AI inflation spike

Anthropic is raising the price of AI at the pump for banks. It’s going to expose whether lenders’ tools run more like Hummers or hybrids. 

Dario Amodei’s lab last week changed its pricing model for enterprise customers, telling businesses they’d shift from selling per-seat licenses to billing based on how much AI they actually use. This is the age of Claude, and Anthropic feels the coding tool has become so indispensable that they can pretty much set the price – for now, barring the emergence of another, better option – however they want.

This price hike isn’t just adding some extra line items to budgets; it’s forcing banks, or at least ought to be forcing them, to reckon with how they built all their tools in the first place. Back in the good old days – say, a week ago – lenders focused on building AI tools fast, knowing the cost of actually running them largely sat with model-makers (see: “Stat of the week,” The Brief, April 16). But if banks are now on the hook for the usage bill, every clunky workflow comes with a real cost that compounds with every use. That could end up limiting how far banks actually scale their tech.

Until now, the inefficiencies in their tools were largely invisible, at least to AI’s bottom line. Banks could build a Rube Goldberg machine or a straight ramp, and it didn’t really matter as long as it moved the ball from one place to the other. When each added step – a different model used, a handoff to a different system – has a cost that the bank needs to pick up though, gains get eroded quickly.

Capital One has been working to get ahead of that shift. The bank put together a group of engineers to audit how AI tools work, much as they did for cloud in the past, with a sole focus on cost discipline. “The fun part for my team is, what levers can we pull to leverage this technology in a more efficient way?” said Jerzy Grzywinski, a senior director of software engineering who leads financial operations (FinOps) for the bank. “How do I get engineers to think about efficiency as part of how they build, what they build, how they maintain their products?”

BNY is focused on the risks if banks don’t. With AI, you could see “the same story that we’ve seen before with cloud,” CEO Robin Vince warned during the bank’s earnings call this past week, before saying the bank’s scale allows it to spread its resources between labs. “If you allow yourself to get locked in and you don't have the breadth of access, you're taking a real risk.”

That flexibility helps, and banks that built model-agnostic tools where LLMs can be swapped in and out will avoid some of the pain of the twisting knife. But if other AI labs follow Anthropic’s lead – and it looks like they might – the real problem isn’t vendor choice but how these tools were built. Any bank that doesn’t go back through their tools to strip out the waste will find out the hard way that you can’t go very far on an AI gas-guzzler.

Citi posted its highest revenue in a decade during last week’s earnings, driven in part by an 11% revenue jump from its wealth management unit. In this week’s “Corner,” we sat down with Joe Bonanno, head of wealth intelligence at Citi, who talked us through how the bank is redesigning the wealth management experience around AI tools and his agentic plans moving forward.

Use case: Citi Wealth AI tool suite
Vendors: Google Cloud and Palantir
Bank: Citi

Why it’s interesting: The bank has been rolling out new applications for its wealth managers for the better part of a year now (see: “Wealth helper,” The Brief, Sept. 4), but the updates to existing tools – like the Ask Wealth chatbot and Client360, which gives employees tips on how best to engage clients – are all focused on marrying AI use into one seamless experience, Bonanno said. At the same time, it’s putting AI in front of customers with tools like Portfolio Intelligence, a dashboard that serves them content about their investments. “We took this opportunity to really rethink, redesign, reinvent, essentially, some of the core capabilities that were either lacking or things we wanted to do in a modern sort of world,” Bonanno said.

How it works: The bank built a centralized data layer that ingests all the signals about a client and their portfolio, Bonanno said. Agents monitor for anything that might warrant some kind of interaction for the bank – exposure to stocks affected by the conflict in the Middle East for example, or something as simple as an upcoming birthday. “We have agents that are looking for anomalous patterns, behaviors, opportunities, risks or threats,” Bonanno said. That information is then pushed through the different systems. It might put a client at the top of a call list, as determined by the Advisor Insights tool. Or it might be used to craft a message that shows up on the Portfolio Intelligence dashboard the next time a client logs in. 

How they did it: The bank first rebuilt the underlying architecture, focusing on unifying client data from different businesses around the globe, which had previously been siloed. “We have this warehouse now,” Bonanno said. “Every trade, transaction, position, balance, email, click is organized.” The bank then matched that up with its own internal guidance so that when advisors used tools to figure out how to answer client questions or recommend positions, they were doing it with all the information they had. “It will actually take the brain of [Citi chief investment officer] Kate Moore and all of the collateral we have and make the same recommendation to the advisor,” he said. 

By the numbers: The bank has just started rolling out the tools, but so far, Bonanno said the biggest benefits are coming through in time saved – though the true test will be in whether the cohort using these tools will be able to bring in more revenue. He’s confident in early results: “We're doing very well,” he said. “Let's put it that way.”

Bigger picture: The line between client-facing and employee-facing use cases is getting blurrier as banks redesign full processes around AI, as Citi is. The bank showed just how much this week, announcing Citi Sky, a forthcoming video avatar tool that will let clients make certain financial moves – like moving money into different savings accounts – or schedule time with advisors. “We're building everything with an agent orchestration, or an agentic way,” said Bonanno. Updates to the set of advisor tools will be built with that in mind, he said.

