AI in investment research: verify sources, data, and model output

AI is less suitable as the final authority for live prices, legal interpretations, tax conclusions, transaction instructions, or investment recommendations when the underlying evidence has not been independently verified.

Before a generated number enters a model or thesis, identify the exact reporting period, fiscal calendar, currency, scale, and whether the figure is GAAP, non-GAAP, reported, adjusted, or calculated. A value labeled “revenue” can refer to a quarter, year-to-date period, trailing twelve months, segment, or consolidated total; the label alone is not enough.

Tables create another failure point. Models can shift columns, drop negative signs, confuse percentages with basis points, or read a subtotal as a total. Reconcile material values to the source table, then independently recompute ratios such as margins, growth rates, per-share values, leverage, and free-cash-flow conversions.

• Fabricated citations: a plausible source name or quotation may not exist, or the linked document may not support the claim.
• Stale information: prices, guidance, executive roles, share counts, rules, and market conditions can change after the model’s knowledge or retrieval timestamp.
• Retrieval gaps: the model may see only part of a filing, miss an exhibit, or fail to retrieve the footnote that changes the interpretation.
• Bias and concept drift: skewed training data, historical relationships, or outdated classifications can distort the answer as market conditions evolve.
• False precision: a detailed forecast can look rigorous even when the underlying inputs are uncertain or unsupported.

The practical response is source verification, not confidence scoring alone. A material claim should be traceable to the exact document, page or table, period, unit, and timestamp that support it.

Reproducibility requires more than saving the final prose. Preserve the tool or model, date, material prompt, documents supplied or retrieved, and any filters applied to the source set. When a result matters to valuation or risk, run a second formulation or independent check to see whether the conclusion depends on wording rather than evidence.

Retrieval-augmented systems introduce their own risks: the search layer can rank an irrelevant document highly, exclude a key filing, or retrieve an old version. The model can only reason over what it receives, so corpus completeness belongs in the review.

Do not place material nonpublic information, client data, account credentials, confidential company documents, proprietary datasets, or restricted research into a tool unless the organization has approved the tool and its data handling. Review vendor retention, training-use, access-control, logging, export, and deletion terms rather than assuming a chat interface is private by default.

Cybersecurity risk also extends to connected tools. A system with access to cloud drives, email, data terminals, or code repositories can expose more than the text entered into a prompt. Permissions should match the minimum data and actions required for the research task.

Key risks include acting outside the intended scope, executing a multi-step action that is difficult to trace, using sensitive data in an unintended tool, or making a decision without sufficient domain knowledge. Human approval should remain explicit before consequential actions such as sending external communications, modifying research records, changing account settings, or transmitting orders.

The audit trail should show what the agent was allowed to access, what action it proposed, what evidence supported the action, who approved it, and what actually occurred.

Related learning: Research setup · Reading 10-K & 10-Q · Research and testing.