Meet 3 really innovative bank research chiefs

The number of AI research papers published by banks each year has grown seven-fold over the last five years, the newly-released Evident AI Research Tracker shows. More than ever, these academic pursuits are being directly translated into use cases and returns for the bottom line.

But there isn’t a single blueprint to squeezing real value out of research. Here we look at three bank research chiefs, who take different approaches.

#1: DUAL THREAT

Manuela Veloso, head of AI research at JPMorganChase

Joined JPMC from Carnegie Mellon University in 2018. She created and leads banking's most prolific AI research team.

• The approach: Simultaneously tackle cutting-edge AI research and practical applications in finance – a luxury JPMC can afford as banking’s biggest employer of AI researchers. Among the top priorities for Veloso’s team: agents, synthetic data, human-AI interaction and cracking explainability and interpretability – ways of getting AI systems to detail why they acted a certain way. It’s a critical step for multi-agent systems and scaling AI into more heavily-regulated parts of the bank.
  
• The pipeline: Practical research like FlowMind, a way to use AI to generate workflows, Advanced Messaging Platform (AMP), a way to use AI to process emails, and Dynamic Pricing feed into some of the bank's processes and tools it has put in the hands of bankers.

#2: EXPLAINER IN CHIEF

Foteini Agrafioti, SVP, data and AI & chief science officer at RBC

Joined the Canadian bank in 2016 and co-founded Borealis, the bank’s research arm. She previously founded a biometric security company.

• The approach: Connect research to real life applications and business value as clearly as possible and employ a centralized team to ensure research continually builds on itself and doesn’t stall. Among Agrafioti’s priorities: building foundation models and partnering with more teams around the bank to deliver on CEO Dave McKay’s push to “take the capabilities of the top 10% and expand them to the top 80%.”
  
• The pipeline: When the bank unveiled its foundation model (ATOM) – which powers fraud and risk management use cases across the bank – it highlighted seven publications that led to its creation. The papers range from extending model memory to improving how the model labels and understands data.

#3 AGENT BOSS

Prem Natarajan, chief scientist & head of enterprise data & AI at Capital One

Blended academia with Big Tech experience, leading Alexa AI at Amazon before joining the bank in 2023.

• The approach: Use the bank’s research talent to anticipate use cases two to three years down the line and build a culture that rewards publications and proofs of concept. Among Natarajan’s top priorities: agentic tools and models that can be widely used across the bank.
  
• The pipeline: Capital One’s T1 dataset – which houses realistic conversations and step-by-step tasks that agents might be asked to handle – is a benchmark that’s useful for the bank as its agentic AI development ramps up. Earlier this year, it launched Chat Concierge, an agentic car-buying tool.

Since 2023, the 50 banks we track are adding experienced AI talent to their workforce faster than less-experienced counterparts.

What it means: Recent studies from Stanford and Harvard say AI is leading companies to eschew younger workers. Inside banks, AI hiring shows the same…sort of. AI headcount is still up regardless of experience level, and the pace of hiring is getting faster. It’s not that banks don’t need juniors anymore, but after years of building, there’s a premium on people with enough seniority to execute AI strategies. That's even more true at the most AI-mature banks: For the top three banks in the Evident AI Index, the rate senior employees have been added to the ranks doubles the rate juniors have been over the last two years.

"[AI is] providing a differentiated profile of what the entry worker looks like. We aren’t seeing a disappearance of entry-level roles, but what we are seeing is a reshaping of them."

- Michelle Vaz, MD, AWS training and certification at AWS, in an interview, Sept. 12

At Barclays’ Global Financial Services Conference last week, AI was the guest of honor. This time, it was largely CFOs – not CEOs – feeding analysts the brass tacks on what AI means for the bottom line. These were the three biggest takeaways:

#1 BALANCING ACT

Finance chiefs are still wrapping their heads around AI’s benefits on the balance sheet. “It's difficult to draw a straight line from those hundreds of thousands of hours [saved] to an expense growth rate,” CIBC CFO Robert Sedran said of the bank’s Gen AI platform. “As these tools mature, though, we should start to see more of that structural cost come out.”

