BNY gives 140 AI agents a supervisor, a scorecard and a 16-step approval
BNY runs nearly 140 AI agents like staff: each has a named supervisor and a scorecard, and a new one passes 16 review steps. No cost is given.
At a glance
- Company
- BNY · Financial services · Size not stated
- Team and use case
- Business operations · AI programme management
- Tools
- Eliza (BNY's in-house AI platform), Microsoft Azure, Microsoft 365 Copilot
- Stage
- Scaled across teams, since "Last spring", per Microsoft on 12 May 2026
- Results
- AI agents ("digital employees") in use: More than 100 (16 October 2025) → Nearly 140 (April to May 2026)
- Time to validate a flagged cross-border payment: 5 to 6 minutes → Under 30 seconds
- Open payment investigations: Down nearly 80%; no starting count
- Reply to a client payment inquiry: Days → Same day
- Cost
- Not disclosed for any agent: no build effort, running cost or supervisor time. Microsoft gives BNY's whole technology spend, $3.8 billion last year, nearly 19% of revenue. About 2,300 employees took a 40-hour AI course in paid hours.
- People
- Each agent has a human supervisor; a person reviews each corrected payment and drafted client reply. Staff who repaired payment exceptions now analyse data-quality patterns and help test AI tools, Microsoft says. No team headcount before or after.
- Evidence
- Vendor-reported
Published by the company selling the product, usually with the customer. Wins are selected; cost is usually missing.
What this leaves out: Published by Microsoft, which sells BNY cloud services and Copilot; the figures are BNY's own, quoted from one manager, with no period, payment volume, error rate or cost. The 80% fall has no starting count. The agent counts come from BNY's earnings calls as reported by PYMNTS.
BNY, the US custody bank, runs nearly 140 AI agents that it calls "digital employees" and manages them like staff. Each has login credentials, an employee number in the corporate directory and a human supervisor who assigns its tasks and reviews its output, according to [a Microsoft profile of the bank published on 12 May 2026](https://www.microsoft.com/en-us/worklab/the-making-of-a-frontier-firm-how-ai-is-redesigning-work-at-bny). Microsoft sells BNY cloud services and Copilot, so this is a seller's account of a customer.
The count is the bank's own. Chief executive Robin Vince told investors on 16 October 2025 that BNY had "more than 100 AI agents in use", [PYMNTS reported](https://www.pymnts.com/news/artificial-intelligence/2025/bny-accelerates-deployment-of-ai-solutions-and-agents/). After the first-quarter call on 16 April 2026, [PYMNTS reported](https://www.pymnts.com/earnings/2026/bnys-ai-strategy-signals-a-new-era-of-platform-banking/) roughly 140 digital employees and about 220 AI tools in production. PYMNTS [repeated the 140 figure on 9 October 2026](https://www.pymnts.com/news/artificial-intelligence/2026/banks-put-ai-agents-through-performance-reviews/); I found no newer count on a source I can cite.
One narrow job, with a person checking
BNY's payments team made "roughly thousands" of manual fixes a day, Microsoft writes: missing country codes, malformed addresses, garbled routing fields. Robotic process automation, software that replays a person's clicks screen by screen, kept breaking on the bank's systems. "Maintaining the bots consumed so much effort that the humans were effectively doing two jobs instead of one," says Rachel Lewis, who leads AI enablement for BNY's operations.
The first payments agent, built "last spring", has one task: read the vendor address on a cross-border payment, work out the country, validate the country code and submit the corrected payment for a person to review. Lewis says validations that took five or six minutes now clear in under 30 seconds, and open payment investigations fell by nearly 80%. A second agent drafts replies to client payment inquiries and queues each one for a manager; Microsoft says an inquiry that waited days is now answered the same day.
How the agents are managed
- Entry gate: 16 steps of review boards, compliance gates and model risk evaluations (checks on how a model's errors could harm the bank) stand between an idea and a working agent.
- Scorecard: supervisors watch a live dashboard of processed volume, cycle times, missed transactions and approval outcomes.
