Bloomberg, FactSet and LSEG AI Compared: Which Answers Can an Analyst Trace?

Bloomberg, FactSet and LSEG have each put a generative AI assistant on top of their data, and only one has documented a way to rerun the number behind an answer. Bloomberg's ASKB hands back the Bloomberg Query Language code it used, and about a third of the Terminal's 375,000 users had it by April 2026. For financial services AI, the citation is the product. The three vendors document it very differently.

Short answer. On published vendor material, Bloomberg ASKB offers the most traceable answers: it attributes responses to source documents and news, and returns the query code behind any calculation. LSEG Workspace AI Search promises citations down to each data point. FactSet documents the least. None of the three publishes which subscription tier includes the AI.

This post is for the practitioner: the analyst who has to defend a number in a model, a note or an investment committee memo. That reader converts on one question only, which is whether the answer survives being checked.

What Bloomberg, FactSet and LSEG AI tools actually answer

All three vendors sell the same promise. Ask a question in plain English and get an answer drawn from data you already pay for. The differences sit in the corpus behind the answer, and in what the assistant hands back with it.

Bloomberg ASKB

Bloomberg launched ASKB in beta on 23 February 2026. The launch release describes it as a conversational interface over Bloomberg's structured data, unstructured documents, news, research and analytics. It names hundreds of millions of company documents, 5,000 original stories and more than 1.1 million curated stories a day, plus research from over 800 providers.

Users can also add their own PDF and Word files. The Trade reported that ASKB runs on a mix of commercial and open-weight large language models, coordinated as several agents rather than one model. In April 2026, Third Bridge said its library of more than 100,000 expert interview transcripts had joined the ASKB beta.

LSEG Workspace AI Search

LSEG introduced Workspace AI Search in July 2026. Its product page says the agent draws on market data and filings, alongside Reuters News, Deals and Aftermarket Research. The tasks it lists are retrieving data, screening markets, comparing entities, summarising documents and analysing trends.

Answers come back as summaries, tables or visualisations. The page says the tool is live in Workspace and in Workspace for Microsoft Teams, and will soon reach Microsoft 365 Copilot.

FactSet Mercury and the 2026 agents

FactSet's first public generative AI product was Mercury, released as a beta in December 2023 for junior bankers. FactSet's own write-up says it covered fundamentals, pricing data, and bank and branch regulatory data. The example tasks were a pitch-ready chart, or a SWOT analysis drawn from a company's 10-K.

In June 2026 FactSet announced a partnership with Google Cloud to build agents for portfolio operations, deal advisory and corporate finance. PYMNTS reported that FactSet is embedding Google's enterprise search and Gemini models in the FactSet Workstation to launch those agents.

Every cell is the vendor's own description or a direct report of it. "Not stated" means absent from the material we opened, not proof that the feature does not exist.

How Bloomberg AI cites the underlying data

Bloomberg's launch release says ASKB grounds every response in its own data and "includes transparent attribution to original research documents and news sources". Its AI product page goes further. It promises "direct links to source documents" and the ability to extract cited text and data into your own research notes.

A WSJ CIO Journal report, relayed by AI Weekly in October 2026, adds that each answer cites its original source and a timestamp. That detail is secondary. I could not confirm it in Bloomberg's own pages.

The BQL code is the real audit trail

The most useful line in the whole release is about code, not citations. When an answer includes data analysis, ASKB returns the underlying Bloomberg Query Language code. You can paste it into Excel, BQuant Desktop or BQuant Enterprise and run it again.

A citation tells you where words came from. A query tells you how a number was built. If ASKB reported that a company's margin fell, the code would show which field, which period and which adjustment produced the figure. That is the difference between trusting the answer and checking it.

My view is plain. Of everything the three vendors have published, this is the single most important design choice for an analyst. It turns a generated answer back into a deterministic data request, which is the form your compliance team already knows how to audit.

