams building overlapping tools would create duplication, so Bloomberg created a paved path with a model gateway, an MCP discovery hub, and a standard way to create and deploy tools. Adoption also depended on existing onboarding programs, cross-company communities, and inner-source work. Zhang says individual contributors adopted the tools more strongly than leaders, which led Bloomberg to add leadership workshops. He closes by arguing that AI changes which engineering work is affordable, so teams need to reconsider what high-quality software engineering should involve.
Bloomberg's scale makes AI deployment an engineering systems problem
Lei Zhang places the AI effort inside Bloomberg's broader engineering organization. The company has more than 9,000 engineers, a large private network, and a very large JavaScript codebase. The Bloomberg Terminal supports many applications and functions, including email, news, fixed-income calculations, and trading workflows. Zhang's technology infrastructure group covers global infrastructure, connectivity, developer productivity, source-control tooling, telemetry, and incident response. Bloomberg also has more than 500 employees focused on AI products. That combination means AI tools must work across many technologies, teams, and software assets rather than a single small application.
Coding assistants lose much of their early impact outside greenfield work
Bloomberg began trying AI coding tools about two years before the talk. The first step was to release capabilities quickly and measure how engineers used them. Surveys showed faster proofs of concept, quicker test creation, and many one-time scripts. The measurements dropped quickly when engineers moved beyond greenfield work. Zhang says the organization then had to ask where these tools could make a real difference. Bloomberg also became cautious about applying fast-generating tools to a codebase with hundreds of millions of lines, since changes are harder to verify as system complexity grows.
Uplift agents target maintenance work that engineers often avoid
Bloomberg began exploring uplift agents that scan the codebase for applicable patches and create a pull request when software needs an update. The proposed workflow includes the patch and an explanation of why it was made. Zhang contrasts this with an earlier regular-expression refactoring tool, which worked only to a limited extent. Large language models and related tooling produced better results for these changes. The approach still depends on deterministic verification. Weak tests, linters, or other checks make it difficult to apply patches safely. Bloomberg also saw average open pull requests and time to merge increase because generated code still needed human review.
Incident response agents can inspect systems faster and with fewer assumptions
Bloomberg is developing agents for incident response. Zhang says these systems can move quickly through a codebase, telemetry, feature flags, traces, and other operational data. They may also reduce the bias that can affect human troubleshooting, such as assuming that a suspected component must be responsible before checking the evidence. An incident response agent needs access to metrics and log dashboards, network or service topology, alarms, triggers, and service-level objectives. These connections require many MCP servers, which created a need for a common way to build and deploy them.
A paved path prevents teams from rebuilding the same AI infrastructure
With thousands of engineers and many teams building pull-request bots or incident response agents, Bloomberg saw a risk of duplicated and conflicting work. Its platform approach is a middle ground between complete team autonomy and a tightly controlled central platform. Zhang describes the goal as making the right thing extremely easy and the wrong thing very hard. The paved path includes a gateway for experimenting with models and tracking their use, an MCP discovery hub where teams can find existing servers, and a standard platform for creating and deploying MCP servers. The platform provides runtime infrastructure and handles authentication.
Production quality needs stronger controls than experimentation
Bloomberg wants engineers to try ideas and build proofs of concept with low friction. Zhang connects that freedom to creativity and experimentation. Production systems have a different standard because stability and reliability are central to Bloomberg's business. The paved path therefore separates easy experimentation from controlled production deployment. It gives teams a place to test different models, helps them choose a model for a problem, and reduces the effort needed to deploy an MCP server. This structure lets teams explore while keeping production systems subject to quality controls.
Existing training and communities can move AI adoption through the organization
Bloomberg added AI coding to an established onboarding program that has existed for more than 20 years. New engineers learn how to use the tools with Bloomberg's principles and technologies, then bring those practices back to their teams. Zhang says this can push more experienced engineers to consider new ways of working. Bloomberg also used its existing cross-organization champion and guild programs to build an AI productivity community. The community helped reduce duplicated effort, spread lessons, support inner-source contributions, and move engineers between teams when a project needed focused help.
AI adoption exposes a leadership gap and changes engineering trade-offs
Bloomberg's data showed stronger AI adoption among individual contributors than among leaders. Zhang says some technical leads and managers do not yet have enough direct experience with AI-assisted software work to guide their teams effectively. Bloomberg is responding with leadership workshops. He also argues that AI changes the cost function of engineering. Some work becomes much cheaper, while other work becomes more expensive. That change affects decisions about whether to build something at all. Zhang uses it as a reason for engineers and leaders to reconsider what high-quality software engineering means and how AI tools should support it.
