Company case
Morgan Stanley: RAG / knowledge assistant
What the published sources confirm
Morgan Stanley Wealth Management deployed an internal GPT-4-based assistant ("AI @ Morgan Stanley Assistant") that answers financial advisors' questions from the firm's internal knowledge, and later a meeting-summary tool ("AI @ Morgan Stanley Debrief"). OpenAI's case study describes evaluation of use cases before rollout.
Business problem and AI concept
- Problem
- Advisors needed fast, trustworthy answers from a very large body of internal research and procedures.
- Concept
- Retrieval-augmented generation over permissioned internal documents, with expert-written evaluation questions.
Reported results
OpenAI's case study states that over 98% of advisor teams actively use the Assistant.
Case study page, checked 2026-10-30. Caveat: Vendor-published adoption figure; adoption is not the same as answer accuracy.
Suggested workflow
Written for learning, not the company's system. This flow shows how you could build something similar. It is not a description of Morgan Stanley's private architecture.
Step 1: Collect approved policy and research documents with owners and permissions
Step 2: Split documents into chunks and index them with metadata
Step 3: Filter by the user's entitlements, then retrieve passages
Step 4: Generate an answer that cites passages
Step 5: Score answers against expert-written questions before release
Step 6: Collect advisor feedback and review weak answers
Suggested technologies and alternatives
| Layer | Suggested | Alternatives |
|---|---|---|
| API and auth | ASP.NET Core with Entra ID | FastAPI with OAuth2 |
| Search | PostgreSQL + pgvector with hybrid search | Azure AI Search, Elasticsearch |
| Model access | Microsoft.Extensions.AI | Provider SDK or LangChain |
| Evaluation | Versioned question set + scripted checks | MLflow or Ragas |
| Monitoring | OpenTelemetry traces | Application Insights, Langfuse |
Small example
- Input
- What documents does a client need for a retirement-account transfer?
- Output
- A cited answer listing the required forms, with links to the two procedure passages used, or "not found in your accessible documents".
Rebuild it yourself
Build a permission-aware policy assistant over 20 synthetic procedure documents.
Definition of done
- Every answer cites at least one passage the user may read
- Restricted documents never appear for unauthorised test users
- 30-question evaluation set with recorded pass rate
- Unanswerable questions return a no-answer response
Security and privacy risks
- Leaking restricted research to the wrong role
- Outdated documents served as current
- Advisors over-trusting uncited answers
Sources
Listed by title and publisher. This guide contains no external links; search for the title if you want to read the original.
- Shaping the future of financial services (Morgan Stanley case study). OpenAI. Vendor case study.
- Morgan Stanley's OpenAI-powered solution for advisors has expanded. InvestmentNews, 28 Jun 2024. Independent reporting.