Fat bill coming

AI labs promised enterprises that AI would be both better than humans and cheaper. Now what if only one of those is actually true?

In the last two weeks, OpenAI and Anthropic released new, more-powerful versions of their flagship models in GPT-5.5 and Claude Opus 4.7. OpenAI’s offering scored 30% better on a benchmark testing agentic tasks and coding than last summer’s predecessor. The cost of running it jumped nearly 300%, new analysis from Evident shows. Anthropic’s new model showed the same pattern: The costs are rising faster than performance.

This puts banks in a tight spot: Lenders have spent the last year rewriting their processes around agentic AI on the assumption that costs would eventually come down. If they don’t, the bills from adoption and “tokenmaxxing,” where employees compete on how much AI they use, will keep rising. And banks will need to prove these tools can justify their costs some other way.

You can hear the leading adopters of AI starting to warn against seeing cost-savings as the primary purpose of these tools: “I think it’s a bad idea to think that you’re going to deploy AI and improve your efficiency ratio,” JPMorganChase (#1 in the Evident AI Index for Banks) CEO Jamie Dimon told investors this month during the bank’s earnings call.

That’s not a loss of faith, but a concession to reality. Agentic systems do more because they have the ability to “think” harder and coordinate other systems to get more complex jobs done. Those extra steps add more cost. The only way that changes is if the underlying infrastructure powering them gets cheaper. Don’t count on that. New data centers are coming slower than expected and the chips used to train these models are getting more expensive with growing demand. AI could become “impractically” expensive, Oxford researcher Toby Ord warned.

That possibility is forcing banks to adjust. BNY, for example, says it’s now taking a Moneyball approach to cost discipline: “We’re not spending anything like some of the numbers that are thrown out there by other big corporate organizations,” said Leigh-Ann Russell, the bank’s CIO on a podcast this week. Instead the bank is pushing agentic AI adoption in parts of the bank where tools can have the biggest impact, she said.

The bigger question is whether those tools can outperform humans. On that, there’s a long way to go. But it’s coming: Capital One, for example, built an agentic tool for its call centers that it said isn’t just faster than humans but produces more readable, accurate and complete call summaries than them, too (see: “Stat of the week,” The Brief, March 26).

Until the costs start coming down though, that needs to be the case for every tool a bank has embedded and adopted across the business. If it’s not, banks might not be able to afford to keep using them.

The share of the 50 banks we track that put a new executive in charge of AI in the last 12 months, new analysis from Evident shows. That includes 10 banks that have made or announced changes since January, including one you can see in our talent section below. 

What it means: Almost all of these new hires came from outside the bank’s ranks – or at the very least, outside of the bank’s tech organization. Where exactly a chief AI officer sits within the organization still varies by bank, though an increasing number now report to the group COO. Their job descriptions show a changing understanding of where these executives can add the most value as banks rewire more of their processes around AI. The first person directly overseeing AI at a bank – sometimes a CAIO, sometimes not – largely focused on how the tools are built. Now, as banks push the tech closer to the lines of business, they increasingly need these central AI leaders to ensure tech gets adopted right across the firm.

That new remit comes through in the recent announcements: RBC's new AI head Bruce Ross is focusing on "the shift from early-stage AI projects to scaled, client outcomes.” Sameer Gupta, who was named chief data and AI officer at Lloyds earlier this month, has the “responsibility for how the technology is further scaled across the business.” And David Rice at HSBC, who was previously COO of commercial banking, is being asked to bring business know-how to “simplify processes” with AI rather than just oversee how the tech is actually built.

Still: The chief AI officer role may turn out to be a fun footnote when we look back at the banking industry in a decade. The faster a bank transitions to be AI-first, the less the role of AI chief can actually be defined. We don't, after all, have chief internet officers. And the demise of the chief digital officers and chief mobile officers (remember them?) in the last decade shows the blueprint of how responsibilities end up getting diffused around a business. But as the job descriptions show, banks have a lot of scaling yet to do before that becomes a reality. 

“One of the things software engineers are always trying to do, they hate doing: endless simple maintenance, taking care of dependencies and stuff like that… They try to automate it to get ahead of it so that the system can be made capable of seeing when the problem is coming and immediately get itself to go around it. I’m sure that A.I. is going to bring many more levels of that. That’s the upside. The downside is you spend more and more of your life arguing with robots.”

– Stewart Brand, author and futurist, on Ezra Klein’s podcast, April 24

Citi has a new CIO. Brian Saluzzo joined the bank from Google, where he was VP of core developer engineering. Saluzzo is no stranger to financial services, having spent nearly a decade at American Express and nearly two before that at Goldman Sachs. In his new role, he reports to Tim Ryan, head of technology and business enablement and member of Citi’s executive committee, who noted that Citi brought Saluzzo on because of his “deep expertise building and scaling enterprise technology.” Jonathan Lofthouse, who had served as Citi’s CIO since February 2025, is no longer with the firm, a bank spokesperson confirmed. Saluzzo’s hire is another sign of how close Google and Citi have become: Last September, Joe Bonanno, head of Wealth Intelligence, told us his team went to Google’s offices twice a week. And last week, the two firms unveiled Citi Sky, a new video avatar financial advisor (see: “New AI wealth suite,” The Brief, April 23).

