Choosing an AI Copilot for Equity Research

An equity-research copilot should be chosen by the work it can complete with your licensed data, not by the fluency of its chat window. Start with five tasks that use the same issuer, documents, estimates, and internal note. A useful pilot records whether each answer reaches the correct source passage, preserves the as-of date, handles missing evidence, and produces an output an analyst can review. The product category matters less than those results.

Disclosure and evidence mode. This is a documented comparison based on public product pages and documentation checked on August 30, 2026. We did not test every product under common conditions, so this page does not name an overall winner. We build AllMind, one of the systems discussed, and we have a commercial interest in the decision. Competitor behavior described below is vendor-reported unless a source says otherwise. Statements about our own product are first-party claims.

Choose the layer before the product

"AI copilot" now covers at least four different purchases. A terminal assistant answers inside a market-data subscription. A content-search product retrieves filings, broker research, or expert transcripts. A research system connects external content with a firm's own work. A general assistant drafts and reasons over files or connected services. These can coexist on one desk because they control different data and permissions.

The first procurement question is therefore: Which system owns the evidence required for the decision? A general assistant may write the cleanest paragraph while lacking the licensed estimate that makes the paragraph useful. A terminal may hold the number while lacking the internal thesis note that explains why the analyst disagrees.

Public documentation supports these distinctions. Bloomberg says ASKB is in beta and can expose the Bloomberg Query Language code behind data analysis. FactSet describes Mercury as a conversational knowledge engine connected to structured and unstructured FactSet content. AlphaSense describes Generative Search as a source-grounded research surface. Anthropic's finance agents announcement lists finance-specific connectors and templates, but access to a connector does not grant a separate data license.

A five-task pilot that exposes the real differences

Use one liquid company with a recent earnings event. Give each vendor the same authorized packet:

1. the latest 10-Q or 10-K;
2. the latest earnings release and call transcript;
3. a timestamped consensus snapshot from the firm's licensed source;
4. one internal model tab with the analyst's estimate;
5. one internal thesis note containing a view that differs from consensus.

Do not include material the product is not licensed to process. Record product version, operator, timestamp, entitlements, and every prompt. The point is a reproducible procurement test, not a polished demonstration run by the vendor on a curated corpus.

Task 1: recover a financial fact with its context

Ask for one reported metric, its period, unit, accounting basis, and source passage. A pass requires the value and all four qualifiers. A link to the filing cover page is not enough. Open the citation and verify that the highlighted passage contains the number.

For estimate-specific controls, use the field schema in the consensus estimates evaluation guide.

Task 2: reconcile actual, guidance, and consensus

Ask the system to compare the reported result with the timestamped consensus and the prior guidance range. Require it to show the arithmetic. This task catches fiscal-period mismatches, currency conversion errors, and silent use of a current consensus snapshot in a historical question.

Task 3: explain the internal disagreement

Ask why the analyst's model differs from consensus. The answer should quote the internal assumption and connect it to external evidence without erasing provenance. A product that cannot read internal research may still pass the first two tasks and fail here for an entirely legitimate scope reason.

Task 4: run the question across a small coverage list

Repeat one structured question for five issuers. Require fixed columns for issuer, period, result, consensus, variance, source, and unresolved issue. The goal is to test entity resolution and consistent output, not to reward long prose.

Task 5: create a reviewable deliverable

Request a one-page update with citations and a table of changed assumptions. Check whether each claim survives export, whether source links remain usable, and whether an analyst can distinguish retrieved facts from generated interpretation.

Score evidence behavior, not writing style

Give each task 0, 1, or 2 points on the four controls below. Keep the component scores visible. A total alone hides the reason a product failed.

The maximum is 40 points: five tasks multiplied by four controls multiplied by two points. Do not publish cross-vendor scores unless the products actually ran this common test. A procurement team can also set hard stops. For example, any unauthorized-source result or fabricated value can disqualify a product even if the rest of the run is strong.