The share of the code BNY pushed in Q1 that was generated by AI, the bank revealed during its earnings call. That translated into a 10% increase in how much software the bank was able to release. Other banks have shared how much code is AI-written – like NatWest, which came in at 35%. But no other bank has explicitly compared input to output the way BNY did here.

Go deeper: The back of the envelope math suggests a 4-to-1 ratio between code generated and gains in output. But game that out further: If the bank could hypothetically produce 100% of its code with AI, it would boost software production by 25%. That's a big gain, but it goes to show how much more productivity uplift there still is in software development beyond code generation – testing, deployment and compliance to name a few. Those are still on the horizon as CEO Robin Vince focuses on getting his 47,000-person company “to be able to deliver like one that is, in fact, many times larger.”

Lloyds brought Sameer Gupta on as chief data and AI officer, replacing Ranil Boteju, who left for CommBank earlier this year. Gupta was previously chief analytics officer at Singapore’s DBS and at Lloyds will focus on “managing how AI is adopted,” the bank wrote. He’ll report to the bank’s COO Ron van Kemenade.

BNY hired Adam Waksman to be MD of AI & ML engineering. He was previously global head of developer platform and AI transformation at BlackRock, where he led a 450-person team that was part of the Aladdin business, he wrote on LinkedIn. The bank this month also brought on Chetan Kotwal to be global head of the bank’s architecture hub. Kotwal joins from index company MSCI but spent five years with BNY earlier in his career.

Westpac hired Maggie Shi as chief AI innovation officer, a “newly created role focused on helping drive AI and GenAI adoption at scale,” she wrote on LinkedIn. She was previously a distinguished engineer at CommBank.

Gautam Gorki is now JPMorganChase’s chief analytics officer for Asia Pacific. He was most recently with Deloitte, where he'd been since 2023 and spent nearly a decade with DBS before that.

Lloyds also announced an expansion to its responsible AI team, led by Suzanne Brink, this week. The release highlighted nine “new joiners” across three units under Brink: One group focuses on the bank’s assurance framework; a research and development arm explores the new RAI techniques and a mobilization group that builds automated controls.

Henry Fuheng Wu joined Wells Fargo as executive director on ML/AI and senior manager on quant analytics. He was previously a generative AI lead at Oracle.

“We won't create a lot of science fair projects and kind of use some of this new technology in ways that actually don't deliver real benefits.”

– Bruce Van Saun, CEO at Citizens Financial, during the bank’s earnings call, April 16

Anthropic’s Mythos preview will soon be made available to banks in Europe, Reuters reported this week. U.S. counterparts – like JPMorganChase, Morgan Stanley and Goldman Sachs, all among the contenders for a spot in Anthropic’s future IPO – have spent the last couple weeks experimenting with the lab’s “too dangerous to release” cybersecurity model. In one bank’s case, the model surfaced a year’s worth of work patching vulnerabilities and updating systems (see: “Mythdirection,” The Brief, April 16). The rollout comes amid new reports showing that unauthorized users have been using the model since it was first rolled out – putting even more pressure on non-U.S. banks to get their hands on it.

AI’s academic set are descending on Rio de Janeiro for ICLR, one of the top AI research conferences, which starts Thursday. The bank contingent rubbing shoulders and drinking caipirinhas won’t disappear. But after several years of becoming more of a fixture, the Wall Street presence looks to have plateaued. After years of growth, the number of papers published by the 50 banks we track was flat this year, Evident analysis shows. JPMorganChase, long the most prolific bank at this conference, cut its paper output this year by 75%. Don’t mistake this for a pull back though: Banks are still courting academics and putting more money than ever into university research partnerships (see: “Stat of the week,” The Brief, April 9). “It is only through research that a firm achieves the breakthroughs that will truly distinguish it,” Manuela Veloso, ex-head of AI research at JPMC told us last week (see: “Why AI research really matters,” The Brief, April 16). But as they move that research focus into the lines of business, they’re keeping those breakthroughs closer to the chest.

BlackRock launched RockAI, a new platform where employees can spin up agents suited to tasks. It rolled the tool out to 5,000 in-house developers last week and will make it available firmwide in the future, said Nish Ajitsaria, the asset manager’s senior managing director and head of Aladdin product engineering. The goal, Ajitsaria wrote on LinkedIn, is “AI handling more of the execution and teams focusing on context, oversight, and outcomes.”

Rabobank and BNY invested in Domyn, a European startup focused on data governance and ownership for AI systems. Neither bank was one of the most active venture investors last year, our Ventures Tracker shows. With the investment, Rabobank plans to use the tech internally: “We’ve also signed an agreement to explore how Domyn’s AI can be applied within Rabobank in the coming years,” Alexander Zwart, the bank’s chief innovation and technology officer wrote on LinkedIn.

Intesa Sanpaolo is fighting the mob with AI. The bank this week said it was working with law enforcement, including the anti-mafia investigative unit, to develop an AI tool to probe the accounts and monitor the transactions of more than 40,000 customers. The bank said the pilot – which is the country’s “first significant public-private partnership” to combat financial crime – surfaced 30 potential tax fraud cases.