Understanding Bloomberg’s Agentic AI for High-Stakes Data

In the high-stakes theatre of global finance, information is not just power – it is a deluge. 

Every single day, the Bloomberg terminal processes a staggering 450 billion data points and 1.1 million news items, serving as the central nervous system for the world’s investment professionals. 

But as AI moves from experimental novelty to an essential tool, a critical question has emerged: how do you ensure a machine understands the nuance of a market shaped by human emotion?

Despite the automation gold rush, Amanda Stent, Head of AI Strategy and Research at Bloomberg, is clear that keeping humans in the loop is a competitive edge. 

While AI is masterful at stripping away the drudgery of repetitive data retrieval, it lacks the one thing a veteran analyst prizes most: discernment.

“The true value of AI lies in how it empowers the person behind the screen,” Amanda says. “While AI handles retrieval and synthesis, financial professionals spend more time interpreting and strategising. 

“Markets are shaped by data, incentives, regulation and psychology; a model can detect a pattern, but it takes an experienced human to discern if it’s a meaningful signal or just noise.”

For Amanda, the evolution of the financial analyst is less about learning to write code and more about mastering the art of the inquiry.

“Becoming effective with AI requires domain expertise and critical thinking – knowing which questions to ask and which anomalies to challenge. The ultimate differentiator will be critical thinking, not coding. The analyst of tomorrow will stress-test assumptions and make higher-quality decisions supported by deeply customizable evidence-gathering AI systems.

“Bloomberg’s AI systems support user decision-making but do not offer financial advice; they are underpinned by our focus on developing safe, reliable and compliant Gen AI systems.”

Constructing a ‘ground truth’

The industry’s biggest headache with generative AI is hallucination, where models invent facts. 

To combat this, Bloomberg has introduced ASKB, a conversational AI interface that helps professionals access Bloomberg data. 

Every news article, research document and social media post ingested into the system is automatically enriched with metadata, tagging entities, sentiment and salience to ensure the AI isn’t just reading text, but understanding context.

Amanda explains that building trust in these systems requires moving away from the black box approach of standard LLMs: “It’s actually hard to use generative AI to build trustworthy systems. What I mean by that is it’s hard to build a system that gives you consistently accurate answers over and over again; a system that capital markets finance professionals can use to make crucial business and investment decisions. 

“The problem lies in the fact that the underlying models themselves don't have a ground truth. So we have to give them one.”

That “ground truth” is Bloomberg’s proprietary fortress of data: decades of structured analytics, curated company documents and alternative data – from credit card spend to foot traffic. By tethering AI to these high-value sources, Bloomberg ensures the model doesn’t lose the truth of the original data. 

“The actual challenge is how do you use this reasoning power to make sure that these models don't use their world knowledge to answer a query, but instead only use these trusted sources of information,” Amanda adds. 

“What this means is that you have to build domain-specific checks and guardrails. Often these checks are deterministic systems. 

“Did you call that calculator with the right parameters in the right order? When you summarised a paragraph, did you faithfully ensure that all the information content came strictly from that paragraph and you didn't insert anything extra? 

“These validators, these checks, are something that you just keep building – and they are domain-specific. They check that the information in responses is correct and timely before the output reaches the user. 

“To do this right, you must have knowledge about what is right and what’s not – and that’s why we involve our subject matter experts in helping us design our AI solutions.”

To further ensure transparency, ASKB provides attribution to the source for every answer. If the system provides a data point, it even reveals the Bloomberg Query Language code behind it, allowing users to export that exact logic into Excel or BQuant for further testing.

Privacy by design

In a sector where a leaked document can move markets, privacy is non-negotiable. 

Bloomberg has built ASKB with strict boundaries to ensure client data never becomes training fuel for the wider world.

“Bloomberg does not use client content to train or fine-tune generative AI models for the purpose of generating, displaying, summarising or reproducing such content without additional consent,” Amanda explains. 

“In addition, we have fine-grained role-based access control for all content available to AI systems. In the case of file uploads, only the user who uploaded the document is able to access them within the ASKB thread in which they’ve been analysed.”

“Where third-party model providers are used in ASKB, their access and handling of the data are governed by contracted terms with strict access, data usage and retention restrictions to ensure client and proprietary data is protected from training and fine-tuning by the LLM provider.”

The power of the agentic crowd

Rather than relying on one massive, slow-moving model, ASKB has a multi-agent architecture. Think of it as a room full of specialised experts – one for news, one for filings, one for maths – all working in parallel. This allows for hyper-specific functionality that a single model simply couldn't handle.

“When you search for news, recency matters,” Amanda highlights. “If you fold all the news into a single massive model, responses will blend old and new news stories and you will never get responses based on the latest breaking news after the model was trained. 

“Whereas, when you search for company documents, type matters: a 10K, a company presentation – these provide complementary sources of information. Also, across these and other document types, perspective matters. 

“Our clients want to separate the perspective of the company from those of analysts who may be covering the company and news coverage. With a multicomponent agentic system, each type of data can be handled as it should and combined in the ways the user needs.”

While the tech world often obsesses over speed, Amanda insists that, in finance, being right is better than being fast – though ASKB manages to achieve both.

“With ASKB, Bloomberg is focused more on accuracy than speed – something our clients seem to appreciate, regardless of whether they’re an equity analyst, portfolio manager, or credit analyst,” Amanda says. 

“That being said, it certainly does speed up users’ financial analysis and investment management workflows and how they discover insights on the Bloomberg Terminal.

“ASKB is made up of multiple AI models to help get answers to your questions. In response to your queries, ASKB calls and orchestrates different AI agents and tools that our team has built to discover and distill information and content from across the Bloomberg Terminal. 

“These different individual agents and tools can access Bloomberg’s comprehensive and high-quality library of more than 400 million documents... Different AI agents perform a variety of tasks in parallel, including discovering information, synthesising data and running calculations."

The goal is to eliminate the swivel-chair effect, where analysts spend hours jumping between applications to manually synthesise research. 

By providing contextually accurate, timely answers – down to the exact page and paragraph of the source document – Bloomberg is turning the AI into a bridge between a question and a fully-formed investment thesis.