Bloomberg is not betting on a single model for its Terminal AI: it divides and cites to avoid the costly error

By Zendoric · September 10, 2026.

Bloomberg has built ASKB, the artificial intelligence conversational interface for its Terminal —the software used by hundreds of thousands of finance professionals to trade and analyze markets— on an architecture of specialized agents rather than a single massive model. Wayne Barlow, global head of Terminal products, explained as much in an interview with the specialist outlet Hedge Fund Alpha. The tool has been in beta since early 2026 and follows a series of earlier steps: AI-generated earnings and news summaries, and a "Document Insights" tool that answered questions about a specific document.

The underlying reason for not using a single model is that different types of financial content are not analyzed in the same way. A news story is weighted above all by how recent it is; a regulatory document is weighted more by what type of document it is; and in both cases perspective matters. Bloomberg has chosen to distribute that work across different agents instead of asking a generalist model to solve it all at once.

The second pillar of the design is attribution. Every ASKB answer comes with citations: if the data is structured, the user can see the exact BQL formula (Bloomberg Query Language, the Terminal's own language for querying its databases) that generated it; if it comes from an unstructured source —a news story, a broker research note—, they can click and see the exact source passage highlighted. Bloomberg also extends that traceability to sell-side analysts themselves: many are Terminal subscribers, and buy-side investors can message them directly through Instant Bloomberg to ask about their research. For data that clients upload on their own, the company says it keeps it segregated from the Terminal's own database, without getting —according to Barlow— into the third-party data cleaning business.

The article, which has limited access, leaves out of view sections titled "Avoiding agentic drift and hallucinations" and "Training tomorrow's analysts": we know Bloomberg addresses them, but not the content, so we are not going to speculate about what we have not been able to read.

What the available facts do make clear is a revealing design decision about where enterprise AI is heading in high-stakes sectors. In finance, a misquoted figure or a confused source is not an aesthetic stumble: it moves investment decisions and, in the worst case, ends up in a law firm's office. We have already seen elsewhere the real cost of an AI that generates claims without anyone verifying the chain back to the source. Bloomberg appears to have learned that lesson before it blew up in its face: it prioritizes granular traceability over the eloquence of a single model, even at the cost of an architecture that is harder to maintain.

This fits a thesis we have been repeating at Zendoric: the race for the biggest model and the highest benchmark is not the one that truly determines who wins enterprise adoption in critical verticals. The winner is whoever solves the trust problem —which agent answers what, from which source, verifiable in one click— rather than whoever boasts the most powerful model. It is the same logic we already saw as competition among the big platforms shifts toward the "plumbing" of agent distribution: value moves from the model itself to the layer of governance, permissions and traceability surrounding its use.

For employment in the financial sector, Bloomberg's move reinforces a pattern we already identified when analyzing banking and insurance: AI does not simply replace the analyst, but it does shrink the room for purely administrative data-aggregation work, while raising the value of those who know how to interpret and verify what the machine proposes. That Bloomberg devotes an explicit section to "training tomorrow's analysts" —even if we have not been able to read the content— suggests the company itself sees that reskilling as a central part of the story, not a side effect. In the long run, tools like ASKB point to a landscape where analysts stop losing hours combing through documents and spend them on what genuinely requires human judgment: exactly the kind of abundance of time that, well managed, is the underlying promise of this technology.