ANZ is bringing in Kai Yang as chief data and AI officer  – one of the 10 called out above – beginning in July. He was previously chief data and analytics officer for Asia & Middle East at HSBC and, before that, spent more than 15 years at CommBank, including a stint as chief data officer. ANZ is making a habit of luring talent from HSBC: Yang will report to CIO Donald Patra, who joined the Australian lender from HSBC last year. CEO Nuno Matos is also an HSBC alum.

Australia’s NAB created a new AI science team that will “play an integral part in building rigorous AI agent architecture, evaluation methodology and new AI products,” said Mahya Knox, chief AI officer. The bank tapped McKinsey veteran George Mathews to lead the team as AI science executive. He’ll report to Knox and sit within the broader digital, data and AI team led by Pete Steel, who the bank brought in last year.  

Wells Fargo hired Saba Dossani to be managing director of enterprise data management and governance. She was previously chief data and analytics officer at Comerica.

Rushabh Doshi joined Standard Chartered as director of agentic AI acceleration and deployment architecture. He was most recently an enterprise architect at Unilever. 

BMO hired Ashvin Veligandla as director and product owner of applied AI data integration. He was most recently lead data analyst at the Federal Reserve Bank of Chicago.

Angie Lamarsh is now vice president of data and AI strategy in commercial banking at RBC, where she’ll focus on building client-facing and internal AI tools. She was most recently head of sustainability for personal and commercial banking.

Canada’s Cohere is set to acquire Aleph Alpha, Germany’s largest AI lab. The deal would value the combined company at roughly $20 billion, pushing it past France’s Mistral, which raised a series C last September at a $14 billion valuation. As part of the deal, German retail giant Schwarz Group will make a $600 million investment into the company’s next fundraising round. Aleph Alpha counts Deutsche Bank as both a client and an investor. The bank made a key investment into the lab in November 2024 and the firms are now working on tech inside the bank. “The use cases that we are currently talking about all circle around the topics of legal, risk and similar functions,” said Christoph Rabenseifner, the bank’s chief strategy officer for technology, data and innovation and head of corporate venture capital, at a conference last November. “This is exactly where we see the edge of a model that is secure that is running on-premises, where we have a lot of control around our data and not necessarily have to put the data on public cloud.”

Wells Fargo rolled out its AI platform, CEO Charlie Scharf told investors this week: "We've developed an internal platform which gives employees access to tools from other technology providers and gives access to multiple models or large language models," he said during the bank's annual meeting Tuesday, noting that users can access public and private data to "build and eventually deploy custom Gen AI solutions and agents." The platform has about 25,000 users, though Scharf didn't provide specifics on how often it was being used. CIBC, for example, has about one-quarter of the employees of Wells Fargo and logs 17,000 daily users on its AI platform CAI (see: “Stat of the week,” The Brief, March 5).

Rogo, the startup building AI tools to handle banking’s grunt work, closed a $160 million Series D funding round this week. It values the company at $2 billion, up from $750 million just three months ago. JPMorganChase has been investing in Rogo since its series B round roughly a year ago, and the startup counts JPMC, Wells Fargo and Bank of America as clients.

Customers Bank, a Pennsylvania-based regional lender, inked a multiyear deal with OpenAI this week, expanding on a previous partnership. As part of the new tie-up, OpenAI will be sending forward deployed engineers to work alongside Customers’ employees, similar to what Anthropic is doing with Goldman Sachs (see: “Why Goldman went with Anthropic,” The Brief, Feb. 12). CEO Sam Sidhu got an AI clone to deliver his prepared remarks on the bank’s earnings call this past week; the OpenAI partnership will be a test of what else AI can deliver. Sidhu’s expectation is that AI will improve the bank’s efficiency ratio from 49% down to the low 40s – following in Santander’s footsteps of tying tech to cost-to-income ratio improvements (see: “Stat of the week,” The Brief, Feb. 26).

Goldman Sachs barred its Hong Kong-based employees from using Anthropic tools as of a few weeks ago, the Financial Times reports. The move is aimed at curbing model "distillation,” where heavy use of one system can be used to train a rival. It comes on the heels of the White House accusing China of industrial-scale theft of intellectual property from U.S. AI labs.

ICLR, one of the top academic AI conferences, wrapped up Monday with a daylong workshop on AI in finance hosted by RBC and JPMorganChase, among others. 

Eric He, director of machine learning research at RBC, was on the ground for the conference and led the organization of the workshop. 

We texted him there to get his personal take (not RBC’s) on the event.