Why agentic is so hard

We’re halfway through 2025 – the year Sam Altman promised AI agents would join the workforce and “materially change” companies. So far, banks are overpromising and underdelivering on agentic AI.

Other sectors don’t seem to have that problem. This month, Roche subsidiary Genentech built an agentic medical research tool with Stanford researchers that develops and tests scientific hypotheses all on its own. Siemens has an agentic system that autonomously begins repairs on factory machinery. And Walmart is experimenting with agentic systems that proactively manage inventory or prioritize tasks for in-store workers.

Despite hyping up the future of agentic AI, banks aren’t embracing the autonomy part. Even when places like BNY and Capital One (#14 and #2, respectively, in our Banking Index) launch agentic tools, they keep humans in the loop at every turn.

There are several hurdles to going fully agentic that have slowed adoption. One exec told us their risk team is already slammed preventing humans from inadvertently letting AI get too close to sensitive data or clients’ money. When you suggest that agents make decisions about clients' finances themselves – and add in worries that these systems are vulnerable to hacking by crooks – you can understand some of the hesitation.

Another exec told us the struggle with agentic was understanding why agents act the way they do: The more complex tasks agents take on, the harder it becomes to monitor them and find where they go wrong – which won’t sit well if regulators come calling. Still another said that integrating agents with the bank’s APIs and mainframes was the hardest part of ramping up to agentic. But without getting legacy systems and agents to cooperate, giving them more autonomy would just be letting them loose with half a brain.

That’s not to say banks have lost faith – or dropped the hyping. “Moving beyond a handful of agents into really implementing agentic in multiple ways,” is still JPMorganChase CIO Lori Beer’s top priority this year. But, “there’s a lot of work to really make that happen.”

For now though, the financial industry seems to agree more with Altman’s OpenAI cofounder Andrej Karpathy, who last week, in defiance of Altman, said, “This is the decade of agents, and this will take quite some time.”

In boxing, it was Ali vs. Frazier. In politics, Lincoln vs. Douglas. Now, in AI, Claude vs. Devin.

Well at least in the fight over agentic.

Anthropic (which makes Claude) and Cognition AI (which makes coding agent Devin) are taking dueling approaches to building agentic systems. The crux of the fight: What’s better, one agent that does everything in order or teams of specialized agents that tackle tasks all at once?

Anthropic makes the case for multi-agent systems. Here’s how that works: A lead agent breaks down a problem, spins up specialist agents that can handle different tasks and runs them all in parallel. Take writing a financial report: One agent might crunch market data, another might scan headlines and a third might check citations – all at the same time. Breaking a complex task into the sum of its parts and doing them simultaneously cuts down the time (and cost) it takes to complete it, the firm says.

Not so fast, argues Cognition AI. The hub-and-spoke approach means each agent only sees its individual slice of the overall task. Without “understanding” the objective, these systems can’t adapt on the fly and are more likely to fail. Oxford professor Michael Wooldridge, who wrote the textbook on multi-agent systems, actually agrees with the single agent argument: Today’s LLM agents aren’t trained to work together, he argues, and trying to get them to will make hallucinations worse.

Bottom line: Banks are, for now, taking the multi-agent approach – following Salesforce, Microsoft and OpenAI’s lead. There are good reasons for this: They can test agents on the cheap with straightforward tasks and use off-the-shelf LLMs to power them. But we don’t actually know if multi-agent is the right approach long-term, and businesses will probably have to be quick on their feet and adjust as needed.

While banks keep their agentic systems closely supervised by flesh-and-boned staff, these three use cases highlight ways they’re starting to loosen the restraints.

#1 OMNISCIENT ADVISOR

Use Case: Ask David
Vendor: LangChain
Bank: JPMorganChase

Why it’s interesting: Instead of punting client questions to the investment research team, financial advisors can “Ask David” to help them build strategies in real time using decades of price and performance data along with analyst notes, internal emails and presentations.

