Wealth Management

Client communication in wealth work is supervised communication. A model does not know when a sentence crosses from education into a recommendation. This starter screens the returned wording for that crossing and expects the firm to supervise the rest.

An adviser drafts a market note or a client email, and the model produces advice shaped language without meaning to. Output that promises guaranteed returns is blocked, and output that says buy this stock is blocked too. The required notice states plainly that the content is informational and not financial advice. Suitability and best interest analysis stay with the adviser.

Any answer that raises insider trading, or that names material non-public information, is blocked on return. That screen reads generated text, so it catches a phrase the model added rather than a phrase the adviser typed.

A client national identifier number is masked out of the prompt. The refusal list covers a secret, a client account export and wire instructions. Prompt screening sits ahead of them, which protects the later stages from a crafted instruction.

Scoring blocks any answer under 0.86 overall, under 0.78 for relevance, or under 0.80 for accuracy. Relevance and accuracy contribute in equal parts. Text below 120 characters fails the same check, and so does a single sentence. None of that scoring applies to a streamed answer.

The oversight entry has no configuration of its own, so supervisory routing is work that remains for you. The closing audit marker adds no storage and no retention, so records duties stay with the firm. Supervision under FINRA rules, care obligations under Regulation Best Interest and privacy duties under GLBA remain firm responsibilities. This starter shapes what a model may return, and it never approves a recommendation.