- Correction: when reviewers kept sending back the payments agent's responses, its team answered with "coaching, recalibration, retraining". Microsoft says the responses were not wrong; reviewers had habits from the old way of working.
- Ownership: the team of Michelle O'Reilly, BNY's head of talent, is developing a "Digital Employee Agency" to decide whether a job goes to an employee, a contractor or an agent, with onboarding, scorecards and offboarding for agents. It is in development, not running.
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
The people side
About 2,300 employees have completed a 40-hour AI course in paid work hours since April 2025, each required to build a working prototype: 2,300 x 40 = 92,000 paid hours. Staff who repaired payment exceptions now analyse data-quality patterns and help test new AI tools, Microsoft says. It gives no headcount for that team before or after. BNY "still hires thousands of human employees every year" and expects its workforce to change shape "from a pyramid to a diamond", with repetitive processing work at the base going to agents.
The catch
No cost is given for building or running any agent. The only money figure is BNY's whole technology spend, $3.8 billion last year, nearly 19% of revenue. There is no error rate, no count of missed transactions, no measurement period and no payment volume, and the 80% fall has no starting count. The sources do not say how long the 16 steps take, how many proposed agents fail them, or how many agents one supervisor carries.
What to do if you run an AI programme or an operations team
- Name a human supervisor for every agent already in production, and write down what that person reviews and how often.
- Give each agent a four-line scorecard before you add another: volume processed, cycle time, misses, and the share of outputs a reviewer approves.
- Log why reviewers reject an agent's output, and separate wrong answers from "not how we used to do it". BNY says its rejections were the second kind, and retrained on that basis.
Take this to work
Forward this
BNY runs nearly 140 AI agents as "digital employees", each with a named human supervisor and a live scorecard, and a new one passes 16 review steps before it starts work, according to a Microsoft profile of 12 May 2026. BNY publishes no cost or error rate, but the management model can be copied at any size: no agent should run without a named supervisor, a scorecard and a record of why its output gets rejected.
Questions to ask
- Who is the named supervisor for each agent we run today, and how many agents can one person supervise alongside their own work?
- What does it cost to build and run one agent, and how many hours a week does its supervisor spend reviewing output?
- How often do reviewers send an agent's output back, and how many of those rejections are wrong answers rather than old habits?
- How long does an approval gate like BNY's 16 steps take, and how many proposed agents fail it?
- What happened to the headcount of the team whose exceptions the agent now handles?
The arithmetic
payments needing a manual fix per day x minutes saved per validation / 60 = staff hours freed per day; set against build cost + running cost + supervisor and reviewer hours per day
BNY's figures give at least 4.5 minutes saved per validation (5 minutes before, 0.5 after). Microsoft says only that the payments team made "roughly thousands" of manual fixes a day, and gives no build cost, running cost or review time, so the sum cannot be finished from the source. The training side can: about 2,300 graduates x 40 hours = 92,000 paid hours in the AI course.
Questions this article answers
How many AI agents does BNY have?
Nearly 140, which the bank calls digital employees, according to a Microsoft profile of 12 May 2026 and PYMNTS's report of BNY's 16 April 2026 earnings call. BNY's chief executive had cited more than 100 AI agents on 16 October 2025.
How does BNY manage its digital employees?
Each agent has login credentials, an employee number and a human supervisor who assigns tasks and reviews output, and supervisors track volume, cycle times, missed transactions and approval outcomes on a dashboard. A new agent passes 16 steps of review boards, compliance gates and model risk evaluations first, Microsoft reports.
What results has BNY reported from AI agents in payments?
BNY's Rachel Lewis told Microsoft that payment validations that took five or six minutes now clear in under 30 seconds and that open payment investigations fell by nearly 80%. BNY has not published the payment volume, the period measured or an error rate.
What did BNY's AI agents cost?
Not disclosed. Microsoft reports only BNY's whole technology spend of $3.8 billion last year, nearly 19% of revenue, and that about 2,300 employees took a 40-hour AI course in paid work hours.
Vercel put an agent on inbound leads: one rep now does the work of 10
Vercel says an agent researches and sorts its website leads, a rep approves every email, and nine of ten reps moved to other sales work. No cost is given.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.