Morgan Stanley: the GPT-4 assistant used by 98% of financial advisor teams

• Morgan Stanley uses AI evals to shape the future of financial services — OpenAI (customer story)
• Morgan Stanley kicks off generative AI era on Wall Street with assistant for financial advisors — CNBC, 2023-09-18
• Launch of AI @ Morgan Stanley Debrief — Morgan Stanley (пресс-релиз), 2024-06-26
• Morgan Stanley wealth advisors are about to get an OpenAI-powered assistant to do their grunt work — CNBC, 2024-06-26
• Morgan Stanley Research Announces AskResearchGPT — Morgan Stanley (пресс-релиз), 2024-10-23

Background

Morgan Stanley Wealth Management is one of the world's largest wealth management operations: about 15,000 financial advisors and $5.5 trillion in client assets (as of March 2024; the firm's target is $10 trillion). The division's key intangible asset is its internal knowledge base: decades of research, market analysis, procedures, and playbooks packaged into a corpus of roughly 100,000 documents.

The catch: a knowledge base is only as valuable as the speed of access to it. An advisor on a client call cannot afford twenty minutes of searching internal PDFs — which meant a huge share of the firm's 'intellectual capital' simply did not work at the moment of sale or advice.

In March 2023 Morgan Stanley announced it was building an assistant on GPT-4, becoming one of OpenAI's first enterprise partners. The competitive context matters: Goldman Sachs and JPMorgan Chase had also announced generative AI projects by then, but Morgan Stanley was the first major Wall Street firm to put a bespoke GPT-4 solution into employees' hands — as CNBC reported, citing Jeff McMillan, then head of analytics, data and innovation at wealth management (later Head of Firmwide AI).

For a regulated financial industry this was a bold step: banks have zero tolerance for invented answers, and any system that can 'hallucinate' a recommendation is a compliance risk. That is why the Morgan Stanley case became the reference point not because of the model but because of the implementation methodology: systematic evals that turned trust in AI from a matter of faith into a matter of measurement. McMillan put the motivation plainly: 'I've never seen anything like this in my career, and I've been doing artificial intelligence for 20 years. We saw a window of opportunity that was just completely disruptive, and I think as an organization, we didn't want to get left behind.'

Problem

The firm's knowledge was locked inside a huge corpus of internal documents. By the company's own account, advisors could effectively reach only a fraction of it: document accessibility was estimated at 20%. The previous internal tool answered a limited set of questions — around 7,000 — and did not scale to the firm's full body of knowledge.

The mechanics are familiar to anyone who has used corporate search: classic keyword search finds documents, not answers. An advisor asked a client question about market strategy or an internal procedure had to craft a query, open several reports, scroll dozens of pages, and synthesize the answer themselves — minutes at best, and in practice often 'I'll get back to you'. On a call that meant a pause and lost momentum; after the call, a delayed answer to a client who is already comparing the service with competitors.

The second, less obvious pain was meeting post-processing. The division hosts about a million Zoom calls a year, and a typical advisor — per pilot participant Don Whitehead — runs four to six meetings a day. After each one someone has to write notes, capture commitments, update the CRM, and send a follow-up email. Advisors or their junior staff did this routine by hand, and follow-up preparation could stretch into days.

Both pains add up to one strategic problem: the time of the firm's most expensive specialists went into information retrieval and record-keeping — instead of the client work they were hired for.

Solution

Morgan Stanley built the 'AI @ Morgan Stanley Assistant' — a GPT-4 assistant over the internal knowledge base using a RAG scheme: the model answers only from the firm's vetted documents, not 'from memory'. Before launch the team spent months curating the corpus and testing answers with human experts (McMillan described this to CNBC). The full rollout to all advisors and their support staff happened on September 18, 2023 — CNBC obtained the memo from co-President Andy Saperstein: 'Financial advisors will always be the center of Morgan Stanley wealth management's universe. We also believe that generative AI will revolutionize client interactions... and ultimately help free up time to do what you do best: serve your clients.'

The key implementation element is the evals system, which OpenAI later detailed in a dedicated case study. The team started with three target scenarios: faster information retrieval, automation of routine report summarization, and insights tailored to client needs. Each got its own evaluations: advisors and prompt engineers manually graded model answers for accuracy and coherence, and prompts and retrieval were refined accordingly. Translation evals for multilingual clients came next, and together with OpenAI the team iteratively tuned retrieval methods over the growing library. In production, a daily regression suite of sample questions catches degradations before compliance would notice them. The data question was settled separately: a zero-data-retention policy guarantees the firm's proprietary data is not used to train public models — David Wu called that condition critical to the deal.

