Time to pay the AI piper

Banks spent the last two years pushing AI into every corner of the firm and urging every employee to use it. Now they have a new problem: It worked.

AI, it turns out, costs a lot on top of the billions banks have already spent building their tools: RBC’s token usage – the meter that runs based on AI use – is up 500% on the year, CEO Dave McKay said. At JPMorganChase, some employees are “spending more on tokens” than what they earn, payments CDAO Zachery Anderson said.

The bills are only getting more unpredictable. As tools become more agentic, a prompt may return a one-line answer or kick off hours of autonomous work. “Reasoning, the access to tools, the amount of context that you can put into it, your token costs do not scale on a linear basis,” CommBank CEO Matt Comyn said last week.

For risk-averse banks, the next move would seem obvious: Follow the lead of Uber or Coinbase and limit how much AI employees can use. Instead, lenders are making a different bet: that they can engineer a way to make the economics work faster than employees can run up the tab.

For CIBC, that means taking some choice away from users even as they get more functionality. The bank is piloting CAI 2.0, a new, more agentic version of its flagship AI platform that lets users complete longer-running tasks. So far, some users have let it do work that’s taken four hours and used between 10 million and 100 million tokens (for context, a 75-word response in a chat would typically burn about 100). “In this new product, the user no longer has a choice of model to use for their work,” Chris Patterson, head of enterprise AI platforms and solutions, told us. Instead, the bank built a way to classify each prompt by the type of task at hand – Excel model building, deck creation, research. Based on that, the tool auto-selects the model (and the version of the model) that’ll do it best and most economically. “You don't need a baseball bat to swat a fly when a fly swatter will do the job just as well and costs a lot less to swing,” he said.

TD Bank meanwhile has focused on rolling out the right tools to the right teams and meticulously tracking what happens. “You will have a shock if how you rolled it out is like, ‘we gave it to everybody to see what would happen,’” Jeff Martin, EVP of procurement, data and corporate technology, told us. The bank’s approach means it’s training managers to think about AI the same way they would other expenses. “Make sure that you've educated your team to use the right model for the right task,” he said. “So you're not just using it just because you can; you're using it because you should.” The bank also has a robust AI FinOps function (short for financial operations) that monitors trends in token usage and flags when the bank’s AI arsenal isn’t being used most efficiently. “We believe that efficiency is what is going to make AI scalable,” Martin said.

The longer-term push for banks is to stop renting every token from AI labs, at least for some tasks. Labs have been changing the way they price tokens and pushing the cost of work up in recent months (see: “Fat bill coming,” The Brief, April 30). Banks have long run some AI models locally, but the spikes in usage and cost are forcing them to lean further into open-weight options that don’t need to ping an OpenAI or an Anthropic API for every task. “We will have our own GPU compute,” said PNC CEO Bill Demchak at a conference this week. “We will not be as reliant on burning external tokens than what we will do internally.” The added benefit: External models don’t get close to their data.

For now banks still say the value is worth the AI bill. But as Uber showed, costs can go from manageable to alarming quickly. That calculus is now “part of the worry set of a risk-adjusted CFO,” Dermot McDonogh, BNY’s CFO, said this week. Each day is a test of that balance.

“The impact that AI can have on the productivity of a bank…can be taken away by the cost of tokens,” said PNC’s Demchak.

BNP Paribas renewed its partnership with Mistral AI two weeks ago, extending a relationship dating to 2023 from model access into co-developed software and solutions. One of the first products of the new arrangement is an agentic KYC tool for corporate onboarding in Belgium. In this week’s “Corner,” we sat down with Sophie Heller, chief transformation officer of the bank’s commercial, personal banking and services division to see how it works and why one of Europe’s largest banks is betting on Europe's largest model maker.

