Invest1 publisher3 min readPublished Updated
Bank of America plans to double AI spending that must clear a 16-pillar risk review
Bank of America will double its AI budget next year, CEO Brian Moynihan said; about 140 uses now cost $400 million and yield $800 million in benefit. Its technology chief, Hari Gopalkrishnan, says every AI project goes through a review covering 16 categories of risk.
The Investor · Invest desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Chief technology and information officer Hari Gopalkrishnan told Fortune's AIQ Summit that rushing to AI is one of the biggest mistakes he sees when deterministic models work well enough.
- The bank maps the steps behind each client request first and often picks a mobile app or a real-time decision rule over an AI model.
- An orchestration layer called Orchestra sends simple classification to open-weight models on the bank's own GPUs and harder reasoning to proprietary models.
- Gopalkrishnan said the Erica assistant has handled 3.6 billion transactions and that without it the bank would need 11,000 more people answering calls.
- S&P Global's Sally Moore said clients need accuracy, citations and traceability back to the source, since much of S&P's data feeds regulated workflows.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost Keeping the two-to-one return on a doubled budget would take about $1.6 billion of benefit, twice the $800 million the bank claims now.
- constraint Because requests are screened for simpler fixes before any model is chosen, the number of AI projects is set by what survives that screen and the risk review.
- decision Running classification on open-weight models on its own GPUs makes each task a cost choice, so token spending tracks the share of work sent to proprietary models.
Moynihan's September figures imply a ratio: about 140 uses, $400 million of cost and $800 million of benefit [2]. Each dollar spent returns two. The average use costs roughly $2.9 million and yields about $5.7 million [1][2]. If the $400 million is the base being doubled, next year's AI expense lands near $800 million [3].
Gopalkrishnan said the bank turns AI down "plenty of times" [3]. Projects that go ahead pass a review of 16 "pillars" of risk, among them privacy, bias, workforce impact and intellectual property [4]. "We're not going to implement a chatbot that only answers to certain accents," he said [5].
All four pillars Fortune named are risk tests [4]. A project can pass on privacy and bias and still earn well under the $5.7 million average [2]. Cost control sits somewhere else, in the Orchestra routing layer. Gopalkrishnan said keeping classification on open-weight models helps control token costs [8].
Where the extra money goes decides whether the ratio holds. It can scale uses the bank has already proven, such as the Erica assistant [6]. (A March bank press release counted more than 3.2 billion client interactions [7], a different measure from his 3.6 billion transactions.) It can fund a longer tail of projects that clear the risk review but return less, pulling the ratio toward one. Or it can go to agents, where the bank is staying with assistive tools for now. "There is so much juice to be squeezed right now with assistive agents that are actually working with humans in the loop," he said [9]. The bank will go further as control infrastructure improves, he said [10].
I'd expect the first case to take most of the money, because the process inventory and the routing both push easy work to cheap tools before any model budget is spent [3][8]. The counter-thesis is that a doubled budget creates its own pressure to find uses, and a screen built on risk does not stop a safe project with a thin return. If benefit roughly doubles alongside spending, the gate worked on return as well as risk. If benefit stays near $800 million while spending grows to the same size, the review screened for safety only [3].
Fortune's account covers one bank's process and does not show whether other regulated firms run a comparable review. Sally Moore, S&P Global's chief client officer, described the demand from the supplier's side. "Data is the currency within AI," she said [11]. In her example, an unnamed tier-one bank with 8,000 bankers went from about 60% accuracy to 98% with S&P's help, on S&P's own figures [13]. Its error rate fell from roughly 40% to 2%, a twentyfold drop [4].
What to watch
- Whether Bank of America names the other 12 of its 16 risk pillars, and whether any of them measures financial return.
- When Bank of America moves past assistive agents, which Gopalkrishnan tied to better control infrastructure, and what that spending returns.
- Whether other banks publish project-level AI review criteria comparable to Bank of America's.