BofA’s struggles with AI adoption reflect a broader problem in banking

Processing Content

• Key insight: The banking industry has underestimated how much structure AI requires to be useful at scale.
• Supporting data: Eighty percent of U.S. banks are planning to increase their AI spending.
• Forward look: Until perspectives and expectations change, the most powerful AI technology will fall short.

Leaked Bank of America emails reveal employees are struggling to use Nvidia’s enterprise AI software complaining they were sold “a Formula 1 race car” but expected to operate it with “local car mechanics.” With 80% of U.S. banks planning to increase their AI spending, the fact that one of the world’s largest banks can’t make advanced AI work might give the industry pause.

But Bank of America’s experience isn’t evidence that enterprise AI doesn’t work. Banks are fundamentally misjudging how AI should be deployed, who should be responsible for making it work, and how much variation a regulated industry can tolerate.

One of the biggest failings I see is treating AI like a tool that employees should adapt to. Maybe there’s some training, a few best-practice prompts get circulated or analysts are left to figure it out on their own. In theory that sounds efficient, but in practice, it creates inconsistency, fragility and new forms of operational risk.

AI isn’t like traditional software. Two analysts can use the same model, with the same data access, and arrive at different outcomes. Not because one is wrong and the other is right, but because experience dictates how narrowly the question is framed, what assumptions are embedded in the prompt, and how the output is interpreted.

Training, like Citi’s rollout to 175,000 employees last year, helps, but it doesn’t solve the biggest problem. Consider what effective training actually requires: Our data shows underwriters need to process approximately 20 loans before they develop genuine proficiency with AI tools. That’s not a one-hour workshop or a best-practices document. That’s sustained, structured practice with feedback loops built in. Most banks aren’t designing for that learning curve.

You can teach employees how to write better prompts, but you can’t train away subjectivity, and each bank and lender will have different policies. Institutional knowledge that AI won’t know unless it’s baked in. When outputs depend heavily on individual skill, you don’t get scalable productivity, you get analyst-to-analyst variance. 

This inconsistency gets compounded by how individuals respond to the technology itself. Some employees see AI as amplifying their capabilities while others view it as threatening their expertise or adding complexity without clear benefit. When adoption is optional or implementation is unclear, the skeptics often shape the narrative, and promising pilots stall out before they reach meaningful scale.

What’s going wrong is that banks are deploying extremely powerful systems but expecting generalist employees to operate them safely and consistently without redesigning workflows around the technology. The result is impressive demonstrations, followed by hesitation, uneven adoption and internal concern about reliability.

The real risk isn’t that AI produces bad answers. It’s that it produces answers that look reasonable, but are arrived at through undocumented logic, inconsistent prompts and ad-hoc usage. If two analysts reach different conclusions using the same AI system, which one is correct? Which assumptions are approved? Which process is defensible?

These aren’t hypothetical concerns. They’re exactly the questions regulators will ask as AI becomes embedded in credit decisions, risk assessments and investment analysis.

Seen through that lens, Bank of America’s struggle isn’t unique. Most banks experimenting with enterprise AI run into the same friction.

The solution isn’t more training or better prompts. It’s a reframing of responsibility. In banking, AI shouldn’t behave like a blank canvas. It should be constrained, productized and embedded into workflows in ways that minimize individual interpretation. The system should adapt to banking’s requirements for consistency and explainability, not the other way around.

That means fewer general-purpose tools and more domain-specific AI systems. More guardrails, not fewer. And a shift away from the idea that every employee needs to become an expert prompt engineer.

The lesson from Bank of America’s internal emails isn’t that AI is too complex for banking. It’s that banking has underestimated how much structure AI requires to be useful at scale. Until perspectives and expectations change, even the most powerful AI technology will fall short.

However, here’s the kicker; those that do master this now, will unlock unprecedented growth in a way others struggle to understand or match. New market leaders will emerge as major disrupters, not because AI failed to support the generalists, but because others approached AI as a fundamental change to how they work. They avoided treating it like a bolt-on and embraced the notion that people management, and how our organizations are designed, is at the heart of success for every major digital transformation.