What the current product families document

Terminal assistants inherit terminal strengths and boundaries

Bloomberg says ASKB can query Bloomberg data, news, research, documents, and analytics, provides attribution, and exposes BQL for data analysis. Its April 2026 roadmap also describes scheduled and trigger-based workflows, proprietary-knowledge integrations, and entitled expert intelligence. Those capabilities make ASKB a logical first pilot for a desk already centered on Bloomberg. The same evidence also establishes its boundary: the experience is part of the Bloomberg environment and remains tied to Bloomberg access.

FactSet's public releases describe a broader AI layer across its Workstation. The company expanded AI-enabled Document Search in beta to more than 85,000 users in March 2026, and its Portfolio Analytics MCP limited release brings governed performance, attribution, and risk outputs into agentic interfaces. A FactSet-heavy buy-side team should test those governed outputs directly, then separately test whether its research-management notes and custom models retain lineage through the same workflow.

Search platforms should be tested on passage retrieval

AlphaSense's value proposition begins with its external content corpus, search, and source-grounded answers. That makes a transcript or broker-research question a fair pilot. It does not establish that every customer's warehouse schema or model assumptions are connected. Ask the vendor to demonstrate the exact internal repositories, sync method, and permission propagation in your proposed configuration.

Research systems should prove the join

AllMind is the system we build. Our financial ontology connects companies, estimates, filings, research, and a firm's internal material, with agents operating under inherited permissions, and our data page names partner data and warehouse connections. Those are our own claims, so hold them to the pilot rather than to this paragraph: verify the hardest join directly, with one licensed estimate, one filing passage, and one internal-model assumption resolved to the same entity and fiscal period. AllMind is quote-based, has no self-serve checkout, and deeper internal-data onboarding requires a data conversation.

General assistants are useful when the packet is controlled

Claude and other enterprise general assistants can be productive for reasoning, drafting, and file-based analysis. Their finance connectors expand the reachable set of services. The procurement trap is to treat reachability as ownership. The firm still needs valid entitlements, and the assistant must preserve the source identifiers and access rules delivered by each connected system.

Rogo documents a finance-specific platform oriented toward analysis and deliverables. Its April 2026 funding announcement is not proof of a research result, so keep the pilot centered on source fidelity and output review rather than company scale.

Questions the vendor demo must answer

Send these in advance and require written answers:

• Which data is included, which data uses the firm's entitlement, and which data cannot be used by the AI layer?
• Can a user open the exact source passage and see the as-of timestamp?
• How are ticker changes, dual listings, fiscal calendars, and restatements resolved?
• What happens when the source set does not contain an answer?
• Does the exported memo preserve citations and calculation inputs?
• Can administrators reproduce who asked a question, which sources were accessed, and what was exported?
• Which capabilities are generally available, beta, limited release, or roadmap items?

These questions also protect against a common comparison error: treating a vendor's announced roadmap as current product availability.

What public evidence cannot settle

Public pages cannot establish latency under a firm's entitlements, retrieval quality on its documents, or accuracy on its fiscal-period conventions. They also do not reveal the total contract cost after data feeds, implementation, storage, and support. Those items require the common pilot and a written commercial proposal.

The sensible outcome may be two products. A terminal can remain the source of record while a research system connects internal work and automates a coverage task. A general assistant can handle bounded drafting. Consolidation is valuable only when the replacement passes the same evidence controls as the system being removed.

Sources and methodology

• Bloomberg AI and ASKB product page, product scope and attribution claims, accessed August 30, 2026.
• Bloomberg ASKB roadmap, April 16, 2026, beta status and planned workflows.
• FactSet Intelligent Platform announcement, Mercury and internal research integration.
• FactSet AI-enabled Document Search announcement, March 26, 2026 beta rollout.
• AlphaSense Generative Search product article, public description of source-grounded search.
• Anthropic finance agents announcement, finance templates and connectors.

If you run this pilot with AllMind, bring the five authorized artifacts and the expected answer for each task. Keep the raw runs, failed citations, and score components so the decision can be reviewed after the demo.