Investment & financial research assistant (filings, earnings calls, broker research)

Equity analysts, private-equity and venture teams, corporate-development and treasury staff spend most of their time finding, reading and re-keying numbers from filings, call transcripts and broker notes. A finance-specific research assistant searches a licensed corpus, answers with citations to the exact passage, fills cross-company comparison grids and tracks KPIs across quarters. It accelerates the reading and spreading work; the investment judgement and any decision stay with the human.

Problem & value

• Today: analysts read hundreds of pages per company per quarter (annual report, 10-K/10-Q, ESEF reports, call transcripts, broker notes), copy figures into Excel, and rebuild the same peer comparisons before every earnings season, IC meeting or pitch. Content is spread over Bloomberg/LSEG/FactSet terminals, broker portals, expert-network transcripts and internal shares.
• AI value: natural-language search across all of it, answers that link to the source sentence, multi-document extraction into grids, and alerts when management language or a KPI changes. Typical gain is hours saved per company per quarter; the vendors above do not publish audited, comparable figures, so treat any quoted "x% time saved" as marketing until you measure it in a pilot.
• Accuracy is the real constraint: on FinanceBench (10,231 open-book questions on public companies) a GPT-4-Turbo + retrieval setup answered incorrectly or refused about 81% of questions in 2023 [3]. Agents have improved a lot, but on the Vals AI Finance Agent benchmark (537 entry-level-analyst questions) the best model scored 64.37% as of June 2026 [2]. Numeric grounding, citations and review are therefore design requirements, not extras.
• Not the same as neighbours: KNOW-015 does open-web desk research; DOC-012 spreads one statement at a time; this use case adds licensed premium content, multi-document numeric grids, transcripts and sentence-level citations.

Variants

• Public-markets research (buy/sell side): coverage universes, earnings-season workflows, consensus vs. actual, tone and guidance tracking across call transcripts.
• Private equity / VC / M&A: data-room Q&A, due-diligence checklists, comp sets with PitchBook-type data, CIM and teaser drafting (human-reviewed).
• Corporate development, treasury, CFO office: competitor and customer monitoring, counterparty/bank-credit reading, peer benchmarking from annual reports.
• Investor relations: the reverse view, preparing Q&A from analyst questions and competitor calls.
• Wealth/asset managers and family offices: summarising house research and fund documents for advisers (note: personalised recommendations to clients are a different, regulated activity).
• Private investors: retail apps that explain filings and reports; useful for reading, but not investment advice and regulated if framed as personal recommendation.

How it works

1. sources (licensed feeds, EDGAR/ESEF filings, transcripts, broker PDFs, internal files)
2. ingestion + entitlement tags (who may see which licence/restricted list)
3. parsing (layout/table extraction, XBRL facts, speaker-separated transcripts)
4. hybrid index (BM25 + embeddings + reranker, metadata: company, period, doc type)
5. agent loop (plan
6. retrieve
7. read tables
8. compute
9. verify)
10. answer / grid cell with sentence-level citation + numeric check against structured data
11. human review (analyst)
12. export (Excel, PowerPoint, memo)
13. archive

• Structured first, text second: filings in XBRL (SEC company-facts and frames APIs are free and key-less [4]; EU issuers file in ESEF/iXBRL) should supply numbers directly; the LLM only explains and compares. Computations run in code, not in the model's head.
• Grid pattern: rows = companies, columns = questions ("net leverage", "segment guidance", "change of auditor"); each cell is an independent retrieval + extraction job with its own citation. Hebbia's Matrix popularised this layout [6].
• Citation verification: a second pass checks that the quoted span exists in the cited document and supports the number; failures are flagged, not hidden.
• Human-in-the-loop: analysts approve before anything leaves the firm; output is labelled as AI-assisted draft; investment recommendations are never auto-published.

Data preparation

Data needed: licensed filings, earnings-call transcripts and broker research, plus internal models, deal notes and investment memos. These live with data vendors, on research portals, in e-mail, in SharePoint or file shares and in analysts' Excel models.

