Bloomberg AI: What ASKB Does on the Terminal and What It Changes

For decades the Bloomberg Terminal rewarded the people who memorized it, down to the function codes some users keep on sticky notes. In February 2026 Bloomberg gave it a different front door.

ASKB, Bloomberg's AI interface for the Terminal, takes a question in plain English, sends a network of AI agents through Bloomberg's data, news and research, and returns an answer with its sources attached and, where it has analyzed data, the query code behind the numbers.

For a junior analyst, that moves the job from finding the data to checking what ASKB assembles and supplying the idea it cannot. The legwork before a company reports, lining up peers, transcripts and consensus in one view, becomes something ASKB drafts in minutes. ASKB is also the latest turn in a story that includes BloombergGPT, the finance-trained research model Bloomberg unveiled in 2023.

Feb 2026

ASKB enters beta

on the Terminal and the Bloomberg mobile app

Several

AI models behind ASKB

commercial and open-weight, per Bloomberg

50bn

Parameters in BloombergGPT

the 2023 research model

Since 2009

AI work in finance at Bloomberg

as Bloomberg dates it

What ASKB changes for a junior analyst

Picture earnings season from the junior seat. For every company on the list, someone has to line up the share price against peers, read transcripts and filings, check the fundamentals and work out what the Street expects. Bloomberg's chief technology officer, Shawn Edwards, described that grind to WIRED in April 2026 in the first person, as if he were the analyst: "During earnings season, I'm not sleeping."

His answer is a template. "With ASKB, I can create workflow templates," he said. "I can write a long query, and say, 'Hey, here's all the data I'm going to need. Give me a synopsis of the bull and bear cases, what the Street is saying, what the guidance is.'"

From functions to questions

The old Terminal tested recall. Its power users, Edwards said, "pride themselves in knowing the path through the Terminal, remembering which functions to run". ASKB tests the question instead: the user "gets to ask the high-level question", the thesis already in their head, rather than asking for "particular data points".

For a first-year, that changes what counts as fluency. Knowing where a number lives on the Terminal matters less than knowing which number the argument needs, and what a complete answer should contain before you accept one.

What the machine cannot supply

Edwards is blunt about the limits. "These tools are not magical. They don't make an average [employee] all of a sudden great. The difference will be your ideas." In an expert's hands, he argued, the tools let an analyst "sift through 10 great ideas when they might have only had time for one. If you're a mediocre analyst, they'll be 10 mediocre ideas."

Asked whether automating a junior analyst's legwork would ripple through the labor market, he did not pretend to know: "I do think there's a big question of how to educate, train, and mentor junior analysts coming through the workforce. At least for the next few years, you can't just accept what some [AI] system says. You still have to have that rooted understanding of your craft. I don't know if the world has the answer."

The same split, drafting handed to the machine and judgment kept by the person, runs through what the big banks say about their own analyst classes, set out in will AI replace investment banking analysts.

What is ASKB on the Bloomberg Terminal?

ASKB is Bloomberg's conversational AI interface for the Terminal: one box for questions in everyday language. Bloomberg calls it agentic, and the word is doing real work. Behind the box, a coordinated network of AI agents works in parallel across Bloomberg's data, documents, news and research, retrieving, interpreting and synthesizing information before an answer is assembled.

Three design choices matter most to the person using it:

• Every answer carries its sources. Bloomberg promises transparent attribution to the documents and news behind a response, so a claim can be traced to the page it came from.
• Data comes with its query. When an answer includes data analysis, ASKB hands over the underlying Bloomberg Query Language (BQL) code, which the user can run in Excel or in Bloomberg's BQuant tools to extend or audit the work.
• Your files and colleagues can join in. Users can upload PDF and Word documents to be analyzed alongside Bloomberg's content, and message the analysts behind a piece of research through Instant Bloomberg, the Terminal's chat.

ASKB Workflows

The feature behind that earnings-season template is ASKB Workflows. A user describes a multi-step task, such as preparing for a company's results, analyzing them afterward or getting ready for a client or management meeting, and ASKB assembles the data, news, research and analytics into a structured output within minutes.

The user sets the objective, the structure and the tone, saves the workflow as a reusable template, re-runs it across companies or time periods, and can share it with colleagues at the same firm.

What ASKB draws on

An answer is only as good as what the agents can reach. Bloomberg lists these sources:

Which AI models does ASKB use?

ASKB does not run on one model. Bloomberg says it was built using multiple commercial and open-weight large language models, and Fortune reported in April 2026 that the mix includes models Bloomberg built itself alongside frontier models from companies such as Anthropic, with each query routed to the cheapest model that can handle the task with the reliability that particular job demands.

JPMorgan built its in-house LLM Suite on the same principle: its data chief said in 2024 that the plan was not to be "beholden to any one model provider", as the page on JPMorgan's LLM Suite sets out. The difference is what sits behind the window. On the Terminal, it is the archive Bloomberg sells.

How analysts use Bloomberg's AI features

Before ASKB put one interface over that archive, Bloomberg rolled generative AI onto the Terminal one feature at a time, each aimed at a specific chore in an analyst's week.

