BNY's CTO: AI's Real Test Is Systems, Not Apps
BNY CTO Leigh-Ann Russell says AI's adoption phase is over and the next challenge is engineering whole systems — as Gartner predicts 40% of agentic AI projects will be canceled by 2027.
By Grace Kim
4 min read

What's News
- Gartner predicts more than 40% of agentic AI projects will be canceled by end of 2027.
- BNY CTO Leigh-Ann Russell spent 18 years at BP before joining America's oldest bank.
- William Henry Merrill founded Underwriters Laboratories in 1894; the first National Electrical Code followed in 1897.
- Economist Paul David showed electric motors supplied under 5% of US factory mechanical drives by 1899.
- The Herald of Free Enterprise capsized in March 1987, killing 193 people.
By Grace Kim
Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. That figure frames the argument BNY's chief technology officer Leigh-Ann Russell is making: the AI industry has finished its adoption phase and now has to engineer whole systems, not scattered apps.
Russell, global head of engineering at America's oldest bank and a mechanical engineer who spent 18 years at BP, calls the current moment a turning point. "It's clear to me that there is a generational opportunity taking place right now with AI and, like other transformational technology before it, the early innings are critical. However, there will be many inflection points. Phase one was adoption and now we're moving to adaption. To engineering a system that both compounds in value over time but also reinvents itself," Russell said. "To me, this is the moment to be principled about what we build, who it benefits, but most critically how it builds trust."
Why the 1893 Chicago World's Fair matters to AI
The piece opens with a historical parallel. In spring 1893, a 24-year-old engineer named William Henry Merrill arrived in Chicago to inspect the World's Columbian Exposition, which displayed more than 100,000 incandescent lamps. Fires kept breaking out. Merrill ran his first product test in a small lab above a Chicago fire patrol station in 1894; that lab became Underwriters Laboratories. By 1897, insurers, engineers, and the electrical industry had agreed on the first National Electrical Code. Neither the lab nor the code produced a single watt of power — but electricity could not have transformed daily life without them.
Russell brings an energy-industry term to the argument: the interlock. In 1922, as London electrified and switchgear injured workers, engineer James Harry Castell borrowed the railway concept: a cabinet key stays trapped in the switch as long as power is on. Shipping learned the same lesson in March 1987, when the ferry Herald of Free Enterprise capsized off Zeebrugge with its bow doors open, killing 193 people. International regulators mandated bridge door indicators just over a year later.
The takeaway: mature industries never assume that because every part works, the whole will.
Edison's system and the slow productivity payoff
What Thomas Edison switched on at Pearl Street in lower Manhattan on September 4, 1882, was a complete system — dynamos, 80,000 feet of underground conductors, switches, sockets, meters, and safety fuses. Historian Thomas Hughes built his classic study, Networks of Power, around that observation: Edison's real invention was the system.
The payoff was still slow. Economist Paul David showed that by 1899, nearly two decades after Pearl Street, electric motors supplied less than 5% of mechanical drives in American factories. The productivity surge came in the 1920s, after factories were redesigned with a small motor on every machine.
Most organizations are in that swap-the-engine phase with AI, the article argues. One team builds an agent, another a chatbot, a third pilots a workflow tool. Together they produce fragmented workflows, inconsistent controls, duplicated costs, and new security exposure.
The four disciplines
Russell's four standard questions — Will customers love it? Can we build it? Will it create value? Does it survive in the broader ecosystem? — aren't enough for AI. She adds a fifth: Can we build and operate it safely as part of a larger system? Her framework:
- Intent. Start with a real problem. When building gets cheap, deciding what not to build becomes more valuable.
- Integrity. Clear boundaries, meaningful evaluation, visibility into behavior, and a human able to step in — the AI equivalent of the fuse box.
- Infrastructure. The layers around the model — data, integration, compute, controls, monitoring — matter more than which model is best.
- Impact. Measure what changed, and whether benefits covered the full cost of integrating, operating, and governing the tool.
The four disciplines form an interlock, not a checklist: intent shapes infrastructure, infrastructure makes integrity practical, and impact tells you whether the intent was sound.
For leaders, the practical advice is blunt: count systems, not demos; treat governance as engineering, like a circuit breaker; measure the whole operating model; and leave room to rewire. Merrill's motto — "Know by test, and state the facts" — is the standard Russell implies AI programs will eventually have to meet.
Original: ul.org
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,387 articles