Typical gaps: PDFs with tables that break text extraction, research behind licences that forbid indexing or AI use, inconsistent company identifiers and tickers, outdated versions of internal models, notes without date or deal tag, and unclear access rights between deal teams (information barriers).

Preparation steps: clarify per source whether licence terms allow AI processing ([PREP-025]); extract and structure tables with an OCR and layout pipeline ([DOC-014]); map companies to one identifier and add metadata such as period, document type and sector ([PREP-010], [PREP-012]); restructure Excel models into tidy tables ([PREP-011]); transfer permissions so the assistant respects barriers and need-to-know ([PREP-023]); build a test set of 50 to 100 analyst questions with verified answers ([PREP-018]). Tools: a document parser, a vector database, the existing DMS.

Plan about 40 percent of the effort for data work, much of it licence clarification and access rights rather than technology.

Implementation paths

Which path for whom

• smb: A - small advisers and corporate-development teams get licensed content plus assistant bundled from data vendors; building their own would not pay off.
• enterprise: B - banks, asset managers and PE houses mostly already hold data licences, so a retrieval layer on cloud APIs in an EU region lets them bring their proprietary notes and models into the loop.
• developer: B - integrators build on cloud LLM APIs and the customer's licensed corpora; self-hosting needs HW3 hardware and rarely pays off.

Path A — Buy

• AlphaSense: Generative Search and Deep Research agents over a library of "500M+ premium documents", Excel/PowerPoint add-ins, enterprise-intelligence over internal files [1].
• Hebbia (Matrix): multi-document grids and decks; connects to SEC filings, transcripts, FactSet, S&P Capital IQ, PitchBook and SharePoint/Box [6].
• Rogo: finance-specific agents for banking and investing; lists SOC 2 Type II, ISO 27001 and ISO 42001, and 450+ client firms [7].
• Quartr (Sweden, EU): first-party investor-relations data (live calls, transcripts, slides, filings), AI chat with citations, an API and an MCP connector for Claude and other tools [8].
• Incumbent terminals: Bloomberg, LSEG, FactSet and S&P Capital IQ ship their own AI assistants; the pragmatic choice if you already pay for the terminal.
• Foundation-model suites: Claude for Financial Services offers pre-built MCP connectors to FactSet, S&P Global (Capital IQ, transcripts), PitchBook, Morningstar, Daloopa, Box, Databricks and Snowflake, and states that data is not used for training by default [5].
• Choose A when: you already license the content (licensing is the moat, not the model), have fewer than ~100 users and need value within weeks. Check EU data residency, DPA, SOC 2/ISO 27001, entitlement mapping and exit options.

Path B — Build on cloud APIs

• Models: a frontier reasoning model with tool use (Claude Opus/Sonnet, GPT, Gemini) for agent steps; Mistral (EU provider) where EU-provider status is required. Use EU regions via AWS Bedrock (Frankfurt), Azure OpenAI (EU data zone) or Google Vertex AI (EU).
• Stack: MCP connectors to licensed data, hybrid search (OpenSearch/Elasticsearch or pgvector + BM25), reranker, document parser (Azure Document Intelligence, Mistral OCR, Docling) and a deterministic calculation tool (Python sandbox).
• Data: SEC EDGAR APIs (free) [4], national registers and ESEF reports from issuer sites/OAMs; ESAP, the EU single access point for public company data, is being phased in and should become a unified source over the coming years.
• Choose B when: you need custom grids, integration with an internal deal database or model library, and have 2-4 engineers.

Path C — Self-host (open source)

• Models: open-weight reasoning/instruction models (Qwen3, Llama, Mistral/Mixtral families, DeepSeek) at 30-120B on a single pro GPU (HW3) for sensitive deal material; FinBERT-class encoders for sentiment/classification; BGE or multilingual-e5 embeddings with a cross-encoder reranker.
• Stack: vLLM or SGLang for serving, Qdrant/pgvector/OpenSearch, Docling or Unstructured for parsing, LangGraph or similar for agents; FinGPT (MIT licence, academic-purpose notice) offers fine-tuning recipes and sentiment models but is a research toolkit rather than a product [9].
• Limits: licensed broker and vendor content usually may not be pushed into your own index without an explicit licence; open-weight models are weaker at long numeric reasoning than frontier models, so keep calculations in code.
• Choose C when: deal confidentiality, MNPI handling or regulators forbid third-country processing.

