Bank of America’s GenAI plan wants to avoid ‘sins of the past’

Last week, I sat down with Bank of America’s Ashok Krishnan and Duncan McInnes for a wide-ranging conversation on our Waters Wavelength Podcast. I found them to be open and insightful, though as their adept interviewer, I may be biased, so I think you should listen to the full conversation. But in case podcasts aren’t your thing, I wanted to highlight one part of the interview.

Most of our talk had nothing to do with artificial intelligence; it focused on how Krishnan, head of global markets platform and electronic trading, and McInnes, head of global markets technology, work together to build a culture that promotes efficiency and provides the tools to generate alpha.

We did, though, eventually get to AI, machine learning, generative AI, and agentic AI. We all agreed that while much of the conversation today revolves around GenAI, it’s more of a research assistant and a means to reduce time-consuming, repetitive tasks. That’s not a bad thing, but as McInnes noted, it “doesn’t necessarily lend itself to time-series analysis.” That’s not a knock; it’s simply to say that more traditional forms of AI are being used to improve trade decisions, manage risk, or improve back-office processes—at least for now.

Nonetheless, an important piece of the work that the pair’s teams are working on is being “consistent” when building GenAI tools that “everybody” will want to use, such as ingesting PDFs, distilling emails, and other everyday tasks not unique to sales or trading.

“[This way], hopefully we will get ourselves out of the problem of sitting here in 10 years’ time saying, ‘Ok, how are we going to unwind 14 of those different solutions to the same problem?’” McInnes said. “It’s an interesting opportunity for us to try and avoid some of the sins of the past as we roll this through the Reimagine program.” (More on the “Reimagine program” in a bit.)

I said that “sins of the past” is an interesting phrase. Whenever I speak with bank technologists, there are common refrains. Technical debt is a struggle to manage, so they need to sunset legacy systems. But as they build new, sophisticated systems, they need to make sure there’s some semblance of interoperability. Otherwise, they end up in the same situation they had hoped to leave: saddled with technical debt, aging systems, and a maximum speed of slow.

In a meandering way (as to which I am prone), I asked about the challenges of building AI systems while reining in tech sprawl and technical debt. After all, it sounds easier said than done. And McInnes told me this:

“We have a whole separate thought process about just grinding out stale and legacy applications. It takes time, and you have to get behind it and gradually pull them out of the ecosystem. But they’re almost two orthogonal problems embedded in your question.”

I nodded my head. After the call, I grabbed my trusty Merriam-Webster dictionary and looked up the word orthogonal: intersecting or lying at right angles; having a sum of products or an integral that is zero or sometimes one under specified conditions; having a matrix that is orthogonal: preserving length and distance; statistically independent.

(People assume that because I’m a journalist, I have an expansive vocabulary; I do not. It took me almost eight years to graduate with a bachelor’s degree in journalism from Plattsburgh State University. Meanwhile, McInnes holds a PhD in computational theoretical physics and a Master’s in physics from Oxford. Orthogonal careers, amirite? No, seriously, am I?)

Anyway, the “right angle” of my question referred to what we had been talking about before my wandering question. (Thank you, Merriam-Webster, for giving me a different word for meandering.)

Krishnan, whom I’ve interviewed several times, including for this “Voice of the CTO” profile, was explaining the company’s “Reimagine” program. The aim is to take the bank’s various constituents—trading, sales, risk, operations, and finance—and put them in a room with the platform, technology, and quant teams and talk. If all of them were to start from scratch, how would they do things differently? What are the outcomes they would be looking for?

“Look at the outcome you’re looking for—and don’t worry about the how—but let’s look at what the outcome is, and let’s try and drive to that outcome in an effective fashion,” Krishnan said. “Most of our alpha generation is still under traditional AI, machine-learning techniques. But [with] generative AI and agentic, [that’s] largely [about] taking away toil and making sure that people are able to work at scale. So if you take a trader, he’s able to back-test 10 strategies at a time, or he’s able to have access to all of these datasets instantaneously. That’s the piece that is becoming really, really valuable.”

In the world of trading, maybe GenAI isn’t a direct alpha generator. But if it helps a trader be more efficient or improve results for clients, the definition of “alpha generator” (which does not yet appear in Merriam-Webster) starts to split hairs.

In that same vein, Krishnan mentioned salespeople. The “outcome” that they want is “to look intelligent and productive in front of a client.” To do that today, a salesperson needs to review a number of call notes, examine the client’s trading history and the trade history of their peers, and assess the market’s activity. They should also check whether the client has appeared on a program like, say, CNBC’s The Exchange. 

GenAI—and agentic AI along with the Model Context Protocol (read more on that here)—could theoretically distill all that information onto one sheet of paper. So, as Krishnan put it, you reimagine the day-to-day processes that a person has, and combine the intelligence of the platform, tech, and quant functions to figure out more efficient workflows.

This is just a snippet of a much broader conversation. But I believe the key is that, as AI continues to evolve rapidly, those left behind are those who are continuously playing catch-up to the present moment, rather than thinking ahead. It’s easy enough to say, but it’s more difficult in practice, especially at a massive organization.

I’ll end on this: One of the things Krishnan, McInnes, and I talked about was the idea that, sure, it’s great to create a culture where everyone buys in, but what happens when people like Krishnan or McInnes leave? How do you keep that discipline? And what happens when things go wrong? What happens when deep learning turns into generative AI turns into agentic AI turns into…whatever is next? Quantum AI?

Those questions might seem simple, but I’m sure they are orthogonal. Well, pretty sure…

The image accompanying this column is “Gardener’s House at Antibes” by Claude Monet, courtesy of the Cleveland Museum of Art’s open-access program.