Use case: KYC agents
Vendor: Mistral
Bank: BNP Paribas

Why it’s interesting: The bank has a limited amount of time with Mistral engineers and saw KYC as a process with enough upside to warrant extra focus. “I'm responsible at CPBS (commercial, personal banking and services) level to select the project on which we are going to use the Mistral AI,” Heller said. “We won't want to waste skills on projects that will not be valuable or not complex enough to justify.”

How it works: The tool splits corporate onboarding between two orchestrated agents. The first extracts and structures client information and pre-fills the application. The second runs consistency and completeness checks before the file goes to the back office. The relationship manager stays in control throughout and remains the client's sole point of contact, and no agent ever talks to the customer.

How they did it: The pilot was one of the first "sprints" under the renewed partnership: Three Mistral engineers were paired with three BNP Paribas engineers for 12 weeks. The bank brought the domain knowledge. “There has been a lot of work before to really look at the process, look at the numbers, the number of documents, the timing, the cost, the problem of quality to determine what were the steps that needed or could be automated first,” Heller said. Everything runs on-prem on the group's centralized GPU infrastructure, using the smallest Mistral models that can do the job, she said. With this setup, banking data never leaves the bank and costs are controlled.

By the numbers: “What used to take weeks is more around days at the moment,” Heller said, though cautioned it’s only in a proof of concept phase.

Bigger picture: Like any pilot, the solution isn't guaranteed to scale, and BNP is running rival corporate-KYC experiments elsewhere in the group. Which one ultimately wins out “will be a balance between performance and cost,” Heller said. Regardless, the bank will get multiple uses for the agents. “[KYC] was complex enough to say, okay, if we can do it on that process, which is very complex, probably we can do it also on other processes,” Heller said. “We are running lists of all the future candidates, the idea is each quarter to have new experiments with them.”

That’s the amount an AI agent – spun up by ING and Mastercard – spent on orchestra tickets this month in the first agentic transaction run entirely on infrastructure based in Europe. Agentic commerce – the futuristic process where an agent will search for products on your behalf, compare options and ultimately pull the trigger – has been gaining steam in 2026. Morgan Stanley projected these agent-led transactions could make up 20% of all ecommerce, or $385 billion in sales, by 2030. Banks and payment providers are racing to plant their respective flags as first movers: Nordea was the first in Finland this week, Santander had Europe’s first back in March, and DBS and Visa partnered up to do the same in Asia earlier this year (see: “Agents hold the purse strings,” The Brief, March 26).

Yes, but: These “live” purchases all still happen in a sandbox environment, and as trust in AI dwindles, the public and merchants are going to need a lot more convincing than proof that this kind of transaction can happen. “I don't think people are gonna delegate their purchasing to agents just yet; let's not forget that the ability to automate everyday purchases isn't new,” said Marianne Lake, JPMorganChase’s CEO of consumer & community banking this week. “You remember the smart refrigerator that's going to do all your shopping for you. The Subscribe & Save, where you now open the closet in the corridor and you have 200 years' worth of toilet paper…I'm not being dismissive. I'm just saying that customer choice is important, customer protections are important.” 

JPMorganChase is bringing on Tahir Zafar, who was the international head of AI at Japan’s Nomura. Zafar will report to Deep Thomas, another former Nomura executive, who joined the bank in August as chief data and analytics officer for the Asia-Pacific region. It’s another APAC hire for the bank, who also brought on Gautam Gorki as chief analytics officer for the region in April (see: “New chiefs,” The Brief, April 23).

Standard Chartered promoted Guillermo Veiga to group chief information and operations officer, where he’ll be tasked with bringing “technology and operations together under one mandate, covering Wealth & Retail Banking, Corporate & Investment Banking, and our Functions,” he wrote on LinkedIn. Veiga has been with the bank since 2023 after stints with AWS and Santander. He’s based in Singapore.

Goldman Sachs lost its head of AI research, Bing Xiang, to Siemens. He’d been with the bank since 2023 and led the AI research and applied AI teams. In his new role, he’ll be leading a new team “focused on physical AI for industrial applications,” he wrote on LinkedIn.

