How to Evaluate AI-Assisted Work in Asset Management

Start with the investment-work decision

Name the precise decision supported by the work. Examples include whether a research memo is ready for internal challenge, whether a portfolio-monitoring exception needs investigation, or whether a client communication accurately reflects approved source material. Record audience, time horizon, relevant mandate, source cutoff, consequence of error, and accountable reviewer.

The US Securities and Exchange Commission's Division of Investment Management describes its remit as protecting investors and promoting informed investment decisions through oversight of investment companies and advisers (Division of Investment Management). A 2026 SEC speech on AI and investment management discusses opportunities alongside risks such as conflicts, opacity, data quality, and governance (SEC speech). These sources establish a current regulatory context, not a legal conclusion about a particular workflow.

Sources: [1] [2]

Build a provenance chain for every material statement

Separate source facts, calculations, assumptions, model-produced transformations, and analyst judgments. A citation should resolve to the exact dated document or dataset version used. For numerical work, retain input fields, formulas or code, unit conversions, and reconciliation checks. For narrative work, mark unsupported statements and distinguish a source summary from an interpretation.

Record the tool, plan, permissions, model or feature version when available, and relevant configuration. A capability observed in one environment should not be generalized to another. If the system cannot expose an adequate record, design an external evidence log or keep the workflow out of scope.

Worked research-memo example

Imagine a synthetic internal memo comparing two listed infrastructure businesses. The AI assistant produces a concise margin comparison and claims that Company A has more stable cash generation. The evidence packet contains two annual reports, a dated market-data export, a spreadsheet calculation, and the generated draft.

The reviewer traces each reported figure to the source period, discovers that one margin uses adjusted earnings while the other uses a statutory measure, and marks the comparison as non-equivalent. The stability claim has no defined metric or observation window. The repaired memo labels the measures, recalculates a comparable series where possible, states the remaining limitation, and routes the interpretation to the accountable analyst.

The evidence artifact includes a claim ledger with source, date, transformation, assumption, verification, and reviewer disposition. The exercise assesses evidence navigation and verification in a bounded task. It does not determine whether either security should be purchased or whether a person is suitable for a role.

Review controls and failure modes

Test stale data, corporate-action adjustments, survivorship effects, inconsistent currency, missing fees, look-ahead information, fabricated citations, and authoritative language unsupported by evidence. Include a scenario where the assistant declines or omits a relevant source, not only one where it gives an incorrect answer.

NIST's AI RMF Core recommends measuring systems in context, documenting limits, evaluating human-AI configurations, and maintaining monitoring and feedback processes (AI RMF Core). Apply those ideas to the specific workflow and consequence. Do not turn a generic checklist into a sector-wide assurance claim.

Sources: [3]

Asset-management evidence checklist

This method supports an evidence-based review of AI-assisted work. It does not establish regulatory compliance, investment performance, or the validity of a product across every asset-management context.

• The task, audience, mandate context, cutoff date, and reviewer are explicit.
• Material facts resolve to dated primary sources.
• Calculations preserve inputs, transformations, units, and reconciliations.
• Assumptions and analyst judgments are separated from source facts.
• Tool version, permissions, and configuration are recorded where relevant.
• Conflicts, uncertainty, missing sources, and model limitations remain visible.
• Accountable qualified reviewers retain the final decision.