The limit is just as plain. Attribution to a sell-side note proves the note said it. It does not prove the note was right, or that the summary kept the note's caveats. The BQL route only covers numbers built from Bloomberg's structured data. Anything synthesised from documents still needs a human to open the source and read the paragraph. The general failure pattern is covered in the audit of where deep research tools invent their sources.

How LSEG Workspace AI cites its sources

LSEG's July 2026 page makes the boldest claim of the three. It says Workspace AI Search provides "transparent citations that allow users to trace every data point back to its source". Every data point is a higher bar than every answer.

The provenance theme is consistent in LSEG's messaging. LSEG's Nej D'Jelal told FinTech Global in March 2026 that data used to make decisions "needs to be linked to its underlying provenance". That is the right principle, and it sets the standard LSEG should be held to.

The strongest promise, the least shown

The page shows no worked example of what a citation looks like. It does not say whether a figure links to a filing page, a Workspace data item, a Reuters story or all three. It does not say whether a calculated ratio cites its inputs or only the final value.

I would not treat that as a weakness of the product. I would treat it as the first question in a demo. Ask the rep to pull one figure that LSEG derives from a filing, then click through to the exact line in the filing. If that works live, the claim holds.

LSEG's corpus also has one asset the others cannot copy, which is Reuters News. For event-driven questions, a citation to a Reuters story with a time on it is often more useful than a citation to a research note. The open question is whether AI Search preserves that timestamp in the answer. The page does not say.

How FactSet AI cites its sources

FactSet's clearest public statement on sourcing is from its December 2023 Insight post on Mercury. It said the beta could "generate provable, source-linked company research from FactSet data". The release was part of FactSet Explorer, an early-access programme for beta products.

The same post said a buy-side release was planned for spring 2024. I could not confirm from material I opened whether that release shipped on schedule, or what its citation format became.

FactSet's main AI announcement this year was the June 2026 Google Cloud partnership. PYMNTS reported that FactSet is "directly integrating with Google grounding" and that its MCP and agent-sharing functions will connect the Workstation to Gemini Enterprise. FactSet's release boilerplate puts its base at more than 9,000 clients and over 241,000 individual users. That is a large audience for a tool whose citation behaviour is barely described in public.

Grounding is not citation

Grounding means the model checks its answer against retrieved material before replying. It reduces invented facts. It does not, by itself, hand the analyst a link to the FactSet data item behind a figure.

I disagree with the common reading that grounding solves the audit problem. An answer can be grounded and still be untraceable to the person reading it. For an analyst, the visible link is the feature, not the hidden check.

AI financial research tools compared on citation behaviour

To compare the three fairly, separate four levels of traceability. Each level is harder to build than the one before, and each one is more useful to an analyst.

Four levels of traceability

Level 1 names a source, such as "Reuters" or "company filing". Level 2 links to the specific source document. Level 3 points to the passage or data point inside that document. Level 4 returns a query you can rerun to rebuild the number from scratch.

Most consumer chat assistants stop at level 1 or 2. Professional research needs level 3 for text and level 4 for numbers. On published material, only Bloomberg describes level 4, and only LSEG explicitly claims level 3 for every data point.

On paper, each vendor stops somewhere different. Bloomberg stops short of a clear data-point claim for document answers. LSEG stops short of showing the format, or offering a rerunnable query. FactSet stops at a 2023 beta description, so its current behaviour cannot be judged from public material.

That ordering is about documentation, and documentation is a product decision too. A vendor that publishes exactly how its citations work is inviting you to test them. A vendor that does not is asking you to take the demo on trust. If your team is already checking AI output inside spreadsheets, the test of AI accuracy on Excel financial models shows how quickly unchecked numbers spread.

Which subscription gets you each financial research AI tool

This is where the vendor material goes quiet. None of the three states a list price for its AI features. None says plainly whether the assistant is included in a standard seat or sold as an add-on.

