From Spreadsheets to a 230,000-Person AI Brain

JPMorgan did not start with AI tools and work backward to strategy. It started with three domains where intelligence could measurably change the economics: knowledge-worker productivity, stronger risk and control, and better customer experience. Every AI investment was tested against those filters. If it did not move a P&L line, it did not get funded.

The structural shift was making business leaders - not technologists - own AI outcomes. As Chief Analytics Officer Derek Waldron told McKinsey, a little under half of JPMorgan employees use gen AI tools every single day in tens of thousands of role-specific ways. That breadth came from business owners identifying where AI could move their metrics, not from IT forcing adoption.

The bank tracks ROI at the initiative level rather than through platform-wide vanity metrics. Benefits have been growing 30-40% year over year. Investment bankers create pitch decks in seconds instead of hours. Coach AI improved client response times by 95% during market volatility and contributed to a 20% increase in gross sales across asset and wealth management between 2023 and 2024.

LLM Suite - JPMorgan's proprietary generative AI platform - is the backbone. Developed in-house to maintain security and regulatory control, it now reaches more than 230,000 employees and has already gone through multiple major upgrades. The platform is model-agnostic and can safely incorporate models from OpenAI, Anthropic, and others within a tightly controlled environment.

The important point is that LLM Suite is not a chatbot. It is a shared intelligence ecosystem that connects AI to the bank's data, documents, applications, and workflows. Analysts use it to summarise research. Developers use it to accelerate coding. Credit teams extract covenant information instantly. Call-centre staff use context-aware Q&A tools. Every new use case inherits the same data access, security, and governance model instead of starting from scratch.

That shared architecture compounds. AI-powered fraud detection has prevented about US$1.5B in losses in real time. The bank already has more than 450 AI use cases in production and is targeting 1,000. The scale works because the platform was built before the organisation chased endless disconnected pilots.

Banking cannot afford to improvise on trust. JPMorgan designed governance into the operating model from the start: access policies, logging, audit trails, and mandatory review paths for higher-risk use cases. In trading and credit, human oversight remains non-negotiable.

The bank's stance is clear: consumer-grade AI tools like ChatGPT and Gemini are banned for internal use, not because the bank fears AI, but because it demands control. Waldron has warned openly that if AI performs correctly 85-95% of the time, human reviewers may stop checking carefully and error can compound at scale. The guardrails were designed before the first major failure, which is why the bank could scale without a trust crisis.

JPMorgan's edge is not a single model or one famous use case. It is the way strategy, business-owned leadership, in-house platform capability, shared architecture, and governance work together as a system.

The lesson is simple: platform before pilots. They did not win by collecting experiments. They built the backbone first, then let every new use case add value without rebuilding the stack from scratch.