Opening: Context and Inflection Point
The Bank of New York Mellon (BNY) is not an institution that can afford to “experiment at leisure.” It operates at the infrastructural core of the global financial system—asset custody, clearing, and the movement and safeguarding of data and cash. As of the third quarter of 2025, the value of assets under custody and/or administration reached approximately USD 57.8 trillion. Any error, delay, or compliance lapse in its processes is therefore magnified into systemic risk. ([bny.com][1])
When ChatGPT ignited the wave of generative AI at the end of 2022, BNY did not confine its exploration to a small circle of engineers or innovation labs. Instead, it elevated the question to the level of how the enterprise itself should operate. If AI is destined to become the operating system of future technology, then within a systemically important financial institution it cannot exist as a peripheral tool. It must scale within clearly defined boundaries of governance, permissions, auditability, and accountability. ([OpenAI][2])
This marked the inflection point. BNY chose to build a centralized platform—Eliza—integrating model capabilities, governance mechanisms, and workforce enablement into a single, scalable system of work, developed in collaboration with frontier model providers such as OpenAI. ([OpenAI][2])
Problem Recognition and Internal Reflection: The Bottleneck Was Not Models, but Structural Imbalance
In large financial institutions, the main barrier to scaling AI is rarely compute or model availability. More often, it lies in three forms of structural imbalance:
Information silos and fragmented permissions: Data and knowledge across legal, compliance, business, and engineering functions fail to flow within a unified boundary, resulting in “usable data that cannot be used” and “available knowledge that cannot be found.”
Knowledge discontinuity and poor reuse: Point-solution proofs of concept generate prompts, agents, and best practices that are difficult to replicate across teams. Innovation is repeatedly reinvented rather than compounded.
Tension between risk review and experimentation speed: In high-risk industries, governance is often layered into approval stacks, slowing experimentation and deployment until both governance and innovation lose momentum.
BNY reached a clear conclusion: governance should not be the brake on AI at scale—it should be the accelerator. The prerequisite is to design governance into the system itself, rather than applying it as an after-the-fact patch. Both OpenAI’s case narrative and BNY’s official communications emphasize that Eliza’s defining characteristic is governance embedded at the system level. Prompts, agent development, model selection, and sharing all occur within a controlled environment, with use cases continuously reviewed through cross-functional mechanisms. ([OpenAI][2])
Strategic Inflection and the Introduction of an AI Platform: From “Using AI” to “Re-architecting Work”
BNY did not define generative AI as a point-efficiency tool. It positioned it as a system of work and a platform capability. This strategic stance is reflected in three concrete moves:
Centralized AI Hub + Enterprise Platform Eliza
A single entry point, a unified capability stack, and consistent governance and audit boundaries. ([OpenAI][2])From Use-Case Driven to Platform-Driven Adoption
Every department is empowered to build first, with sharing and reuse enabling scale. Eliza now supports 125+ active use cases, with 20,000 employees actively building agents. ([OpenAI][2])Embedding “Deep Research” into the Decision Chain
For complex tasks such as legal analysis, risk modeling, and scenario planning, multi-step reasoning is combined with internal and external data as a pre-decision thinking partner, working in tandem with agents to trigger follow-on actions. ([OpenAI][2])
Organizational Intelligence Re-architecture: From Departmental Coordination to Integrated Knowledge, Workflow, and Accountability
Eliza is not “another chat tool.” It represents a reconfiguration of how the organization operates. The transformation can be summarized along three linked pathways:
1. Departmental Coordination → Knowledge-Sharing Mechanisms
Within Eliza, BNY developed a mode of collaboration characterized by joint experimentation, shared prompts, reusable agents, and continuous iteration. Collaboration no longer means more meetings; it means faster collective validation and reuse. ([OpenAI][2])
2. Data Reuse → Formation of Intelligent Workflows
By unifying permissions, controls, and oversight at the platform level, Eliza allows “usable data” and “usable knowledge” to enter controlled workflows. This reduces redundant labor and gray processes while laying the foundation for scalable reuse. ([bny.com][3])
3. Decision Models → Model-Based Consensus
In high-risk environments, model outputs must be tied to accountability. BNY’s approach productizes governance itself: cross-functional review and visible, in-platform controls ensure that use cases evolve from the outset within a consistent risk and oversight framework. ([bny.com][3])
From HaxiTAG’s perspective, the abstraction is clear: the deliverable of AI transformation is not a single model, but a replicable intelligent work system. In product terms, this often corresponds to a composable platform architecture—such as YueLi Engine (knowledge computation and orchestration), EiKM (knowledge accumulation and reuse), and vertical systems like ESGtank—that connects knowledge, tools, workflows, and auditability within a unified boundary.