Bloomberg's rollout is a queue, not a tier

Bloomberg's release frames ASKB as part of a Bloomberg Anywhere subscription, with no separate fee mentioned. The WSJ report says Bloomberg is extending it from a beta subset to the full subscriber base, without a date. In practice, access is decided by rollout order rather than by plan.

LSEG ties AI access to existing data rights instead. The October 2025 Microsoft release says customers can build Copilot Studio agents with data licensed through Workspace and Financial Analytics. A March 2026 LSEG article adds that "data rights controls are embedded in the MCP connectors".

That is a sensible design. It means the AI cannot hand you a dataset your contract does not cover. It also means two analysts at the same firm may get different answers to the same question, depending on what each one is licensed to see.

FactSet's access is the hardest to read. Mercury started as an early-access beta. The 2026 agents are announced, with Google, but no access terms appear in the coverage I could open. If you hold a FactSet contract, the only reliable source for what your seats include is your account team, in writing.

When the AI answer leaves the terminal

All three vendors are now pushing data out to general-purpose assistants. LSEG's data reaches Microsoft 365 Copilot and Copilot Studio. FactSet is connecting the Workstation to Gemini Enterprise. Bloomberg's product page offers a standard MCP interface for firms' own enterprise AI applications.

MCP, the Model Context Protocol, is an open standard that lets an AI assistant call outside data sources through one common interface. The mechanics are covered in the explainer on MCP as the agent interoperability standard.

Licensing travels, citations may not

Here is the risk nobody in these announcements addresses. When LSEG or FactSet data answers a question inside Copilot or Gemini, the host assistant decides how the answer is shown. The data vendor controls the licence. It does not obviously control the citation.

None of the LSEG or Microsoft pages I read describes how citations appear on the Copilot path. The FactSet coverage describes grounding and interoperability, not citation display. I think that gap matters more than any feature in this post, because the Copilot path is where most non-specialist staff will meet this data.

If your firm is choosing a general assistant for finance work, read the comparison of Claude and Copilot for finance teams alongside this one. Then check how each handles access rights, using the analysis of permission gaps in enterprise AI search as a guide.

Freshness is a second, quieter gap. A citation to a figure is only useful if you know when it was true, and market data moves by the second. The WSJ report says ASKB timestamps its sources, which addresses this for Bloomberg. Neither of the other two vendors' pages mentions timestamps. Stale answers from agents are a known failure, described in the breakdown of why agents give stale answers.

Where this comparison is weakest

This post compares documentation, not products. I did not test any of the three tools hands-on. A vendor can document poorly and build well, and the reverse happens too.

FactSet is the most likely to be treated unfairly by this method. Its investor relations releases could not be opened during research for this post, which leaves its 2024 and 2025 product releases out of the evidence. Its live Workstation may cite as well as either rival. The honest verdict on FactSet is "not shown", not "worse".

Bloomberg also has a structural advantage in this test. BQL is a well-known, exposed query language, so returning it is natural. LSEG and FactSet both have formula and API layers of their own. Their AI material simply does not say whether answers return them.

Two details in this post are secondary. The one-third access share and the source timestamps come from a WSJ report relayed by AI Weekly, and Bloomberg's own press pages could not be opened. Treat both as directional until Bloomberg confirms them.

Finally, the market is moving month by month. LSEG's Copilot availability was "soon" in July. FactSet's Google agents were announced in June. Anything in the access table may be out of date by the next renewal cycle.

Frequently asked questions

How to run a trace test before your next renewal

You do not need a procurement project to compare these tools. You need one afternoon and five questions you already know the answers to.

Score each answer against the four levels of traceability above. Write down the level each tool reaches on each question. That table becomes your evidence when the renewal arrives.

Then send one email to each account team. Ask which subscription tier includes the AI assistant, whether it will be priced separately at renewal, and how citations display when the data is accessed through Copilot, Gemini or MCP. Keep the replies. For the wider vendor questions, use the AI agent procurement checklist as the template.

Related reading