Bloomberg shifted from building its own large language model to building products on top of available models after ChatGPT and the open-weight community changed the trade-offs.
2
Agent systems are fragile because errors compound across composed language models, so downstream systems need their own checks instead of assuming upstream agents are always accurate.
3
Teams should start with a collapsed software and organizational structure for fast iteration, then split out shared foundations such as guardrails once the product and agent boundaries become clear.
Summary
Anju Kambadur describes how Bloomberg moved from building its own large language model in 2022 to building generative AI products on top of available models. He focuses on research analysts, whose work combines search, summarization, structured data, communication, and modeling. Financial products impose strict requirements around precision, speed, availability, data protection, and transparency. Kambadur argues that current systems are best treated as semi-agentic because some steps still need mandatory guardrails and human trust has not reached the point where everything can run autonomously. Scaling also requires a different approach to software and team structure. Language models and agents are stochastic, and their errors compound when they are composed. Downstream consumers therefore need independent safety checks. Teams can begin with a single group and a collapsed stack while the product is still changing, then factor out shared capabilities and smaller agents after their roles are understood.
Bloomberg moved from training its own model to building on available models
Bloomberg spent 2022 building its own large language model and wrote a paper about the work in 2023. That effort taught the company about model construction, dataset organization, and evaluation. After ChatGPT and the growth of the open-weight and open-source community, Bloomberg changed its strategy. It continued investing in similar problems, but began building products on top of models that were already available. Kambadur presents this as a practical response to the pace of model development and to Bloomberg's many product use cases.
Financial AI products have requirements that do not change when AI is added
Kambadur uses the research analyst as the talk's main example. Analysts search and summarize unstructured information, work with structured data and analytics, communicate with colleagues, and sometimes write programs and build models. Bloomberg's products must meet requirements around precision, comprehensiveness, speed, throughput, and availability. The company also has to protect contributor and client data and provide transparency. These conditions apply whether a product uses AI or not, so they shape how Bloomberg builds agentic systems.
Production summaries need monitoring and remediation because errors are public
Bloomberg began applying generative AI to earnings-call transcripts in 2023. The system could answer questions that matter to analysts, helping them decide whether to investigate a company more deeply. Out-of-the-box performance was not good enough on precision, accuracy, and factuality. Since the summaries are published for everyone to see, an error has a larger impact than a mistake in a private chat. Bloomberg therefore built monitoring, remediation workflows, circuit breakers, and CI/CD around the product, with repeated checks that make the summaries more accurate over time.
Bloomberg treats current products as semi-agentic because guardrails remain mandatory
Kambadur says Bloomberg does not yet trust every part of a system to operate autonomously. Some parts can act on their own, while other steps must always run as checks. A guardrail should catch requests that amount to financial advice, since Bloomberg does not offer financial advice. Other checks enforce factuality. These controls are coded as required steps rather than optional behavior, and they may be called at multiple points in a workflow.
Composed language models multiply small errors into fragile behavior
Traditional model APIs had some stochastic behavior, but teams could usually define their input and output distributions, create held-out tests, and monitor deployment risk. Agents make this harder because they compose multiple language models. Kambadur gives a structured-data example involving US CPI for the last five quarters. One missed character caused a monthly query to be treated as quarterly data. A downstream analyst looking at the underlying table might catch the problem, but a workflow that only exposes the final answer could return a plausible number without revealing the error.
Downstream systems need their own safety checks instead of handshake-based releases
Kambadur argues that upstream tools and agents should be treated as fragile and continually changing. An improved API can be better on average while becoming worse for a particular downstream consumer. Each consumer should therefore add its own checks rather than rely on every upstream team to explain each change and obtain sign-off before release. This approach lets agents evolve independently and reduces the coordination required to promote a new version to beta or production.
Agent and team boundaries should change as the product becomes understood
Early in a product's development, the design is unclear and fast iteration matters most. Kambadur recommends collapsing the organizational and software structure so one team can share code, data, and models while it figures out what to build. Once an agent's purpose and limits are clearer, teams can factor out foundations, improve performance, reduce cost, and make the system more testable and transparent. Shared functions such as guardrails can then become horizontal capabilities, while monolithic agents are split into smaller pieces.
A research agent separates question understanding, information gathering, and answer generation
Bloomberg's current research-agent design takes the user and session context, interprets the question, and determines what information is needed. That understanding is factored into its own agent. Answer generation is also separated, with explicit rules for what makes an answer well formed. Guardrails remain mandatory and are called at multiple points. The system builds on Bloomberg's existing data infrastructure, including sparse, dense, and hybrid indexes.
"In the beginning you don't really know much on what the product design is going to be and you want to iterate fast."16:35
Who should watch
You are shipping LLM or agent features into products where incorrect answers have operational or reputational costs.
Your team is deciding whether to keep a monolithic agent and organization or split out shared capabilities.
Downstream services depend on frequently changing models, data APIs, or agent workflows and need a release strategy that does not require constant coordination.
