The signal dropped via Anthony Scaramucci, relaying a lunch conversation with Mike Novogratz about BNY Mellon CEO Robin Vince’s view on AI agents:

“Stop asking how AI can make each person more productive. Start looking at every open job and asking: Can an agent do this?”

Not tools. Not copilots. Employees.

BNY Mellon — America’s oldest bank, custodian of $52 trillion in assets — now has 134 “digital employees” deployed through its Eliza platform. They handle payment remediation, code repair, and operational workflows. They have performance reviews. They have human managers. They have email addresses and system logins.

And in the background, human headcount has dropped from ~53,400 in 2023 to 48,100 — a decline of 5,300 positions over two years.

The Official Story vs. The Math

The CFO’s official line: “Our head count has trended down a little bit, but that’s not really anything to do with AI yet.”

CEO Vince frames AI as a “superpower” that unlocks capacity without replacing workers. He cites an MIT study finding that 95% of companies building AI get no return from it. BNY’s bet: they’re in the 5%.

But the operational reality tells a different story:

  • 134 digital employees doing work that was done by humans last year
  • $3.7-4 billion technology spend in 2025 — 19% of revenue, highest among large-bank peers
  • Nearly every staffer trained on agentic tools
  • The bank’s own platform strategy is called “AI for everyone, everywhere, and everything”

When your digital employees have performance reviews and your human headcount is down 10% in two years, the “nothing to do with AI” claim requires increasingly creative accounting.

What Makes BNY’s Approach Different

Most enterprises deploy AI agents as tools that assist human workers. BNY went further: agents as organizational peers.

These digital employees aren’t chatbots answering questions. They:

  1. Have organizational identity — email addresses, system logins, defined roles
  2. Have management structure — assigned human managers who oversee their work
  3. Have performance accountability — reviewed like human employees
  4. Execute real workflows — payment remediation, code updates, operational tasks

This is the same trajectory we documented when G2 started hiring AI agents for enterprise roles — agents moving from “software” to “workforce” in organizational design. BNY is the first major financial institution to do it at this scale.

The Workforce Implications

The Korean AI education boom we covered — where workers are spending millions of won learning AI agent skills out of replacement anxiety — suddenly looks prescient.

BNY’s approach suggests the displacement won’t come as mass layoffs announced in earnings calls. It’ll come as:

  • Attrition without backfill — employees leave, their roles get a digital replacement
  • Task decomposition — complex jobs broken into agent-automatable components
  • Gradual reframing — “we’re augmenting, not replacing” until the headcount numbers speak for themselves

The Meta 20% layoff plan to fund AI infrastructure was explicit. BNY’s approach is more subtle — and potentially more widespread.

The Technology Stack

BNY’s Eliza platform is the internal infrastructure powering these digital employees. Details are scarce (enterprise banking infrastructure isn’t exactly open-source), but the observable capabilities include:

  • Multi-step task execution — not just answering questions, completing workflows end-to-end
  • System integration — agents operate within existing banking systems with real credentials
  • Agentic tool use — accessing payment systems, code repositories, operational databases
  • Enterprise-grade governance — the regulatory environment of a $52T custodian bank demands it

This is conceptually similar to what Snowflake’s Project SnowWork is building for the data platform layer and what Nvidia’s NemoClaw targets for general enterprise deployment — but BNY built it internally for one of the most regulated industries on earth.

Why This Matters Beyond Banking

BNY Mellon is the canary. If the oldest bank in America — regulated, risk-averse, culturally conservative — is treating AI agents as employees with performance reviews, every other enterprise is either doing the same or about to.

The implications:

  1. Identity and access management becomes existential — Okta, ConductorOne, and SailPoint building agent identity products isn’t just a governance exercise. When agents have email addresses and system logins, identity is security.

  2. Governance and compliance can’t lag — when HiddenLayer reports that 1 in 8 AI breaches are linked to agentic systems, financial institutions running 134 agent “employees” need runtime governance that matches human employee oversight.

  3. The “not replacing anyone” period has an expiration date — BNY’s headcount went from 53,400 to 48,100 while deploying 134 digital employees. Watch the next few quarterly earnings for when the CFO’s language shifts.

What OpenClaw Users Should Take Away

The BNY model validates a key thesis: agents are moving from tools to workforce. For OpenClaw users building agent systems:

  • Agent identity matters — give your agents distinct credentials, defined roles, and ownership. BNY does it. You should too.
  • Performance monitoring isn’t optional — if BNY performance-reviews its agents, your production agents need equivalent observability.
  • The workforce conversation is here — whether you’re building agents for your organization or deploying OpenClaw as infrastructure, the end result is autonomous systems doing work humans used to do.

The question Scaramucci surfaced isn’t going away: “Can an agent do this?” BNY is answering it 134 times over. Every enterprise will eventually answer it for every open role.