A sensible first assignment is organizing public educational material. Give an agent a small set of documents and ask for a comparison of their explanations, with the source beside each factual point. Do not begin with a client’s portfolio or a request for a personalized recommendation.
For example, compare two public explanations of an investment account’s features. Ask the agent to record the stated eligibility, fees, restrictions and publication date, leaving missing fields blank. The output is a research aid for a professional to check, not advice to a client.
Define the brief before supplying documents
Specify the reader and purpose. A staff training note, a client-facing educational draft and a recommendation involve different obligations and review. The agent should not choose among them silently.
List the sources it may use and the fields to extract. Require a source location for every figure and condition. A number without its date, unit or qualifier can be misleading even when copied correctly. If two documents disagree, preserve the disagreement rather than asking the model to pick a winner from memory.
Run a public-source exercise
“Using only these two public documents, create a comparison table of stated features. Include dates, fees and restrictions exactly as stated and link each item to its source. Mark absent information as ‘not stated.’ Do not recommend an option, infer client suitability or create a client communication.”
Open the cited sections and verify the fields. Check that introductory rates have not become permanent rates, that an example return has not become a forecast, and that a restriction has not disappeared in the summary. These are proposed review checks, not observed failures from a product test.
After the table is correct, ask for a plain-language explanation of one difference. Keep material conditions in the explanation. Shortening a paragraph is useful only when a beginner still understands what affects the decision.
Follow the rules that apply to your organization
FINRA’s Regulatory Notice 24-09 reminds its member firms that existing obligations apply when they use generative AI, including relevant supervision and communications requirements. That is a US member-firm scope, not a claim that every financial planner is regulated by FINRA.
Use your firm’s policies and the authorities applicable to your work. Get the necessary review before introducing a tool into client communications or records. Do not treat a generic approval prompt, privacy setting or vendor assurance as evidence that the whole workflow meets those obligations.
Keep client access out of the first test
Public documents are enough to test extraction and explanation. Client identity, holdings, account numbers and confidential notes introduce additional duties. Do not upload them merely to make a demonstration more realistic.
If a later approved pilot uses sensitive material, define the system boundary, permitted inputs and reviewers in advance. Preserve what the agent was asked, which sources it used and which human decisions followed. Avoid claiming that a conversation is a compliant record unless your organization has established that it is.
Evaluate usefulness honestly
Count the fields that were extracted accurately, the missing qualifiers and the time spent reviewing. Keep rejected sentences with a reason. If the agent writes beautifully but needs every condition restored, change the task to extraction only.
A successful first exercise gives you a table and explanation you can trace back to public sources. Moving from that result to individualized advice, client messaging or transactions requires a separate professional decision. None of those actions is performed or endorsed by this guide.