Path D — Custom training

Rarely worth it for the whole assistant. Fine-tuning pays off for narrow components: a reranker or embedding model on your own query/click logs, a classifier for document types or guidance statements, or a domain extraction model for sector-specific KPIs (e.g. insurance combined ratio, SaaS ARR). Needs a few thousand labelled examples and an evaluation set built from your own past analyst questions.

State of the art (2026)

• Agentic research is the standard pattern: planners that issue many searches, read tables and recompute figures. Top accuracy on the Vals Finance Agent benchmark (v1.1) was 64.37% (Claude Opus 4.7), with Claude Sonnet 4.6 at 63.33% as of June 2026 (newer benchmark versions and models have since changed the ranking); the original benchmark paper reported 46.8% for the best model at an average of $3.79 per query [2][10]. Accuracy rises with number of tool calls, and cost per question matters.
• Connector ecosystems (MCP) let one assistant reach FactSet, S&P, PitchBook, Quartr and internal stores without copying data [5][8].
• Excel-native agents that build and audit models are a fast-moving area; Anthropic reports a partner agent passing 5 of 7 levels of the Financial Modeling World Cup [5].
• Failure modes: hallucinated or mis-scaled numbers (units, thousands vs. millions), period/restatement mix-ups, fiscal-year confusion between peers, citing a passage that does not support the claim, over-reading management tone, stale data, and prompt injection from external documents. Never use model output as the sole basis for a trade or a published recommendation.

Shared resources & combinations

• Shared building blocks: hybrid index (vector-db, search-engine, reranker), document parsing (ocr-docparse), the agent framework with MCP connectors, entitlement-aware access control, and a code sandbox for calculations. A HW3 GPU or a private EU cloud endpoint serves all of these.
• KNOW-015 (open-web deep research) supplies news and web context; KNOW-002 (company-document RAG) covers internal memos; KNOW-013 (GraphRAG) adds supplier/ownership relationship graphs.
• DOC-012 (financial statement spreading) and DOC-014 (table extraction) provide the numeric extraction layer; PRED-009 (credit scoring) and OPT-010 (portfolio optimisation) consume the structured outputs but are separate regulated systems.
• ANOM-012 (market-abuse surveillance) shares the transcript/news ingestion and is the compliance counterpart; AGENT-002 runs scheduled earnings-season workflows.

Effort, cost & KPIs

• A: weeks to pilot, per-seat licences (typically enterprise pricing, quote-based) plus content fees; the largest cost is data licensing.
• B: 2-4 engineers for 2-4 months; costs are model tokens (agentic runs can cost dollars per complex question, see [10]), parsing and index hosting.
• C: one HW3 GPU server plus ops; 1-2 engineers ongoing; licensed content integration is the hard part.
• D: only for components; adds labelling effort of weeks.
• KPIs: citation precision (share of claims whose cited span supports them), numeric accuracy on a gold set of 100-200 own questions, time per company-quarter update, analyst adoption, share of answers needing correction, cost per answer, and incident count (leaked restricted-list names, wrong figures reaching clients).

People & change

Involve management and the research head from the start, then the analysts as daily users, compliance for information barriers and licence terms, IT security for access and logging, and the DPO if client or personal data appear ([GUIDE-005], [GUIDE-013]). Data owners at the content vendors are part of the contract discussion ([GUIDE-010]).

Works council: depends. Where usage is logged per analyst, or search queries could show performance, §87(1) no. 6 BetrVG applies under the objective-suitability test of the BAG, so agree the logging scope in a works agreement ([GUIDE-004]). Inform the council in good time under §90(1) no. 3 and consider an AI expert under §80(3).

AI literacy: employers must take measures that support staff AI literacy under Art. 4 AI Act; offer analysts training on citations, hallucinated figures and checking against the source ([GUIDE-006]).

