Concept card

BNY Built Its Digital Workforce Backward — And It’s Working

As of 2026-07-24T14:00:00Z

Edited by Cédric-Olivier Mâry, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is BNY Built Its Digital Workforce Backward — And It’s Working?

BNY designed its digital workforce first, then asked which humans the agents needed — each agent got a job description, SLA, and escalation path, and the numbers back it: 18% adjusted CAGR in 2025.

What happened, and why it is not just a case study

In a Forrester analysis published in 2026, analyst Brian Hopkins describes how BNY Mellon inverted the usual sequence for deploying AI agents. Most organisations start from an existing human workflow and ask an AI agent to sit inside it — augmenting a clerk, assisting an analyst, drafting alongside a case handler. BNY did the opposite: it designed each digital worker as a standalone role first, complete with a job description, a work queue, a service-level agreement, and an escalation path, and only then asked which humans that digital worker needed to interact with. The ordering sounds like a footnote, but it solves a problem that quietly kills most agentic AI programmes: accountability drift.

The problem the backward ordering solves

When an AI agent is dropped into an existing human role, it inherits everything ambiguous about that role — undefined ownership, informal escalation habits, no independent success metric. Three months into the programme, a steering committee asks whether the agent is actually delivering value, and nobody can answer cleanly, because the agent's performance was never separated from the human's satisfaction score or workload perception. The agent becomes a permanent pilot that nobody can kill and nobody can fully credit. BNY avoided this by treating every AI deployment the way an organisation treats a new hire: define the job, the queue it works from, the service-level agreement it is held to, and the on-call path for when it cannot resolve something — before a single human role is touched.

Why it matters

Deploying an agent into an existing role inherits that role's ambiguity — unclear ownership, no SLA, no way to measure the agent's value separately from the human it assists.
Eliza now serves roughly 50,000 employees (97% of the bank); 130 digital employees were in production by end-2025 (up from 70), with 160+ AI solutions deployed — a 200%+ year-over-year jump.
The financial corollary — 18% adjusted CAGR, 21% pre-tax income growth, 13% dividend increase in 2025 — was reported as one integrated story with the AI programme, not a separate chapter.

Key points

Design the agent’s job-to-be-done, queue, and SLA before defining which human roles it will escalate to — not the reverse.
BNY’s Eliza platform: 50,000 employees served, 130+ digital employees in production (2025), 200%+ year-over-year growth in deployed AI solutions.
Scope Joule agents with ownership language: ‘resolves 80% of invoice exceptions’ rather than ‘assists the AP clerk’ — ownership changes the integration surface, the metric, and the org chart.
Agents require policy authority in SAP Build Process Automation (write access to decision rules), not just read access to queues — without it the agent cannot close a case.
The agent’s reporting line is to the business lead, not IT — the org chart is a design deliverable of the agent programme.
Mandatory trigger for this pattern: programme has been live 90 days and the steering committee cannot distinguish agent resolutions from human resolutions.
Avoid for narrow, single-user copilot use cases where the human retains full action ownership — the full digital workforce design overhead is not warranted.
SAP’s Autonomous Suite (Sapphire 2026) deploys 200+ agents across five domains with defined KPIs tracked through SAP AI Agent Hub — the same accountability logic at platform scale.
BNY’s 2025 financial results (18% adjusted CAGR, 21% pre-tax income growth) were reported alongside the AI programme as an integrated story, not a separate annex.
Training at scale is non-negotiable: 1,400+ employees completing 40-hour AI bootcamps and 170,000+ cumulative training hours at BNY underpinned the agent success.

Terms used on this page

Digital workforce: Set of agents managed as a team, with queues, SLAs and on-call.
JTBD: Job-to-be-done — outcome the agent owns end-to-end.
Build Process Automation: SAP's workflow engine where agents need policy authority.
Decision owner: The accountable person who accepts the trade-off and funds the next action.
Semantic contract: The shared definition of business terms, metrics, entities, and access rules used by tools and teams.
Control plane: The layer that applies policy, access, lineage, monitoring, and escalation across the operating model.
Evidence grade: A label that separates verified fact, directional signal, modeled assumption, and field observation.
Adoption metric: The measurable behavior proving that the concept changed actual work after go-live.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →