Investment Research Copilots: Source Versions, Financial Calculations and Review

Define the research-preparation task

A research copilot can prepare material from approved statements, earnings documents and research notes. Define an output, such as a period comparison or a summary of management statements, then identify sources and calculations.

Analyst judgment, report approval and publication retain their respective authorities. An extraction draft is not approved analysis. Efficiency comparisons must include checking, calculation, corrections and review.

Compare source paths, not model brands

General-purpose models can process supplied documents; configured retrieval can support multi-source lookup. Either can omit, misinterpret or produce unsupported material. Anthropic's PDF documentation describes supplied text/page images and limits. Microsoft's RAG overview describes retrieval design choices.

Index presence does not guarantee retrieval, parsing or applicability. Upload, processing completion and search visibility differ. Microsoft's indexer scheduling documentation explains these processing boundaries.

Retain reconstructable evidence

For sources, record entity, document type, version, date, accounting period and permitted audience. Check company/consolidation scope, currency, units, amendments and restatements. Do not subtract periods with incompatible definitions.

Separate output into three classes:

• Source-reported material: retain the value/text, location and context; check parsing completeness.
• Derived calculations: retain inputs, sources, formula, units and comparison basis; recompute.
• Analysis/inference: identify analyst assumptions, limits and questions, distinguishing management statements from independent judgment.

An older framework or management's causal narrative may be relevant evidence, but a citation does not establish a current conclusion. Assess support for the specific claim, period and use.

Test missing, conflicting and denied evidence

Build cases with adequate evidence, missing data, conflicting versions, inconsistent units, retrieval misses and denied access. Separately assess coverage, extraction, calculations, citation support, omissions and review work.

A missing hit can reflect coverage, parsing, query or access failure. Describe the pending state within authorized context rather than asserting that evidence does not exist. A retrieved passage still needs support checks. Evaluate refusal too; refusal on one task does not validate other answers.

Anthropic's hallucination guidance offers grounding/citation checks without eliminating errors. Microsoft's access overview describes identity, permissions and query enforcement that must be implemented. Permission to read research is not automatically permission to transfer, publish or transact.

Assign version and failure ownership

Record source, processing and model versions for analyst review. Assign source-update approval, parsing/concurrency failure handling and fallback to direct documents. Retain correction reasons and publication decisions. Links and usage frequency do not replace analysis approval.

NIST AI RMF (January 26, 2023) is a voluntary risk/monitoring framework, not research regulatory clearance. Sources reviewed October 4, 2026. Microsoft's RAG, access and scheduling dates are August 4, September 17 and July 8, 2026; Anthropic displays no update date.

Check knowledge-source readiness, then discuss existing sources, access and evaluation through the RAG service.