AI wouldn’t be “a big expense efficiency machine in the next year or two,” BMO CFO Tayfun Tuzun said. But the long term effects would be “significant.” At RBC, significant means generating up to $1 billion in value from AI by 2027 – which would involve “driving cost down or driving efficiencies as well as top line productivity,” CFO Katherine Gibson noted.

#2 LABORED DISCUSSIONS

“Personnel costs [are] up, but personnel headcount [is] largely flat,” said PNC CEO Bill Demchak of his changing workforce. “The cost of people going up is the degree of expertise.” In other words, finding people who can execute the bank’s AI strategy is expensive.

Among those costs (and needs) are engineers. NatWest upped its engineering force from 3,000 to 5,000 this year, CFO Katie Murray said. PNC’s Demchak highlighted the need for “engineers and people who can describe the outcome they want,” a shift away from outsourced coders as agents handle more work. BNY CFO Dermot McDonogh said that instead of talking about AI taking jobs, the bank talks about using it “to create capacity so we can do more projects with the same resources.”

#3 JANE’S (AND EVERYONE’S) ADDICTION

Top execs are becoming obsessive about granular AI data. At Citi CEO Jane Fraser’s management meeting that week, “the first 20 pages of the discussion” was a “business by business, function by function” report of AI adoption at the firm, CFO Mark Mason said. Top of mind: the percentage of the workforce with access to tools and how it can “push the use of that more aggressively.”

The Citi C-Suite wasn’t alone being metric-obsessed: NatWest’s Murray noted that she keeps tabs on engineering productivity and “30% to 40% of that first level coding is all done via AI” at the firm.

In this week’s use case corner, we look at how agents are taking on different roles. A Lloyds tool uses an agent as a second set of eyes, while Mastercard is trying to let them be personal shoppers.

#1 AI-UDITOR

Use Case: Audit chatbot
Vendor: Aveni
Banks: Lloyds

Why it’s interesting: The bank is testing an audit chatbot built on FinLLM, a UK-specific financial LLM it developed with startup Aveni. It uses an “agent as a judge” approach: One AI model generates an answer, and a second set of models independently review and score the response before it reaches a user.

How it works: The chatbot answers auditor questions by searching Atlas, Lloyds’ internal audit data and documentation system. The answer is passed to an agent, which validates the answer against regulatory guidelines and internal compliance rules before returning it to the user. Humans remain in the loop, but the introduction of agents helps the banks scale quality control.

By the numbers: The tool is still being piloted, but the bank says the system cut the time it takes auditors to find relevant information and improved accuracy checks by layering AI reviewers. The bank plans to extend the “agent judge” setup beyond the audit tool to integrate financial expertise with other models like ChatGPT or Gemini in other use cases.

#2 PUT IT ON MY CARD

Use Case: Agent Pay
Vendor: Mastercard
Banks: Citi, U.S. Bank

Why it’s interesting: In April, Mastercard debuted “Agent Pay,” a way for agents to complete purchases on customers’ behalf. This week, the credit card company revealed that Citi and U.S. Bank customers can be the new tech’s guinea pigs.

How it works: Agent Pay is the infrastructure that lets agents send money from your account to a retailer. If a cardholder were planning a party with an LLM, a chatbot might ask for a budget and put together a list of supplies and a menu. Agent Pay lets it take the next step and actually place the orders. To increase the ecosystem where agent payments are accepted, the company also launched an Agent Toolkit, a set of tools that make connecting to the card company’s APIs easier.

By the numbers: Mastercard says it will give all cardholders access to Agent Pay by the holidays. The real hurdle is customer trust: Only 24% of people said they feel comfortable with AI making purchases on their behalf, a Bain survey found earlier this year. Still, Paypal and Visa are forging ahead with similar tools, and Google this week introduced Agent Payments Protocol, a way to authenticate this growing form of payment.

In this segment, we explore trends in AI development and adoption beyond the largest banks or, as we did last time, in regions our Index doesn't currently cover.