Ashish Garg, head of data architecture and agentic AI platforms at JPMorganChase, is leaving the bank. In his four-year tenure, he “helped establish the foundation for agentic AI at the firm through the agent builder platform,” he wrote on LinkedIn.

Citi hired Gela Fridman as head of AI and services for U.S. consumer cards. She was previously a director of new initiatives at Amazon Retail. The bank also brought on Nadiya Konstantynova as COO for Citi Wealth. She was a partner at McKinsey.

Mani Iyer joined Vanguard as chief AI and technology officer. He was most recently global head of infrastructure, data and AI technology and developer platforms at PayPal.

“They call it the fourth narcissistic wound that was inflicted on humanity, the first one was Copernicus when he found out that the Earth was not the center of the universe. The second one was with Darwin when we had to accept that we were just another kind of animal. The third one was Freud and having to accept that our ego was not actually fully in control of our minds. This is the fourth one with AI. Is conversation, creativity, intelligence something that is specifically human or not? It is intimidating, and we experience it in the bank with colleagues asking, ‘Hey, not only how do I use this, but is this going to replace me? Will I still have a role?’”

–Marguerite Bérard, CEO at ABN AMRO, at a conference, June 3

JPMorganChase is getting into physical AI, the much-hyped crossover between robotics and AI. At its lockbox sites – places where the bank receives checks or remittance documents – it has a robot that “can open envelopes, remove and unfold the contents, separate pages, remove staples, track content and scan the underlying documents,” the bank announced last week. The robot, which was built with start-up Ripcord, can handle 4,000 different types of envelopes and documents and chips away at the manual work of opening and scanning roughly 480 million documents and checks per year, the release said.

Anthropic released a “Mythos-level” model called Fable 5 to the general public, two months after saying it was “too dangerous to release,” and limiting access (see: “Much ado about Mythos,” The Brief, April 16). It’s not all about cybersecurity: On one coding benchmark that tests how well and how long models can complete software development tasks, the model was five and six times better than OpenAI’s GPT 5.5 and Google’s Gemini 3.1 respectively. It won’t come cheap: Its token prices are twice as expensive as Anthropic’s other top-tier model, Opus, and from June 23 onward, all subscribers will need to pay for it on a per-token basis rather than just paying for a blanket seat license.

Crédit Agricole earmarked €500 million ($578 million) to invest in AI through 2028, the bank announced Wednesday. As part of the push, it’s creating an AI Enterprise group that will establish “common technological foundations, enabling each entity to drive its AI transformation,” the release said. Core to the strategy is making sure that AI-driven efficiency gains are “not negated by hidden costs or uncontrolled dependencies,” a problem many banks are working through (see: “AI blows the budget,” The Brief, April 23).

One quarter of the 251 ultra high net worth individuals – those with more than $10 million in investible assets – surveyed by BNY said AI has made their relationship with their wealth advisor “more collaborative.” Collaboration is normally a good thing with an advisor, but that number may be worth keeping an eye on long-term: Wealth management is one of the hottest areas for new AI tools, and most are focused on making advisors feel irreplaceable to clients (see: “Not your father’s advisor,” The Brief, May 21). That relationship becoming more collaborative could cause some to question the value advisors bring. Then again, it could also make them more engaged and attached to the relationship.

ING built an agentic tool that assesses mortgage applications that haven’t been filled out correctly or have an atypical way of financing and generates a series of suggestions for how they can be fixed, the bank announced this week. It cut the amount of time employees spent manually scanning for issues and coming up with solutions down to 58 seconds, said Jacques Schram, senior manager of housing at the bank. Mortgage financing is a hot area for AI tools; TD Bank rolled an agentic tool out last month (see: “House rules,” The Brief, May 28). And CommBank is approving 70% of its mortgage applications automatically within the same day thanks to an “automated credit rules engine,” the bank reported last year.