The users Bloomberg quoted when it launched Document Insights describe the trade: less time hunting, more time on the question. An equity research associate at ING said the tool made it quicker to find specific information inside earnings call transcripts than searching by hand, and easier to keep up with more of the peers of companies in their coverage: "This was definitely a more tedious task before."

A credit trading strategist at NatWest Markets described asking a manufacturer's transcript directly about the impact of tariffs, calling the tool "a godsend on a day when you have multiple earnings calls and cannot attend them all".

Look at the design habit running through the table. The earnings summaries link every point back to the transcript; the document tool highlights the passage and lets you replay the audio. Feature after feature hands the user a way to check it, and ASKB inherits the habit.

A worked example: earnings prep with ASKB

Here is how the pieces fit on an ordinary day in earnings season. The company and the numbers are illustrative; the steps and the checks are the part to learn.

Say you cover a listed restaurant chain that reports in two weeks. You build the prep once as an ASKB Workflows template and reuse it for every name on your list:

1. Ask for the setup in the terms Bloomberg's own CTO used: the bull and bear cases, what the Street expects, the latest guidance, and how the shares have traded against peers.
2. Ask for the card-spending data your firm is entitled to. Bloomberg's April 2026 roadmap uses this very query as an example: four weeks of year-on-year growth in Bloomberg Second Measure card data for a coffee chain, compared with revenue estimates for the quarter.
3. Read the synthesis ASKB returns, with its citations and the query code behind each figure, and decide what to check before any of it reaches your associate.

Suppose the answer says card spending at the chain is up 9% year on year over the last four weeks, against consensus revenue growth of 4% for the quarter. That reads like a beat. It is not one yet.

A quarter is 13 weeks, so four weeks is 4 ÷ 13, about 31% of it. If the other nine weeks grow at 2% and weekly sales are roughly even, the quarter blends to (4 × 9% + 9 × 2%) ÷ 13, which is 54 ÷ 13, about 4.2%. That is within a whisker of consensus. The strong month tells you where to dig, not what the quarter will print.

Before any of it leaves your desk, four checks:

The last step is the one the template cannot do. Write your own one-line view of what the Street is missing before you read the bear case the machine drafted, so you know which of the two you actually believe.

What does Bloomberg's AI get wrong?

Bloomberg does not claim ASKB is infallible, and its CTO is unusually specific about where it can slip. "We can never say it's perfect," Edwards said. "More of the problem is that we might not answer a question fully." The harder error, in other words, is an answer that is incomplete rather than invented, because nothing in it is false.

Behind the validators sit people. In a separate interview in April 2026, Edwards called evaluations "the make-or-break of building a useful, trustworthy system", built with bond covenant experts, equity analysts, market structure specialists and Bloomberg's own journalists. Models can grade the easy cases; for everything else, Bloomberg uses human assessors.

The news summaries have had corrections. The New York Times reported in March 2025 that Bloomberg had corrected at least three dozen AI summaries of articles published that year, including one that put a tariff decision in the wrong year and others with incorrect figures or attribution.

Edwards draws one more line that matters for anyone using the output at work: "Bloomberg never gives a buy or sell signal to somebody." The decision, and the responsibility for it, stays with the user. The systems, he said, "should be used to drive users to sources, not to hide them".

How many people use Bloomberg's AI?

All of this runs on the Terminal, and the Terminal's base is large. WIRED reported in April 2026 that the ASKB beta was open to roughly a third of the Terminal's 375,000 users. Fortune put the subscription at about $30,000 per user a year that month, and wrote that the Terminal is still considered "the de rigueur tool of every trader, investment banker, and hedge fund quant."

Bloomberg built ASKB for people doing company and equity research: equity analysts, portfolio managers and other investment professionals. In August 2026 it tied the desktop and mobile versions together, so a question started at the desk can be finished mid-thread on an iOS or Android phone or tablet.

Bloomberg's launch materials also name early users. Schroders, Bernstein and Mizuho Securities were early beta users of the document search and analysis tool announced in June 2025, and Colin McGranahan, then Bernstein's head of research, called it "a useful addition to how we navigate and summarize research on the Bloomberg Terminal."

Where Bloomberg's AI came from: BloombergGPT

Bloomberg dates its AI work in finance back to 2009, long before chatbots. Entity extractors fed sentiment functions such as BSV and TREN, which estimate how positive or negative a piece of news is for a company, and a natural-language query interface let users ask questions in plain English. "But it was brittle," Bloomberg's CTO said of that interface. "You had to ask the Terminal in very specific ways."

On March 30, 2023, Bloomberg published a research paper describing BloombergGPT: a 50-billion-parameter language model built from a corpus of about 700 billion tokens, roughly half of it financial text drawn from an archive Bloomberg has collected over four decades. The paper's own breakdown shows what "trained on finance" meant in practice.

Financial web pages

298bn

News from other outlets

38bn

Company filings

14bn

Press releases

9bn

Bloomberg-authored content

5bn

The Pile (public)

184bn

C4 (public)

138bn

Wikipedia (public)

24bn

From Bloomberg's 2023 paper. Highlighted bars are the financial dataset (363bn tokens), the rest public data (345bn); each group's categories sum one higher because of rounding.