“I do not think we are behind the curve [on AI],” proclaimed Associated Bank (1) CEO Andy Harmening last week. As declining model costs lower AI’s barrier to entry, his rallying cry is getting truer for smaller firms.

The Wisconsin-based bank is using AI in document processing, legal work and call centers, and recently set up an AI council to better connect the tech to business cases. That drive to use AI to solve real problems is familiar for big and small banks alike, but it’s perhaps even more immediate for smaller firms without billions earmarked for tech.

After finding they were “overwhelmed with phone calls,” Fifth Third Bank (2) built an AI chatbot for call center employees that saves time by recognizing customer intent 90% of the time, according to Michelle Grimm, the bank’s senior director of conversational AI. Ally Financial (3) has gone further, launching an enterprise-wide Gen AI platform that houses multiple tools: one, for example, that lets marketers build creative campaigns, another that lets auditors design new risk metrics (see: “Multitooling,” The Brief, August 7). Huntington Bank (4) CFO Zach Wasserman weighed in on value, saying Gen AI will cut costs and grow revenue by up to 15%.

Others have turned to big bank veterans to build AI strategies: M&T Bank (5) hired Deutsche Bank’s Andrew Foster, who has focused his two years at the bank on new training programs, governance guardrails and data partnerships. And Raymond James (6) last month plucked David Solganik from RBC to head up AI strategy.

Bottom line: Smaller banks are taking the AI lessons (and talent) from big banks and making it work in their own way.

Citi hired longtime IBM-er Shobhit Varshney as head of AI. Varshney was the head of data and AI for IBM’s consulting arm. At Citi, he’ll report to Anand Selva and will work “in lockstep” with CTO David Griffiths. It’s more AI muscle for Citi, which reorganized its AI leadership in June.

Krzysztof Uryn joined Santander as the bank’s U.S. CTO. Uryn spent 22 years at Accenture in Poland.

KKR is continuing to recruit AI talent to private equity. The firm hired Gaurav Hind from Google in May to lead its AI efforts. Now, it’s looking for a head of AI product management and a head of AI platforms.

Claude can now create spreadsheets – and does it well enough that Microsoft is now using Anthropic to power certain Office tools. Among the fans of Claude’s new skills is Ethan Mollick: “Claude’s new ability to work with Excel files is the best I have seen so far,” he wrote on X after the launch. It comes as the company reports that 77% of enterprise customers are using their tech to fully automate tasks rather than collaborate.

BNY is setting up an AI lab at Carnegie Mellon University to do joint research and recruit new talent. The five-year, $10 million partnership will focus on “developing technologies and frameworks to ensure AI systems are robust, resilient, and trustworthy” and aims to “accelerate the path from research to real-world deployment,” a BNY spokesperson told us. It follows several other high-profile academic partnerships – like RBC’s tie-up with MIT and Capital One’s with the University of Illinois – as banks look to shore up talent pipelines (see: “Bookish bankers,” The Brief, July 24).

BMO wants to understand where AI helps employees most, writes Kristin Milchanowski, the bank’s chief data and AI officer. The bank is surveying employees about tool usage to get an AI-specific net promoter score, examining all points of AI contact and scoring tools’ impact and recording how often employees override suggestions to gauge trust in the tech.

Revolut is all in on agentic AI, noting on job postings that it’s building call center tools, outbound sales tools and voice assistants “from the ground up.” It’s been a busy month for the U.K.-based firm, which brought aboard former SocGen CEO Frederic Oudea as its West Europe chairman and is reportedly considering buying a U.S. bank.

Last time, we asked you if you agreed with futurist Brett King’s assertion that humans wouldn’t be operating the banking system in 30 years. Was that timeline spot on? Too bullish? Completely off?

Your verdict: Don’t lose sleep – at least not on King’s timeline.

We live to fight another day, dear banker.

Mon 22 - Weds 24 Sept

Responsible AI Summit, London

Mon 29 - Thurs 2 Oct

Sibos, Munich

Mon 17 - Tues 18 Nov

Momentum AI Finance 2025, New York