Performance and Quantified Impact: Proving That Scale Is More Than a Slogan
What makes BNY’s case persuasive is that early use cases were both measurable and repeatable:
Contract Review Assistant: For more than 3,000 supplier contracts per year, legal review time was reduced from four hours to one hour, a 75% reduction. ([OpenAI][2])
Platform Scale Metrics: With 125+ active use cases and 20,000 employees building agents, capability has expanded from a small group of experts to the organizational mainstream. ([bny.com][3])
Cultural and Capability Diffusion: Training programs and community-based initiatives encouraged employees to see themselves as problem solvers and agent builders, reinforced through cross-functional hackathons. ([OpenAI][2])
Together, these indicators point to a deeper outcome: AI’s value lies not merely in time savings, but in upgrading knowledge work from manual handling to controlled, autonomous workflows, thereby increasing organizational resilience and responsiveness.
Governance and Reflection: Balancing Technology and Ethics Through “Endogenous Governance”
In financial services, AI risks are tangible rather than theoretical—data misuse, privacy and compliance violations, hallucination-driven errors, permission overreach, and non-traceable audits can all escalate into reputational or regulatory crises.
BNY’s governance philosophy avoids adding yet another “AI approval layer.” Instead, governance is built into the platform itself:
Unified permissions, security protections, and oversight mechanisms;
Continuous pre- and post-deployment evaluation of use cases;
Governance designed to accelerate action, not suppress innovation. ([bny.com][3])
The lessons for peers are straightforward:
Define accountability boundaries before autonomy: Without accountable autonomy, scalable agents are impossible.
Productize governance, don’t proceduralize it: Governance trapped in documents and meetings cannot scale.
Treat training as infrastructure: The real bottleneck is often the distribution of capability, not model performance.
Overview of AI Application Impact in BNY Scenarios
| Application Scenario | AI Capabilities Used | Practical Impact | Quantified Results | Strategic Significance |
|---|---|---|---|---|
| Supplier Contract Review | NLP + Retrieval-Augmented Generation (RAG) + Structured Summarization | Faster legal review and greater consistency | Review time reduced from 4 hours to 1 hour (-75%); 3,000+ contracts/year ([OpenAI][2]) | Transforms high-risk knowledge work into auditable workflows |
| HR Policy Q&A | Enterprise knowledge Q&A + Permission control | Fewer manual requests; unified responses | Reduced manual requests and improved consistency (no disclosed figures) ([OpenAI][2]) | Reduces organizational friction through knowledge reuse |
| Risk Insight Agent | Multi-step reasoning + internal/external data fusion | Early identification of emerging risk signals | No specific lead time disclosed (described as pre-emptive intervention) ([OpenAI][2]) | Enhances risk resilience through cognitive front-loading |
| Enterprise-Scale Platform (Eliza) | Agent building/sharing + unified governance + controlled environment | Expands innovation from experts to the entire workforce | 125+ active use cases; 20,000 employees building agents ([bny.com][3]) | Turns AI into the organization’s operating system |
HaxiTAG-Style Intelligent Leap: Delivering Experience and Value Transformation, Not a Technical Checklist
BNY’s case is representative not because of which model it adopted, but because it designed a replicable diffusion path for generative AI: platform-level boundaries, governance-driven acceleration, culture-shaping training, and trust built on measurable outcomes. ([OpenAI][2])
For HaxiTAG, this is precisely where productization and delivery methodology converge. With YueLi Engine, knowledge, data, models, and workflows are orchestrated into reusable intelligent pipelines; with EiKM, organizational experience is accumulated into searchable, reviewable knowledge assets; and through systems such as ESGtank, intelligence is embedded directly into compliance and governance frameworks. The result is AI that enters daily enterprise operations in a controllable, auditable, and replicable form.
When AI is truly embedded into an organization’s permission structures, audit trails, and accountability mechanisms, it ceases to be a passing efficiency trend—and becomes a compounding engine of long-term competitive advantage.