JPMorgan Chase

Investment banks operate across highly heterogeneous business functions — sales desks, trading floors, risk management, finance, technology, and operations — each with distinct data environments, workflows, and compliance requirements. For JPMorgan Chase, the core challenge was not model capability; as chief analytics officer Derek Waldron recognized early, LLMs were already sufficiently capable. The real barrier was connectivity: even state-of-the-art models produce no enterprise value if they cannot reach internal systems, proprietary data stores, and institutional knowledge locked inside CRM, HR, trading, and risk platforms. Without that connectivity layer, AI would remain a proof-of-concept rather than a productivity multiplier across a workforce of hundreds of thousands.

JPMorgan built a connectivity-first internal LLM platform centered on personal AI assistants, treating AI as core infrastructure from the outset rather than a pilot experiment. The technical architecture is anchored by a fourth-generation multimodal retrieval-augmented generation (RAG) system — evolved through successive generations from basic vector search to hierarchical, authoritative, multimodal knowledge pipelines. The platform connects to an expanding ecosystem of enterprise data sources: CRM, HR, trading, finance, risk systems, structured data stores, and document knowledge bases, with new connectors added monthly. Rather than limiting employees to prompt interfaces, the platform allows workers to build and customize role-specific AI assistants using reusable building blocks — RAG, document intelligence, and structured data querying — assembled into tools with specific personas, instructions, and roles. This "one platform, many jobs" architecture enables deep customization across functions without fragmenting into disconnected point solutions.

Adoption grew from zero to 250,000 active users within months of launch, driven entirely by voluntary peer-to-peer sharing of use cases — no mandates were issued. Today, over 60% of JPMorgan's workforce across sales, finance, technology, operations, trading, and risk actively uses the platform, representing one of the largest voluntary enterprise AI deployments on record. The firm has iterated through four RAG generations, each materially improving retrieval fidelity and document handling. Qualitatively, the rollout produced a self-reinforcing innovation flywheel:

• Employees moved beyond prompt engineering to building and distributing custom assistants
• Role-specific tools spread organically through internal sharing channels
• Engagement deepened across all seniority levels and business lines

• Treat AI as core infrastructure from day one — connectivity to proprietary systems, not model selection, is the primary engineering challenge.
• Proprietary data access is the defensible moat; the models themselves are increasingly commoditized across the industry.
• Voluntary, bottom-up adoption driven by peer sharing of concrete use cases outperforms top-down mandates at enterprise scale.
• Design for a "one platform, many jobs" model using reusable building blocks so individual roles can self-serve customization without creating fragmentation.
• Plan for RAG to be iterative — retrieval quality improves significantly across generations and directly determines the usefulness of any knowledge-grounded application.