Morgan Stanley — AI @ MS Assistant, AskResearchGPT & Debrief
The reference case for enterprise RAG under supervision. Morgan Stanley built its first proof of
concept in September 2022 — before ChatGPT existed — and the most valuable published detail is not
the architecture but the evaluation methodology, which became the firm's reusable governance artefact
that every new AI use case must satisfy.
>98%of FA teams using it
20 → 80%document access
10–15 hrssaved per FA per week
9 monthspilot before GA
500question regression suite
~5,000/daylive annotations
It is RAG, not fine-tuning
Early 2023 press described
“fine-tune trained GPT-4 on 100,000 documents.” Kaitlin Elliott, Head of Firmwide GenAI Solutions,
corrected this on record: “After a while of experimenting, we realized that we had built what is now
RAG, but we didn't know what RAG was.” Draw a retrieval pipeline, not a training pipeline.
Azure OpenAI is not confirmed
Every primary source points to a
direct OpenAI relationship: MS was a GPT-4 launch organisation and the only strategic wealth-management
client, ZDR was negotiated bilaterally, and they sat with OpenAI to co-tune retrieval. Do not put Azure on
this diagram.
Two products, two corpora
The WM Assistant (~100k internal docs) and
AskResearchGPT (ISG, 70k+ research reports/yr) are separate systems on separate corpora — not one
product with two front doors.
L1
System context
Four products, one content discipline, and in 2026 a genuinely new pattern: opening the
platform to clients' own autonomous agents over MCP, bypassing the human UI entirely.
McMillan's own system-context diagram, in one sentence
“we envision a
world where AI serves as an efficiency enhancing interaction layer that sits between our colleagues and the
many applications they interact with such as execution and order entry, CRMs, reporting tools and risk
analysis.”
2026: agent-to-system access
Mark Mitchell, CPO of MS at Work:
“In the future state, our corporate clients will not be logging into ShareWorks or Equity Edge.
Instead, they will use agentic AI tools within their own companies to interact with Morgan Stanley's
platforms in a purely agent-to-system manner.” First major Wall Street bank to do this.
L2
Component view — RAG pipeline, controls and the eval loop
The eval column on the right is not an afterthought — it gates the retrieval
column on the left. Every change to search rules or prompts re-runs the 500-question regression suite, and a
broken recency rule in production was caught by daily annotation trending, not by an alert.
morgan stanley — L2 components, controls and evaluation
A 25-question head-to-head: AI versus a
human subject-matter expert, both given one hour. “The AI was able to answer all 25 questions in an
hour, and the humans weren't.” Framing used with control partners: human performs the task at ~40%,
AI at ~80%.
An eval approach that failed
1,000 questions scored by cosine similarity.
Elliott: “we realized it literally told us nothing.” They replaced it with SME and
end-user grading, with failures bucketed thematically.
What is genuinely not published
Vector store product. Chunking strategy,
embedding model, chunk size, top-k. Reranking or hybrid search. Per-advisor entitlements filtering. The
compliance archive design. All five are drawn dashed.
REF
The six-layer stack from the person who built theirs
Jeff McMillan led Firmwide AI at Morgan Stanley, then left and published this. It is
his framework rather than an internal MS diagram — but it is the highest-fidelity public proxy
available, and it maps cleanly onto the L2 diagram above.
Critical path — Debrief: client meeting to CRM to draft email
The interesting constraint here is that Debrief does process client PII — it
transcribes real client meetings. Consent and zero data retention are the compensating controls, and nothing
leaves without an advisor pressing send.
