Latest / MBA Training AI / Right context, wrong assumption: what Morgan Stanley learned about prompting at scale

Transcript

Right context, wrong assumption: what Morgan Stanley learned about prompting at scale · MBA Training AI

• 0:00
  Welcome to the Leaders' Insights podcast.
• 0:07
  Today's episode, right context, wrong assumption, what Morgan Stanley learned about prompting
• 0:13
  at scale.
• 0:14
  Morgan Stanley gave its financial advisors an AI assistant, stuffed it with the firm's
• 0:19
  entire research library, and discovered something that should embarrass half the industry.
• 0:24
  More context made the answers worse.
• 0:26
  Explain that.
• 0:27
  The counterintuitive part everyone skips.
• 0:30
  Morgan Stanley wired their assistant to something like 100,000 internal research documents through
• 0:36
  OpenAI's models.
• 0:37
  And the early lesson wasn't, we need more data.
• 0:41
  It was, we're drowning the model.
• 0:43
  When you dump everything into the prompt, the prompt being the instructions and background
• 0:47
  you hand the model before it answers, it starts treating irrelevant paragraphs as if
• 0:52
  they matter.
• 0:53
  But that's the whole pitch, isn't it?
• 0:55
  Feed it your knowledge base and it gets smarter.
• 0:58
  You're telling me the pitch is wrong?
• 1:00
  The pitch is lazy.
• 1:01
  There's a well-documented failure mode researchers call, lost in the middle.
• 1:06
  Models pay attention to the start and end of a long context window and quietly ignore
• 1:11
  what's buried in the middle.
• 1:13
  So if the one document that answers the advisor's question is on page 40 of 60,
• 1:18
  the model may skim right past it.
• 1:20
  More pages doesn't mean more accuracy.
• 1:23
  It means more noise competing for the model's attention.
• 1:26
  So the fix is retrieval, right?
• 1:28
  Only pull the relevant documents.
• 1:30
  That's not new.
• 1:32
  Retrieval is the tool.
• 1:33
  That's RAG, Retrieval Augmented Generation, where you search your documents first and
• 1:38
  only feed the model the top matches.
• 1:41
  But Morgan Stanley's real insight was upstream of that.
• 1:46
  Retrieval only helps if you've decided what relevant actually means for that user in that
• 1:50
  moment.
• 1:52
  An advisor asking about a client's retirement allocation doesn't need last
• 1:56
  week's oil market commentary, even if the words overlap.
• 2:00
  The discipline isn't fetching context.
• 2:02
  It's ruthlessly deciding what to leave out.
• 2:05
  Give me a concrete example of that going wrong.
• 2:07
  Picture an advisor asks, what's our house view on municipal bonds for
• 2:12
  a high net worth client in California?
• 2:15
  A naive system retrieves 20 documents mentioning municipal bonds,
• 2:19
  including a technical piece on bond mechanics,
• 2:22
  a dated 2023 tax note, and a national overview.
• 2:26
  The model blends them and produces something generic and slightly stale.
• 2:31
  The right system knows the query has three constraints.
• 2:34
  The security type, the client tier, the state tax situation, and
• 2:38
  pulls only the current California specific guidance.
• 2:42
  Same model, completely different answer.
• 2:44
  This sounds like a lot of manual curation.
• 2:47
  Doesn't that defeat the point of automation?
• 2:49
  It shifts where the human effort goes.
• 2:52
  You're not answering questions manually.
• 2:54
  You're building the selection logic once.
• 2:56
  Morgan Stanley put real work into evaluation,
• 2:59
  having actual advisors grade responses before rollout.
• 3:03
  That grading is what tells you your retrieval is pulling garbage.
• 3:07
  Most teams skip it, ship the demo, and
• 3:09
  wonder why adoption dies in month two.
• 3:12
  Let's be honest about the numbers though.
• 3:14
  Morgan Stanley says advisors love it.
• 3:16
  That's the firm talking about its own product.
• 3:19
  How much do we trust that?
• 3:21
  Right to be skeptical.
• 3:22
  Vendor and first party numbers on AI adoption are almost always polished.
• 3:27
  What I trust more is the behavioral signal.
• 3:30
  They reported the assistant getting used across the vast majority of their
• 3:33
  advisor teams, and they kept investing in it rather than quietly shelving it.
• 3:38
  That's more telling than any satisfaction percentage.
• 3:41
  Firms kill tools that don't work.
• 3:44
  This one survived, which suggests the context selection discipline
• 3:48
  actually moved the needle.
• 3:49
  So for someone building this at a much smaller scale,
• 3:52
  a lawyer, a small consultancy, what's the transferable lesson?
• 3:57
  Stop thinking, how do I give the model everything I know?
• 4:01
  Start thinking, what are the three or
• 4:02
  four facts that actually determine the answer to this specific question?
• 4:07
  Structure your retrieval around those, not around keyword overlap.
• 4:11
  And test it against real questions from real users,
• 4:14
  not the tidy examples you invented.
• 4:16
  Give me the one thing to do Monday morning.
• 4:18
  Take your last ten prompts, and for each,
• 4:21
  delete half the context you fed the model, the least specific half.
• 4:25
  Run them again.
• 4:26
  If the answers hold or improve, that context was never helping.
• 4:31
  It was diluting.
• 4:32
  You'll learn more about what your model actually needs in 20 minutes than in
• 4:36
  a month of adding documents.
• 4:38
  That's a wrap.
• 4:39
  Fresh AI briefings drop daily at MBA-training.com.