Bloomberg has launched an enterprise AI access layer for its Data License Plus (DL+) offering, giving clients’ AI agents a standardised way to discover, interpret and retrieve licensed Bloomberg data across more than 100 million securities and over 50,000 fields.
Bloomberg Enterprise Model Context Protocol (MCP) combines the data with metadata, semantic context and entity relationships designed to help AI agents understand what individual data points represent and how they should be used. The service also includes workflow-focused “Skills” that allow agents to perform defined tasks using Bloomberg data.The launch comes as financial institutions move generative and agentic AI from experimentation into production, putting new demands on the data infrastructure supporting those systems. Rather than simply giving an agent access to a price, estimate or fundamental data point, production applications increasingly require historical context, identifiers, calculation methodologies, entitlements and other information that conventional data retrieval may leave to the consuming application to interpret.
“Models available to clients today are pretty impressive in what they can do,” Tony McManus, Global Head of Enterprise Data and Indices at Bloomberg, tells Market & Alt Data Insight. “The gap is on the data side: to move from a demo to a real production workflow, an agent needs access to a broad, interconnected set of data, described by rich metadata and context, at enterprise-grade governance, and that applies to internal data as much as vendor data.”
From snapshots to production data
One challenge is the difference between providing an AI model with sufficient data to demonstrate a use case and providing the depth of information required to run the same process reliably in production.
McManus points to snapshot data as a particular weakness in early implementations. “The specific failure mode we see most is snapshot access being treated as production-grade.,” he says. “Plenty of early implementations give an agent today’s price or today’s estimate, which is fine for some use cases but doesn’t cover what financial professionals actually run in production, a quant backtest needs point-in-time constituents and corporate-action-adjusted history, an analyst tracking revisions needs the full history, not today’s consensus. Without that context, the model fills the gap with assumptions, and in finance, assumptions can become problematic quickly.”Bloomberg Enterprise MCP is designed to expose the metadata underlying its datasets alongside the data itself. That can include the meaning and calculation methodology of individual fields, as well as qualifiers such as exchange, currency, price type, period and as-of date. Semantic search allows an agent to identify fields using natural language, while entity-resolution capabilities map names, tickers and descriptions to instruments and their relationships with issuing companies. This approach puts more of the interpretation layer traditionally embedded in market data platforms directly within the information made available to AI systems.
“For an AI agent’s output to be trusted in a production workflow, it needs high-quality data and rich semantic context sitting behind it. Without that context, the model guesses, and guessing isn’t acceptable when the stakes are a trade or a risk decision,” says McManus. “Two providers can both claim to expose ‘financial data’ and produce completely different answers from an agent, because the protocol doesn’t tell you whether a revenue figure is GAAP or adjusted, whether a rating applies to the issuer or the instrument, or whether an identifier is pre- or post-corporate-action.”
MCP standardises the connection
MCP is an open protocol designed to provide a standardised way for AI applications to connect to external tools and data. The protocol was contributed by Anthropic to the Linux Foundation’s Agentic AI Foundation in December 2025, alongside founding projects from Block and OpenAI. For market data providers, such standardisation could reduce some of the differentiation historically associated with the mechanisms used to deliver data. Bloomberg’s view is that it simultaneously places greater emphasis on what sits behind the connection.
“Standardising the connection with MCP is a good outcome for the industry, it means clients can stop worrying about plumbing and focus on what actually matters, the data. But that’s exactly why MCP is table stakes, not a differentiator,” notes McManus. “Once everyone can plug into the same protocol, the differences between providers become more apparent. It’s the quality of the metadata and semantic context behind the connection that determines whether even the most advanced model produces a reliable answer. The firms that lead won’t be the ones who connected to a protocol first, they’ll be the ones whose data was built for this mode of use.”
Bloomberg has also been involved in the development of MCP itself. Its engineers have authored proposals intended to introduce enterprise governance controls and tailored toolsets, while Bloomberg leads the MCP Financial Services Interest Group and holds seats on the board and Technical Committee of the Agentic AI Foundation.
From data retrieval to workflows
Enterprise MCP extends beyond data discovery through a set of reusable workflow-focused “Skills”. The initial release includes capabilities for point-in-time universe retrieval, corporate-action adjustment and revision histories, as well as identifying outliers, reviewing trades against price-deviation thresholds and assessing securities and trades for potential sanctions exposure.
Bloomberg hosts and manages the MCP layer, while clients retain control of the agents and applications connecting to it, including their choice of models, prompts and instructions. Entitlements are checked before data is returned, with requests operating under existing Data License rights.
“This is still the early stage, though it’s evolving quickly. Agents are already automating a lot of manual workflow, but almost always under an analyst or researcher’s supervision before the final decision gets made,” says McManus. “As clients get more comfortable with model outputs and rack up real, day-to-day production usage, fuller automation is going to become the norm rather than the exception. The biggest open question isn’t technical readiness at this point, it’s the governance framework AI agents will need to operate under.”
Bloomberg plans to add real-time data to Enterprise MCP, extending its potential use into areas including intraday monitoring, pre-trade checks, risk refreshes and exception handling.
The resulting distribution model puts the data itself alongside the metadata, semantics, permissions and workflow logic needed by machines to use it. As MCP lowers the friction involved in connecting agents to different sources, those surrounding layers are becoming part of the market data product delivered to the client.
Subscribe to our newsletter
