Agentic AI's generation gap

TOP OF THE NEWS

B OF A'S BIG DAY

Bank of America hosted its first investor day in 14 years yesterday. Here are the three biggest AI takeaways:

1. Adoption is the focus: The bank “deployed a state-of-the-art foundation model” which is now used by 130,000 employees, CTO Hari Gopalkrishnan said. He also noted that 18,000 software developers are now using coding agents, driving 20% more efficiency.
   
   
2. AI foils fraudsters: The bank has more than 270 AI models in production; more than 50 are used to detect fraud. Holly O’Neill, the bank’s president of consumer, retail and preferred, said AI had cut the fraud loss rate in half since 2018.
   
   
3. Integration is the frontier: The future of the bank’s $13 billion in tech investments is “a much more all-encompassing framework where Erica and other platforms can work across the broader enterprise.” That includes “agent ecosystems,” Gopalkrishnan said.

TREND LINES

TALKIN BOUT MY AGENERATION

If you want to know how transformative someone thinks agentic AI will be for their organization, just ask them their job title.

The C-suite is sold: Goldman Sachs is “rethinking every step of what we do” around agents, says CIO Marco Argenti. Ally Financial’s AI head Sathish Muthukrishnan wants to “rehire and re-underwrite” every role for an “AI-first world with agents.” And “every process is powered by AI agents” at the JPMorganChase of the future, says Derek Waldron, the bank’s chief analytics officer.

The skeptics? They are in the engine room, and tend to be both younger and closer to the technology. One study this week shows people below VP-level are only half as convinced as their bosses that their firms were outpacing rivals on AI adoption. Junior bank employees in another are twice as likely to say agentic AI would be “not very valuable” to their organization.

The gap reflects a messy truth: We’re still a long way from execs’ agentic utopia. For an agent to handle even something simple means breaking a task down into its component parts, giving it access to the right systems to handle each element and putting guardrails in at every step. It means small errors snowball fast: “If you have a 10-agent system and each is 85% accurate, that’s 0.85 to the 10th power accuracy by the end,” says Scale AI CEO Jason Droege.

Engineers may be better off ignoring their boss’ visions of grandeur and focusing on the smaller tasks (more on that in Use Case Corner). “Build an AI system that does the minimum, but build it such that it improves over time,” says Manuela Veloso, JPMC’s head of AI research. Or as one engineer claiming to have shipped agents for banks wrote this week: “You have to earn the right to automate the complicated stuff.”

Bottom Line: If banks make and scale agents, they could effectively have “10,000 more people in your business at zero cost tomorrow,” as JPMC’s CDAO Teresa Heitsenrether puts it. The question is how to get the people already there to make that a reality.

USE CASE CORNER

FOLLOW THE MONEY

To find the specific tasks banks are successfully automating with agents, we dug into the Bank of Singapore’s new agentic KYC tool. Donald MacDonald, head of the group data office at OCBC (which owns the Bank of Singapore), gave us the details.

Use Case: Source of Wealth Assistant (SOWA)
Vendor: Meta, Alibaba, OpenBMB
Bank: Bank of Singapore

Why it’s interesting: The agentic system writes “source of wealth” memos, one of the highly-manual KYC tasks for onboarding new customers to the private bank. Relationship managers previously needed to sort through roughly 40 documents to complete the due diligence process; now roughly 10 agents handle different parts of the process end-to-end, and the humans just double check details and intervene when issues arise. “Basically, their job is just to submit the documents, proofread it, and then submit it to their boss,” said MacDonald.

How it works: An orchestrator agent, which MacDonald calls a “choreographer,” taps specialized agents to handle specific parts of the task. One agent, for example, uses visual language models (more on that in our Jargon section below) to extract financial information from documents; another compares the customer’s income or holdings to the bank’s data to see if it matches what they’d expect for their profession. Another writes the memo itself, documenting where information is pulled from so humans can easily check its work when finished.

How they did it: They built the first version of the tool in just four months and run it fully on-prem for security. It’s built so the models underpinning it can be easily replaced, and the bank has a 1,000-question automated benchmarking framework to test whether new releases would improve performance. “The visual models have come a long way in the last six months,” said MacDonald. “The team are constantly evaluating what are the best models out there and bringing those in.”

By the numbers: SOW memos used to take relationship managers 10 days; SOWA can produce them in one hour. All 500 RMs at the bank can use the tool and a combination of a top-down push and the understanding of how much time it saves is driving adoption, said MacDonald.

FROM THE EVIDENT AI INDEX

NOT GUILTY

Since the launch of ChatGPT, job openings are down 30% while the S&P is up 70%. The “scariest chart in the world” is said to be evidence that AI is killing the job market. 

In banking, that isn’t happening – for now, at least. According to Evident data, the banks doing the most with AI have been steadily growing their headcount since last year. It’s the laggards that are shrinking.

