Multi-Agent Investment Research Assistant with RAG and Human-in-the-Loop

j.p._morgan_chase · Finance · year n/a · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Private-bank investment research

J.P. Morgan Chase's Private Bank needed to answer investment research questions across thousands of products and years of data without removing human judgment from high-stakes client-asset decisions.

Constraints:

• The domain involves billions of dollars of client assets.
• Research data spans databases, files, presentations, notes, emails, and proprietary analytics.
• The assistant must personalize answers by user role while preserving last-mile human accuracy.
• Structured and unstructured data require different retrieval and reasoning paths.

Architecture

Ask David uses a supervisor agent to orchestrate specialized subagents for structured data, unstructured RAG documents, and proprietary analytics. It includes short- and long-term memory for personalization, reflection nodes using LLM-as-judge, independent sub-agent testing, and human-in-the-loop escalation for the critical last-mile accuracy required in investment research.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  supervisor-agent["Supervisor Agent"]
  structured-data-agent["Structured Data Agent"]
  unstructured-data-agent["Unstructured Data Agent"]
  proprietary-analytics-agent["Proprietary Analytics Agent"]
  human-research-expert["Human Research Expert"]
  structured-data-systems["Structured Data Systems"]
  research-document-store["Research Document Store"]
  supervisor-agent --> structured-data-agent
  supervisor-agent --> unstructured-data-agent
  supervisor-agent --> proprietary-analytics-agent
  supervisor-agent --> human-research-expert
  structured-data-agent --> structured-data-systems
  unstructured-data-agent --> research-document-store
  human-research-expert --> supervisor-agent

Components

Data flow

A user asks an investment research question. The supervisor interprets intent, routes parts of the question to structured-data, unstructured-document, or proprietary-analytics agents, applies reflection and evaluation nodes, and invokes human review when the answer affects high-stakes decisions or needs real-David expertise.

Why this architecture

The team used a supervisor with specialist subagents because financial research combines structured data retrieval, document RAG, proprietary analytics, personalization, and human review. The architecture augments research experts rather than replacing them because last-mile accuracy carries financial risk.

Models used: unspecified LLMs

Scale

• The research team manages thousands of investment products.
  “The team manages thousands of investment products backed by years of valuable data”summary, stated
• The domain involves billions of dollars of client assets.
  “given the high stakes involved with billions of dollars of client assets.”summary, stated

Alternatives considered (roads not taken)

Replace investment researchers outright

Rejected because: The team frames Ask David as augmentation and still consults human David when needed.

Use one agent for all data types

Rejected because: Structured SQL/API data, document RAG, and proprietary analytics require distinct skills.

Patterns

multi-agent-orchestration (architecture)

A supervisor delegates to structured, unstructured, and proprietary analytics subagents.

rag-retrieval (knowledge)

The unstructured data agent retrieves across investment documents, emails, notes, presentations, and recordings.

human-in-the-loop-gate (safety)

Human experts remain in the loop for last-mile accuracy in high-stakes investment research.

llm-as-judge (quality)

Reflection nodes use LLM-as-judge to evaluate outputs.

Tradeoffs

Use a human-augmented assistant

Chose Ask David with human-in-the-loop over Fully autonomous investment research.

Rationale: The stakes require human last-mile review despite automation.

Specialize agents by data modality

Chose Structured, unstructured, and analytics agents over A single catch-all research agent.

Rationale: Financial research spans SQL/API data, documents, and proprietary analytics.

Gotchas (what bit them)

Investment answers cannot be treated as autonomous final authority.

Root cause: The research domain involves billions in client assets and high accuracy requirements.

Resolution: Ask David keeps human-in-the-loop last-mile validation.

Manual research across scattered systems limited scale.

Root cause: Questions previously required searching databases, files, and materials manually.

Resolution: The assistant orchestrates structured-data, document-RAG, and analytics agents.

Evaluation

The team used evaluation-driven development with independent sub-agent testing, reflection nodes, and human last-mile review.

Outcomes

• Ask David automated research across thousands of investment products.
• The assistant augments, rather than replaces, human private-bank researchers.
• The system combines structured data, RAG, proprietary analytics, memory, and human review.

The case names thousands of products and billions of dollars of client assets.