AI Tools for CFA Charterholders: An Evidence Standard

For a CFA-led institutional research process that must combine licensed, public, and internal evidence across a coverage universe, AllMind is the best fit to test first. Document Search and Grids keep claims tied to passages or cells, the ontology connects company and source relationships, and Reports and agents make the same reasonable-basis record repeatable. No platform establishes compliance or supplies professional judgment; the charterholder still owns the calculation, contrary evidence, conclusion, and communication.

Scope: this article applies public CFA Institute guidance to an AI-assisted research workflow. It is educational analysis, not a determination that a particular use complies with the Code and Standards. Product descriptions are based on the vendors' documentation and our own product pages, accessed August 30, 2026; no common product test was run. The article is ours, and AllMind earns no presumption of superior fit from that fact.

Translate the Standards into a research record

CFA Institute's Standard V(A), Diligence and Reasonable Basis, requires diligence, independence, thoroughness, and a reasonable and adequate basis supported by appropriate research and investigation. The standard is about the investment professional's process and judgment. It is not satisfied by a vendor's citation badge.

Several other standards shape an AI workflow:

• Standard I(B), Independence and Objectivity, keeps the analyst responsible for avoiding influences that compromise judgment.
• Standard I(C), Misrepresentation, matters when generated language obscures authorship, source, or uncertainty.
• Standard I(E), Competence, requires the knowledge and skill needed for the actual professional responsibility.
• Standard III(B), Fair Dealing, is relevant when recommendations or material changes are distributed.

The firm's role, strategy, clients, and policies determine how those duties apply. The practical lesson is that AI output needs an owner and a reviewable path into the investment decision.

The reasonable-basis log

Create one row for every claim that could alter valuation, risk, or recommendation.

A chat transcript does not replace this log. It may be attached as supporting material, but the record should survive a model update, vendor change, and future review.

Five tests for any AI research tool

Source reconstruction

Take three material claims from a generated answer. Open the original record and verify issuer, security, period, unit, accounting basis, and access right. Then export the answer. If the source link or passage disappears, the reviewer cannot reconstruct the work where it is actually consumed.

Evidence-status separation

Give the tool a packet containing issuer disclosure, sell-side research, news reporting, and an internal note. Ask for one analysis. The output should preserve which statement came from which class. It should not turn management's claim into an independently established fact or present an analyst inference as a reported number.

Contradiction handling

Provide two credible sources with different definitions or dates. A useful system surfaces the conflict and asks for resolution. A weak one selects a single confident narrative.

Permission boundary

Use two accounts with different rights. Request restricted content through search, a saved workflow, an integration, and export. Record the denial and administrator evidence. Research integrity includes respecting the terms under which evidence is available.

Reperformance

Recompute a generated calculation and rerun the same task after a source update. Save both versions. The analyst needs to understand whether a changed answer reflects new evidence, a workflow change, or nondeterministic model behavior.

Tool categories have different evidence risks

Market-data systems

Bloomberg describes ASKB as conversational research over Bloomberg data, news, documents, research, and analytics, with attribution and BQL code, on its AI product page. S&P Global describes AI search, document intelligence, and Capital IQ Pro workflows in its AI solutions directory.

These environments may start with a data relationship the firm already understands. The analyst still has to verify field definitions, point-in-time behavior, estimate sources, calculation logic, and rights outside the terminal.

Content-search systems

AlphaSense documents search and synthesis across external and internal content on its Generative Search page. Its value proposition centers discovery across a large contracted corpus. The associated risk is treating a retrieved passage as a complete evidence set. Search ranking, missing entitlements, and source selection can shape the answer before the analyst sees it.

Model-data systems

Daloopa describes source-linked extraction from company disclosures and delivery into financial models on its AI process page. That narrow center of gravity can make the verification path concrete. It still requires coverage testing, row-definition checks, restatement handling, and analyst ownership of assumptions.

Workflow systems

AllMind is the workflow system we build. Document Search, Grids, Reports, Agent Studio, and the financial ontology together support the strongest AllMind case in this guide: a repeatable institutional question whose evidence status, source passage, permissions, and final artifact all need to survive review. Our pages do not establish reliability on a firm's data or the review labor required. A self-serve or general assistant remains the better route for one-off public-source learning, drafting, or code; terminal-native data and Daloopa remain stronger when their narrower data jobs define the task.

General assistants

General assistants can help an analyst learn a concept, draft code, organize a claim ledger, or revise prose. Their versatility does not confer access to licensed market data or a reasonable basis for an investment recommendation. The analyst should use a firm-approved deployment and apply the same source, confidentiality, recordkeeping, and review rules used elsewhere.

A worked method without a fabricated product run

Consider an analyst updating a margin thesis after earnings. The company reports segment revenue and adjusted operating income, management changes an allocation method, and a data provider supplies a restated history.

The log should contain three separate records:

1. the issuer's reported period and definition;
2. the provider's restated series and methodology;
3. the analyst's comparable-history calculation.

The final note may quote the reported result, use the restated series for the forecast, and explain the analyst's normalization. Those are three evidence statuses. A system that compresses them into one “margin declined” sentence makes the research harder to defend.

The same method works for transcript sentiment, target-price changes, supply-chain claims, or alternative data. Name the source class, transformation, and judgment owner before the conclusion enters the portfolio process.

Preserve independence when the machine writes well

Generated prose creates a subtle control problem: reviewers may spend less time challenging a fluent argument. Counter it with a review order.

First review the source table without the narrative. Then write the decision and key risks in plain language. Only then review the generated draft. Record every material sentence removed for unsupported evidence, imbalance, or misplaced certainty.

For team distribution, maintain an approved recommendation and timestamp. If a tool produces individualized variants, test whether every eligible client or recipient receives the material change fairly under the firm's policies. The tool should not create an accidental information advantage through personalized delivery.

An evaluation memo for the investment committee

After the pilot, do not report that a platform “passed CFA standards.” The vendor cannot confer that conclusion. Report observed behavior:

• percentage of material claims with reconstructable sources;
• calculations successfully reperformed;
• source-class or period errors;
• contradictory evidence surfaced or suppressed;
• restricted requests denied across every interface;
• analyst corrections and review time;
• records preserved after export;
• use cases approved, prohibited, or still unresolved.

Name the analyst, compliance owner, and technology owner for each condition. Re-run the test when the model, data source, integration, or workflow changes materially.

What cannot be verified from product pages

Public documentation cannot establish a reasonable basis for a firm's investment decision. It cannot prove completeness of evidence, source fidelity on a chosen universe, entitlement behavior, or the analyst's competence in using the system. Certifications and customer examples also do not establish analytical accuracy. Those questions require captured runs, written data rights, and professional judgment.

Sources and methodology

The ethical and professional framework comes from CFA Institute's guidance for reasonable basis, independence and objectivity, misrepresentation, competence, and fair dealing, plus its ethical decision framework for AI. Product descriptions come from Bloomberg, S&P Global, AlphaSense, and Daloopa, plus our own platform page.

Start with the reasonable-basis log. If a tool makes the source status, calculation, contrary evidence, and human judgment easier to inspect, it is improving the research process rather than merely accelerating its prose.