In Bloomberg's own tests, the model beat open models of a similar size on financial tasks by wide margins without giving up ground on general ones. But it was not built for Terminal users to talk to.

What BloombergGPT was for

"Most of the things that we're doing with BloombergGPT are going to be behind the scenes," Gideon Mann, then head of Bloomberg's ML Product and Research team, told Institutional Investor in April 2023. "You're not going to interact with that." There was no BloombergGPT website and no chat window. Bloomberg was already using the model to generate "silver data" for training smaller models, and CNBC reported that the company was not planning a ChatGPT-style chatbot.

The early job was translation. The Terminal's query language is powerful, and even people who know it, Mann said, are "not using the full power of the language". One early idea was to let a user type "Tesla price" and have the model write the query, get(px_last) for(['TSLA US Equity']). ASKB performs the same step in the open and hands the code back to the user.

Three years on, Bloomberg's archive, once the model's training data, is what ASKB's agents search. Mann's 2023 description of the work under way for clients, "summarization, or monitoring, or being able to ask questions on those news stories or transcripts", reads like a preview of the features that followed.

How Bloomberg's AI is changing

Bloomberg's April 2026 roadmap aims ASKB at a firm's own work as well as Bloomberg's, which brings a junior's questions closer to the firm's own positions and research. The integrations it lays out:

• Portfolios. Through PORT, the Terminal's portfolio and risk analytics, users will be able to query personal and firm-wide security lists to explain past performance and sources of future risk.
• Internal research. Research teams on Bloomberg's RMS platform will be able to upload their own notes and proprietary models, and ask, in the roadmap's own example, where their analysts disagree most with the Street.
• Alternative data. Near real-time datasets for nowcasting company KPIs, like the card data in the worked example above.
• Expert calls. Excerpts from the expert-network interview transcripts a firm is entitled to.
• Automation. ASKB Workflows scheduled to run at a set time or on a market trigger, such as a customized morning brief or a weekly check on a thesis, and shared with teammates.
• Suggested questions. Questions suggested by Bloomberg's subject-matter experts, set to begin appearing inside supported Terminal functions.

"This will be the new Terminal," Edwards said. He does not expect the old screens to vanish: "I don't think GUIs [graphical user interfaces] are going away." But "by and large, people will start their analysis and workflows through ASKB."

What he wants next is proactive work: agent-to-agent workflows and always-on monitoring, with Bloomberg acting as "the eyes and ears" that flag a second-order risk before the client thinks to ask.

What Bloomberg's people have said

The arc, in the words of the people building it:

What to know about Bloomberg AI for an interview

Treat ASKB the way you would any tool on the desk you hope to join: know what it is, know where it fails, and be ready to show how you would check it.

What a good answer contains

• The distinction. BloombergGPT was the 2023 research model; ASKB is the product analysts use. Mixing them up is the fastest way to sound like you stopped at a headline.
• The mechanism. Agents retrieving across Bloomberg's data, answers that cite their sources, numbers that come with the BQL behind them, and research steps you can save and rerun each quarter.
• The limit. Incomplete answers are the hard error, the tool gives no buy or sell signal, and the idea still has to come from you.

How to prepare

1. Take Bloomberg Market Concepts, Bloomberg's self-paced course on the markets through the Terminal. Its eight sections include Terminal Basics, and the course teaches the keyboard, command line, menus and autocomplete.
2. Take Bloomberg Spreadsheet Analysis if you can. It teaches BQL inside Excel, the same language ASKB hands back when it shows its working, and reading that code is how you audit an AI answer's numbers.
3. Practice the manual version of that prep: pick a company, pull last quarter's transcript, write your own three-bullet summary and bull and bear cases, then compare them with any AI tool you can use and note where they differ.
4. Prepare one specific example of catching an AI tool's mistake, and how you caught it. The interview readiness tool walks through questions of that kind, and AI in investment banking covers the tools banks have built for themselves.

For engineers moving into finance

Bloomberg's own account of building ASKB reads like a job description. Beyond the cost-based model routing and the expert-built evaluations described above, Bloomberg pulled those experts off their day jobs to write benchmarks for individual sub-agents as well as to help grade whole workflows.

Bloomberg's CTO said in April 2026 that data ingestion which used to take four and a half months now takes two days, and many of the people once dedicated to data entry and cleaning were redeployed onto building evaluations. Evaluation design, plus fluency in BQL and BQuant, is the bridge between the AI stack and the finance desk.

The bottom line

In 2023 Bloomberg showed it could train a model on its archive; in 2026 it put an interface on the Terminal that takes questions the way people actually ask them. ASKB is what reaches the desk: ask in plain English, get an answer with its sources and its query code, and save the steps to run again next quarter.

What it changes for a junior is where the value sits. Knowing which function to run matters less. Knowing whether the answer is complete, whether four weeks can stand in for a quarter, and what the Street is missing matters more. That last part is the one ASKB leaves to you.