But but but: Banks whose business is doing well these days also happen to be the ones that are spending more on AI and hiring more overall. Those things may not be correlated. The leaders on AI will for sure be looking to use the technology to squeeze out efficiencies across the board. Yet the investment in AI is also clearly creating lots of new jobs. Look to studies about the internet boom: 2.6 jobs were created for every one lost.

STAT OF THE WEEK

The amount people estimate they save annually by using AI tools to manage their personal finances, per Lloyds’ 2025 Consumer Digital Index.

Why it matters: If ChatGPT can help people save that much, banks rolling out their own money management tools now have a number to beat. But the real test is whether AI can help people make money. That’s an uphill battle: Citadel’s Ken Griffin said last month that “for uncovering alpha, [Gen AI] just falls short.” And an experiment where six models used $10,000 to autonomously trade crypto saw GPT-5 losing more than 60% of it in 18 days.

WELCOME TO THE JARGON

DEEPSEEK’S AVID READER

Banks sit on mountains of “dark data” – scanned documents and slide decks saved in formats that AI models are bad at reading. Chinese AI darling DeepSeek has a new model that uses an old-school tech in a new way to fix it. It’s an improvement on…

DeepSeek’s model doesn’t just extract letters or words in a document, it interprets what those words are actually saying, which means it’s more accurate at inferring what something says when it can’t read it that easily. It set a new record on a challenging benchmark, with 88% accuracy. That means one of 10 words will still be wrong, but it’s an upgrade over everything else on the market – and in DeepSeek fashion, it’s cheaper.

Banks devote lots of resources to making their data machine-readable: Citi, for example, digitizes some 25 million pages of trade-related documents a year. Still, firms are only using about 20% of their data for AI, Standard Chartered’s head of Gen AI enablement David Hardoon told the Evident AI Symposium.

Banks need to unlock more of their data goldmine if they want to make their AI tools better. DeepSeek just showed the way, and banks shying away from Chinese lab’s models better hope they (or an AI lab somewhere else) can recreate it.

TALENT MATTERS

INDUSTRY CHANGERS

Wells Fargo hired Chitra Hota as its head of data and analytics and COO technology. She was previously chief data and analytics officer at Carlyle. The bank also brought Nonso Ogbonna on as a managing director in enterprise functions technology. He previously held tech infrastructure posts at Goldman Sachs and JPMorganChase.

Ash Kaduskar is now head of AI at First Citizens and is hiring for four AI roles, including a head of responsible AI. Kaduskar was previously a managing director of data and AI at EY.

Nicolaj Gudbergsen, who led Gen AI at Danske Bank until earlier this year, joined PwC as a partner in financial services. 

Goldman Sachs brought on Ritu Narula as a VP, AI. Narula will be “building the future of responsible AI,” she wrote on LinkedIn. Previously, she worked at the Ford Motor Company.

RBC is hiring a director of agentic AI and automation strategy into its technology infrastructure team.

NOTABLY QUOTABLE

“I think banks are adopting AI faster than we would expect and arguably outpacing fintechs in their AI adoption.” 

– Jenny Johnston, GTM and partnerships at OpenAI, on a podcast, Oct. 23

ON THE HORIZON

AI MOOD BOOSTER

In this new segment, we explore how a bank’s AI patent advances their AI strategy. This week: Truist (fifth-most patents in the Evident AI Index) wants to use AI to gauge if customers are happy and intervene if they’re not.

The patent, explained: Staying on the line for the umpteenth time to fill out a survey about an interaction is a pain. And the payoff for banks isn’t even that great. Instead of relying on customers to rate their own interactions, Truist designed a sentiment analysis system that takes what people actually say when they talk to the bank and scores the interactions based on how positive or negative it is. It gives the bank a running net promoter score, which lets them more accurately and specifically understand when and why things go awry.

How they can use it: For now, Truist can use the system to get a more accurate read on how customers feel about products, branches, agents, even AI tools without needing to ask. In the future, the bank could have changes in the sentiment scores trigger certain actions. For example, if someone were clearly getting frustrated with a chatbot, the system could flag it and either recommend an action or transfer them to a different workflow or a human.

CODA

JUDGMENT DAY

AI practitioners in banks spend lots of time evaluating models. Now models may start evaluating them back.

JPMorganChase and Citi recently said they’d allow managers to use AI tools to write up first drafts of employee performance reviews. It’s a huge time-saver on a task that few enjoy. But there’s more risk than meets the eye. 

New research shows models prefer resumes they wrote themselves compared to ones written by other LLMs. It’s not hard to extend that idea to the performance review process: If an employee uses one model to generate a self-evaluation, and their boss uses a different one to write their performance review, there’s a chance they could get dinged for that alone.

Leaders at Goldman Sachs have been calling out overreliance on AI tools as a growing issue lately. As dull as performance reviews can be, this might be one area where humans ought to always stay in the loop.

Mon 10 - 11 Nov 
FTT Fintech Festival, London

Mon 17 - Tues 18 Nov
Momentum AI Finance 2025, New York

Sun 30 Nov - Sun 7 Dec
NeurIPS, Mexico City & San Diego