BNY Mellon was founded in 1784. It holds custody over $52 trillion in assets. It has 48,100 employees and spent $3.8 billion on technology in 2025 — 19% of revenue, the highest ratio among large-bank peers. When a 241-year-old institution moves this aggressively on AI, it’s worth paying close attention to what they actually did, not just the headline numbers.
This is an attempt to do that honestly.
Why Did BNY Build One Platform Instead of 160 Separate Tools?
The Eliza platform launched in 2023. By 2025, it had 160 enterprise AI solutions running in production and 99% employee adoption. That last number gets cited a lot. The first number is the one that deserves more attention.
Most enterprises don’t build one platform. They end up with a portfolio of disconnected tools — a legal team’s contract tool, an operations team’s exception-handling tool, a dev team’s code assistant — each with its own governance, maintenance overhead, and user experience. The compounding cost of that fragmentation is invisible until it isn’t.
BNY’s architecture decision to build a single platform for all 160 solutions means every solution inherits the same security posture, the same audit trail, the same access controls. For a custody bank handling $52 trillion, those controls aren’t optional. But the architectural discipline also enables everything else that followed: the 175% growth in platform usage in 2025, the ability to deploy autonomous agents at scale, and critically, the ability to let 20,000 non-engineers build on it without the whole thing collapsing.
One platform is a constraint that generates options. Most enterprises choose the opposite — maximum flexibility upfront, maximum fragmentation later.
What Does It Actually Mean to Have 134 Digital Employees?
This is where BNY’s story gets genuinely novel and where the language deserves scrutiny.
The 134 “digital employees” are autonomous multi-agent AI systems with their own login IDs and email addresses. They have human supervisors. They are assigned to specific workflows: payment remediation, code repair, client onboarding. They are not chatbots answering questions. They are agents executing processes.
The practical meaning is this: when a payment exception surfaces at 6am, a digital employee can triage it, cross-reference transaction data, flag anomalies, and resolve or escalate — before a human being has logged on. BNY reports 30% fewer morning exceptions and 40% faster issue resolution in payment operations as a result. Those are not marginal gains.
The email address and login ID are not cosmetic. They exist because any action in a regulated financial system needs an accountable identity. Digital employees create audit trails. When something goes wrong — and at scale, things go wrong — you need to know which agent took which action, under whose supervision, at what time. BNY solved a real compliance problem by treating agents as identity-bearing entities rather than anonymous processes running in the background.
Whether “digital employee” is the right frame or a marketing choice is a fair question. I’d argue the framing matters less than the architecture it reflects.
How Did 20,000 Non-Engineers Become AI Builders?
The number that stopped me when I first read BNY’s reporting was this: of 48,100 employees, 20,000 are actively building agents on Eliza. Most of them are not software engineers.
That is not a training programme outcome. That is a platform design outcome. If the tool requires engineering knowledge to use, only engineers use it. BNY built Eliza so that operations analysts, legal associates, and client service managers could create and deploy agents within guardrails the platform enforces automatically.
There is a lesson here that gets lost in the conversation about AI replacing workers: the organisations that move fastest are the ones where the people closest to the problem can build the solution. A compliance analyst who runs the same manual check 300 times a year knows exactly what that check involves. If they can automate it themselves, without a six-month IT project, the organisation moves at a different speed.
The multi-year partnership with OpenAI — ChatGPT Enterprise, Deep Research, Operator — and the Google Cloud infrastructure underneath Eliza are what make this possible technically. But the decision to democratise access rather than centralise it is an organisational choice, and it’s the more important one.
Where Do the $550M in Savings Actually Come From — and What’s Missing from That Number?
BNY reported $550M in efficiency savings in 2025 and reinvested approximately $500M of that into growth initiatives. The company posted record net income of $5.3 billion on $20.1 billion in revenue, with EPS up 28% year over year.
Here is what BNY’s leadership is careful to say: the $550M comes from a broader transformation that involved transitioning roughly 70% of employees to a new platforms operating model. Eliza is a significant part of that story, but not the whole story. The attribution problem in large-scale transformation is real — when you change the operating model, the technology, and the organisational structure simultaneously, separating out the AI contribution is genuinely difficult.
The specific Eliza-attributable outcomes BNY has disclosed are concrete: the Contract Review Assistant reduced legal review time by 75%, from four hours to one hour, across more than 3,000 vendor contracts per year. Core custody trade unit costs fell approximately 5%. Select operations saw 24% unit cost reductions versus 2024.
What isn’t being disclosed — and this is worth sitting with — is any account of failures, rollbacks, or projects that didn’t deliver. At 160 solutions in production, at this scale, some things did not work. The absence of that data in public communications is not evidence it doesn’t exist. It is evidence that BNY’s communications are optimised for investor relations, not enterprise learning. Every organisation benchmarking against BNY should factor that asymmetry in.
What Does BNY’s Headcount Decline Have to Do with AI, If Anything?
BNY had approximately 53,400 employees in 2023. By the end of 2025, that number was 48,100. That is roughly 5,300 fewer people over two years — the same two years in which 134 digital employees were deployed across core operations.
CFO Dermot McDonogh addressed this directly on the Q4 2025 earnings call: the headcount reduction “is not really anything to do with AI yet,” framing AI as “unlocking capacity.” CEO Robin Vince has called AI “a jobs creator.”
I am not saying BNY’s leadership is being dishonest. I think they are describing what they believe to be true. The headcount reduction is attributable to location consolidation, attrition, and the broader operating model change. The Wexford site closure eliminated 300+ jobs in 2025; BNY attributed it to geographic consolidation, not automation. Reporters noted the timing.
The honest position is this: the causal chain between AI deployment and headcount reduction at BNY is genuinely unclear, even to the people inside it. That is almost always true at this stage of a transformation. What is not unclear is the juxtaposition: 134 autonomous digital workers deployed, 5,300 fewer humans employed, leadership asserting the two facts are unrelated. That juxtaposition is the governance and communications challenge every enterprise will face, and BNY is currently the most visible example of how not to navigate it. The correct approach is not to assert the facts are unrelated. It is to say, precisely and honestly, what you know, what you don’t know, and what you are doing to find out.
What Can a Custody Bank Teach Every Enterprise About AI Deployment?
Three things, in order of importance.
The single-platform architecture is the most transferable lesson. Not because it is the easiest path — it almost certainly wasn’t — but because it is the only architecture that scales governance alongside capability. If you are deploying AI across an enterprise and you do not have a unified platform with centralised identity, audit, and access control, you are accumulating a debt that will become visible at the worst possible time.
The democratisation of agent-building is the second lesson. Eliza’s 20,000 non-engineer builders are not an accident. They are the product of a deliberate decision to make the platform accessible to domain experts, not just technical staff. The enterprises that move fastest will be the ones where the person closest to the problem can also build the solution.
The third lesson is harder. BNY is navigating the headcount/digital-worker tension in public, imperfectly, in real time. The lesson is not to do what BNY did communicatively. The lesson is to develop your position on this question before you are forced to answer it in an earnings call. “AI is unlocking capacity” is not a workforce strategy. It is a holding position. Enterprise leaders need to do better than that — not because of optics, but because the people whose roles are changing deserve a clearer account of what is actually happening.
BNY’s Eliza is a genuine achievement. The platform architecture is sound. The outcomes are real. The honest version of the story, which includes the parts BNY is less eager to discuss, is more useful to every enterprise than the press release version.