🛟 JPMorgan’s Dimon Deploys Multi-Agentic AI at Scale

👋 Welcome, thank you for joining us here :)

The AI Check In is your weekly power play to navigate AI governance and strategy in banking and finance.

What to expect this edition:

This week, we continue our exploration of AI agents in banks. If last week’s focus was on external or customer-facing agents, this week we look at internal AI agents.

• 🛟 Need to Know: Who’s accountable when your AI Agent moves capital or closes a case file?

• 🥷 Deep Dive:  JPM’s multi-agents are miles ahead

• ⚔️ Vendor Spotlight: AutoGen Studio by Microsoft

• 📈 Trends to Watch: Human culture is lagging technology

Rewrite the rules. Outmaneuver competitors. Stay two steps ahead of regulators.

Let the games begin.

🛟 Need to Know: Who’s Accountable When Your AI Moves Capital or Closes a Case File?

Regulators and boards are now asking: when AI agents act inside a bank, who holds the pen?

Internal AI agents, systems that not only analyze data but also trigger actions, are increasingly embedded in back-office operations. These agents are used to classify documents, rebalance liquidity, and even resolve fraud flags without human initiation. The result is a gray zone of accountability.

JPMorgan CEO Jamie Dimon recently confirmed the bank now has over 450 AI use cases in production, including AI agents in post-trade operations and liquidity management.

These systems do not just recommend actions; they execute them. Citi is using AI agents in advisor optimization, document handling, and regulatory analysis, and calls agent deployment "a massive transformation.”

Morgan Stanley’s DevGen.AI has quietly become one of the most effective AI agents in financial services. It has translated over nine million lines of legacy code into plain English and reclaiming over 280,000 hours of developer time. Its real value lies in technical debt reduction and faster modernization cycles—making it one of the most pragmatic internal agent deployments in global banking.

🥊 Your Move:

• Mandate runtime observability and rollback protocols for internal agents, especially those modifying code or workflows.

• Audit existing agent workflows for hidden autonomy and side effects. Document who signs off on agent decisions.

• Assign a C-level sponsor to build an agent governance playbook with clear triggers for escalation and override.

🥷 Deep Dive:  JPMorgan’s Internal AI Agent Ecosystem

JPMorgan Chase is executing one of the most expansive internal AI transformations in global banking. CEO Jamie Dimon disclosed in March 2025 that the bank is running over 450 active AI use cases across divisions, with plans to expand to 800 by the end of the year. These include agentic implementations, autonomous or semi-autonomous systems, that directly influence trading, customer service, fraud detection, and internal decision support.

Scope and Scale

Dimon shared that JPMorgan employs 2,000 AI specialists, with 200 focused on fundamental research. They support a portfolio of applications used by traders, credit analysts, branch managers, and service center staff.

JPMorgan’s AI systems are not experimental. They are delivering hard financial outcomes: an estimated $2 billion in cost savings and revenue generation annually, coincidentally matching the firm's annual AI budget.

Over 200,000 employees now use large language models (LLMs) trained on JPMorgan’s internal data. Dimon noted these tools can answer detailed risk queries such as “Which companies are likely to be downgraded in Argentina?” by scanning tens of thousands of documents in seconds. The firm has not publicly disclosed which LLMs power these tools or whether they are vendor-hosted or entirely proprietary.

Ask David: Research-Grade Agentic Systems

Within JPMorgan Private Bank, the “Ask David” multi-agent system is transforming how internal staff respond to investment queries.

Built to scale employee capabilities, Ask David uses structured and unstructured data (including emails, meetings, and proprietary analytics) to return curated answers in real time. It orchestrates task delegation across domain-specific agents: a supervisor agent oversees intent detection, structured data agents translate queries into SQL/API calls, RAG agents process unstructured content, and analytics agents access JPM’s proprietary models.

The system supports real-time financial advising, even during client meetings. For example, if a client asks why a fund was terminated, Ask David can immediately surface termination reasons, research history, and suggested alternatives. It also personalizes answers by role (e.g., due diligence officer vs. advisor) and uses LLM-based reflection nodes to check coherence before response delivery. Human-in-the-loop oversight is maintained to cover the “last mile” of accuracy.

The Youtube video (linked above), where JP Morgans’ David Odomirok and Jane Xue talk through the rationale and technical aspects Ask David, is highly recommended.

Agentic Implementation Areas (beyond Ask David)

• Call Center Support: AI agents predict customer intent and surface scripts, solutions, or required actions to staff in real time. Dimon forecasts these systems will soon autonomously resolve common tasks, e.g., issuing a replacement debit card, without human intervention.

• Fraud and Scam Detection: Internal agents monitor transaction flows for signs of elder fraud, phishing, or anomalous behavior, and automatically flag or block transactions.

• Equity and Credit Trading: Agents hedge JPMorgan’s equity portfolio dynamically, support trader decisions, and operate in some fully automated trading systems.

• Error Resolution: Autonomous systems already fix back-office errors without human review.

• Marketing Optimization: AI agents identify eligible customers, personalize offers, and optimize distribution channels across marketing campaigns.

• Investment Research: Ask David aggregates documents, executes proprietary model calls, and delivers customized responses to internal staff in real time.

• Governance Structure
  JPMorgan has a dedicated AI executive who reports jointly to Dimon and the firm’s president, signaling AI’s strategic elevation. The firm has adopted a “deployment-first” mindset.

Despite this progressive posture, agentic systems operate within governance constraints at JPM through the Model Risk Management (MRM) framework (2024 Complete Annual Report).

Human approval remains necessary in sensitive workflows. Ask David includes reflection layers, role-based personalization, and memory updates to ensure oversight. The team emphasized evaluation-led development, with iterative accuracy improvement supported by automated testing and SME validation.

Culture and Change Management

Dimon has acknowledged AI will reshape nearly every role. JPMorgan is planning for natural attrition, reskilling, and job redesign to absorb the impact. The Ask David team described fast iterations, subgraph scaling, and progressive specialization as core to deployment success. A key lesson: start with one agent, iterate fast, and layer capabilities only after performance is proven.

Internal Positioning

JPMorgan’s agentic edge stems from proprietary data access and tightly integrated systems. Ask David is being embedded across the Private Bank to deliver answers with the same fidelity as a human analyst at machine speed and scale. The system is expected to extend across other divisions.

Comparative Context
While Citi and Bank of America are advancing agentic capabilities, JPMorgan remains the industry benchmark. Citi’s internal agentic portfolio is still in early deployment. JPMorgan is already operationalizing multi-agent orchestration.

🥊 Your Move