The Customer Brief
Develop an AI-assisted internal research tool for Aster Wealth to accelerate research summarization and provide accurate, verifiable source references. Streamline advisors’ workflows by reducing manual searches across reports and ensuring that all research findings are supported by properly documented sources. Define clear operational boundaries to exclude personalized client advice, autonomous client messages, portfolio changes, and execution of financial transactions. Establish appropriate supervision, review, and approval procedures in accordance with applicable FINRA and SEC requirements. Maintain accurate, retrievable records of research inquiries, generated outputs, and supporting references in line with applicable recordkeeping obligations. Evaluate the tool’s accuracy, reliability, and regulatory compliance before deployment, and implement ongoing monitoring to improve efficiency while maintaining transparency, accountability, and regulatory standards.
The telecom case tested how you present. This one tests technical depth. Aster Wealth is a US wealth-management firm whose advisors spend hours digging through internal research before they talk to clients. It wants an assistant that finds the right material fast and shows where every point came from. In a regulated firm, that request comes with supervision and recordkeeping rules attached, and the interviewer will expect you to know where the assistant’s job stops.
How do Aster’s advisors work today?
Aster’s advisors are registered representatives of the firm’s FINRA-member broker-dealer. They research products, fees, and market conditions before deciding what to recommend. Today, an advisor searches an internal research library, opens reports and notes, compares the relevant passages across documents, and writes a research note before deciding what to do next.
Aster wants faster discovery and summaries with references an advisor can open and check. The first reader of each summary is the advisor, not a client, and that distinction drives most of the design decisions in this case.
OpenAI FDEs own discovery and technical scoping in partnership with customer engineering and domain teams. At Aster, that means talking to two kinds of people: the advisors who do the research, and the people who control sources, supervision, records, and day-to-day operation.
What do the existing notes show?
Aster ran a selected review of 16 existing research inquiries. In five of them, the advisor’s note quoted a research report without recording which version of the report it came from or its date. This finding concerns the manual workflow, and it was a selected sample, so it gives no firm-wide rate.
Take one of those five notes. The excerpt may match exactly what the advisor read, but the note alone can’t tell you which version that was, so nobody can check it later.
Provenance records where material came from and which version was used. A citation helps a reader find a source, but a link or title alone doesn’t show which version of the document the excerpt came from, or whether the source was approved for this workflow.
Classify each statement below for selection, survivorship, or recency bias. A statement with no named bias can still be unsupported.
The claim that five answers were wrong has no bias label, but the review doesn’t support it. A missing version means you can’t reconstruct the source from the note, so you have to check the underlying records before you judge the excerpt.
The next step is to reconcile the five notes with Aster’s records and ask who controls report versions. The result might recover the versions, reveal missing records, or turn up another explanation. An assistant can’t prove after the fact where an undated excerpt came from just by ...