Pilot and concerns: choose a small coverage universe and let two or three analysts compare answers against their own work. Typical concerns are wrong numbers in client documents, job loss and regulatory exposure. Address them by requiring clickable source citations, a human sign-off on everything that leaves the firm, and by presenting the tool as a way to cover more companies with the same team ([GUIDE-008], [GUIDE-009]).

Risks, regulation & compliance

• EU AI Act: a research assistant for professionals is generally minimal risk; Art. 4 AI literacy applies. It becomes high-risk only if reused for Annex III 5(b) creditworthiness evaluation of natural persons or life/health insurance risk and pricing [11], which this use case does not do. Chatbot-style interfaces to external users trigger Art. 50 transparency.
• MAR (Regulation 596/2014): research tools must not become a channel for inside information (Art. 7, 8, 10). Keep MNPI sources (deal rooms, restricted lists, wall-crossed material) in separate, access-controlled workspaces; log queries and keep insider lists (Art. 18) consistent with who can query which data.
• MiFID II research rules and investment-recommendation rules: AI-drafted research distributed to clients must meet the same disclosure, conflict and approval requirements as human-written research; AI does not remove the author's responsibility. For personalised advice, suitability rules apply.
• Licensing and copyright: broker research, terminal data and transcripts are licensed; check whether AI processing, embedding and storage are permitted, and what the vendor's contractual redistribution limits are. EU text-and-data-mining rules do not override licence terms for paid content.
• DORA: financial entities using third-party AI/cloud providers must register them as ICT third-party providers, assess concentration risk and plan exit.
• GDPR: transcripts and expert calls contain personal data of executives and experts; define lawful basis, minimise, and honour expert-network confidentiality terms.
• Mitigations: citations with verification, human sign-off, calculation in code, entitlement-aware retrieval, query logging, red-teaming for prompt injection from filings and web content.

Getting started

1. Pick 20-30 real past analyst questions (with known answers) across filings, calls and broker notes; this is your evaluation set.
2. Inventory content and licences: which sources may be processed by AI, by whom, and where.
3. Trial one Path A product (e.g. AlphaSense, Hebbia, Rogo or Quartr) and one connector-based setup (Claude or equivalent with FactSet/S&P/Quartr connectors) against the evaluation set.
4. Define the control layer: citations required, numbers recomputed in code, MNPI workspace separation, logging, sign-off before external use.
5. Roll out in one earnings season with a single team; measure KPIs before extending.

Sources

1. AlphaSense platform overview (Generative Search, Deep Research, content library) - https://www.alpha-sense.com/platform/ - accessed 2026-10.
2. Vals AI, Finance Agent benchmark v1.1 (537 questions; scores as of 2026-06-04) - https://www.vals.ai/benchmarks/finance_agent - accessed 2026-10.
3. Islam et al., FinanceBench: A New Benchmark for Financial Question Answering, arXiv:2311.11944 - https://arxiv.org/abs/2311.11944 - accessed 2026-10.
4. U.S. SEC, EDGAR application programming interfaces (submissions and XBRL APIs) - https://www.sec.gov/search-filings/edgar-application-programming-interfaces - accessed 2026-10.
5. Anthropic, Claude for Financial Services (connectors, features) - https://www.anthropic.com/news/claude-for-financial-services - accessed 2026-10.
6. Hebbia (Matrix) - https://www.hebbia.com/ - accessed 2026-10.
7. Rogo - https://rogo.com/ - accessed 2026-10.
8. Quartr (IR data, API, MCP) - https://quartr.com/ - accessed 2026-10.
9. AI4Finance Foundation, FinGPT repository - https://github.com/AI4Finance-Foundation/FinGPT - accessed 2026-10.
10. Finance Agent Benchmark paper, arXiv:2508.00828 - https://arxiv.org/abs/2508.00828 - accessed 2026-10.
11. EU AI Act, Annex III (point 5(b), 5(c)) - https://artificialintelligenceact.eu/annex/3/ - accessed 2026-10.