A curious implementation detail was habit change: advisors had to learn to ask questions in full sentences, as they would with a human assistant, rather than keywords, as with a search engine. 'No different than how I would ask you a question, that's how you talk to this machine. People are not accustomed to that,' McMillan explained.

The lineup's second product is 'AI @ Morgan Stanley Debrief' on a Whisper + GPT-4 stack: with client consent (requested for every meeting separately), the tool listens to a Zoom meeting, generates structured notes, captures action items, saves the record into Salesforce, and drafts a follow-up email the advisor edits and sends personally. The pilot started in March 2024, the firm-wide launch was announced June 26, 2024, and the rollout to ~15,000 advisors completed by early July. McMillan noted that tuning Debrief's prompts took months — and that 'this does a better job of taking notes than the average human.'

For the team it was also a development paradigm shift. 'The traditional way in which you would solve those things is you would write code,' McMillan explained. 'In the new world, you give examples of what "good" looks like, and the system learns what good is. It's actually able to "reason" and apply logic that a human would apply.' Wealth management head Jed Finn set the strategic frame: the goal is a technology layer in which an advisor performs all tasks with simple prompts — from drafting proposals to rebalancing portfolios and producing reports.

Crucially, a human stays in the loop at every step: the Assistant retrieves and synthesizes, Debrief drafts, but the final decision and the send button always belong to the advisor. For the regulator and compliance that boundary is fundamental.

Result

More than 98% of financial advisor teams use the assistant — near-ceiling adoption for a corporate tool. The system effectively answers questions across a corpus of 100,000 documents — versus roughly 7,000 questions the previous tool could handle. Document accessibility, by the firm's estimate, rose from 20% to 80%, and information retrieval collapsed from minutes to seconds. McMillan describes the qualitative shift: 'Now, advisors can engage clients on topics they haven't discussed before because the friction between knowledge and communication has gone to zero.'

For Debrief, early public estimates exist: a pilot advisor put the savings at about 30 minutes of work per meeting; with a million Zoom calls a year across the division, CNBC estimated the potential at hundreds of thousands of hours annually. Kaitlin Elliott, Head of Firmwide Generative AI Solutions, noted: 'The feedback from advisors has been overwhelmingly positive... follow-ups that used to take days now happen within hours.' At the same time McMillan candidly called the project a 'grand experiment in productivity' and said a rigorous impact assessment would take at least a year of observation.

The approach is scaling beyond wealth management: on October 23, 2024 the firm launched AskResearchGPT — a GPT-4 assistant over the Morgan Stanley Research corpus (70,000+ proprietary reports a year) for investment banking, sales & trading, and research staff, with a patented 'query → client-ready email with citations and links in one click' workflow.

It is important to frame these numbers: Morgan Stanley publishes usage metrics (adoption, corpus accessibility, speed) rather than a dollar impact — the bank has nowhere claimed 'AI made us X million'. In our view this is a deliberate and honest position for an internal tool: rigorously attributing a copilot's contribution to advisor revenue is methodologically hard, and the firm prefers metrics it controls.

The industry stakes are high: per a Citigroup estimate cited by CNBC, finance jobs are among the most prone to AI displacement, and AI adoption could boost the industry's profit by $170 billion by 2028.

For the market, the Morgan Stanley case is the model of the 'low-risk' path into generative AI, the mirror image of customer-facing chatbots like Klarna's: the AI faces inward, toward employees; a human stays in the loop of every decision; and trust is built on measurable evals, not demos. In our view that combination — a narrow corpus, evals as a process, a human as the final filter — is the case's most transferable part: it works in any industry where the cost of a wrong answer is high, from medicine to law.

Lessons learned

1. In a regulated industry, trust in AI is built on evals: systematic answer-quality scoring on real questions is what convinces compliance — not demos.
2. RAG over your own corpus is the right first bet for a knowledge-heavy business: the model doesn't 'know' the answer, it finds it in the firm's documents.
3. Evals are a process, not a one-off acceptance test: a daily regression suite on sample questions catches quality degradation before users or compliance notice.
4. The 'knowledge accessibility' metric (20→80%) is a more honest value indicator than 'number of bot queries'.
5. 98% adoption follows from embedding into the advisor's real workflow (and months of corpus curation before launch), not from a top-down mandate.
6. Per-meeting client consent and final human editing (Debrief) are the template for bringing AI into sensitive client scenarios without reputational risk.
7. Start with one narrow corpus and one user role (the advisor); scale after quality is proven — Morgan Stanley's path from Assistant to Debrief and AskResearchGPT took two years of sequential launches.

Frequently asked questions