How it works: The tool has a “supervisor agent” that can plan the best way to respond to a question and in turn call on three specialist agents under its command. One agent turns questions written conversationally into SQL for structured databases. One retrieves information from firm documents. The third taps into analytics systems. If at any point the system has low confidence in part of its answer, it brings in a human expert.

Impact: Questions previously passed to another team now get answered in seconds, allowing advisors to improve customer service and freeing investment research teams up for other tasks.

#2: MONEY MENTOR

Use Case: Financial Coach
Vendor: n/a
Bank: BBVA

Why it’s interesting: Most banking apps just show balances – BBVA’s tells customers what to do next. The tool tracks spending and debt levels. It calls out invisible spending (like automatic subscriptions). It creates a personal financial plan and digital piggy banks, where it sets money aside on its own.

How it works: The AI engine uses information about a customer’s income, savings, loans and spending habits to determine which of the bank’s products suit the customer best. As customers use the tool, it learns and adjusts goals as necessary.

Impact: Financial Coach is in its infancy, but the bank says customers using its other AI financial health tools were 40% less likely to leave the bank, 55% more likely to buy other products the bank offers and had NPS scores – a way of measuring satisfaction – of up to 10 points higher than those who didn’t use those tools.

#3 QUESTION & ANZWER

Use Case: Amie
Vendor: AWS
Bank: ANZ

Why it’s interesting: ANZ’s chatbot doesn’t just answer bankers’ questions about clients, deals or market moves on command. It proactively reaches out to bankers to flag trends in that data it notices on its own.

How they did it: ANZ first began prototyping Amie with AWS in 2023. It used AWS SageMaker to develop and train the models that power it and AWS Bedrock to build the actual application. The bank says the tool has separate agents handling different parts of a question – one that translates text to SQL code, one that retrieves news and policy information and one that interacts with external APIs.

Impact: During the tool’s pilot, relationship managers went from minutes to seconds to generate answers, keeping pace with live client calls.

"AI is surprising. I think that that is the single most consistent theme, is that the thing we were picturing, we got something different, but we got something better, more magical."

- Greg Brockman, OpenAI co-founder and president, to John Collison, June 17

There’s one question that has come up often about the Evident AI Insurance Index since it was unveiled last week: How do banks and insurers compare on AI?

As Franck Pivert pointed out on LinkedIn, banks seem to be pulling ahead. The difference in the average scores of the top three banks versus the top three insurers shows a sizable gap. Since the Insurance Index came out six months after the Banking Index and things are changing so quickly with AI, he noted, “the real gap might be wider.”

Indeed. We've seen some fundamental shifts in banking even this year. They are quickly remaking their workforces to allow them to scale AI use cases more effectively across the business (see: “AI efficiency revolution comes to banking,” The Brief, April 17). Insurers are getting there, said Jodie Wallis, global chief AI officer at Manulife, during Tuesday’s Evident roundtable: “I feel like we’re just on the cusp of moving out of [the use case phase] and starting to talk a lot more about workforce transformation.”

The playbooks among breakaway leaders – JPMC and Capital One in banking and AXA and Allianz in insurance – are also similar. Top performers, as Chris Jefferson noted on Linkedin, invest early in the infrastructure and talent behind the scenes to pave the way for AI at scale. Andreas Schertzinger, AXA’s chief data, AI and innovation officer agrees. “We started around 10 years ago in systematically building out technology, data and operations foundations,” he told us during the roundtable. “You need all those three components…to transform and improve your business”

Her name was Lola. She was an AI engineer.

With apologies to Barry Manilow, our point here is: Check out Brazilian banks. They’re AI hot. Itaú Unibanco and Banco Bradesco, two of Brazil’s biggest firms, each employ more than 400 people dedicated to developing AI tools – about as many as BNY or CommBank and above the median at the 50 banks tracked by the Evident AI Index (which doesn't currently include Latin American banks).

They’re putting their talent to work, rolling use cases out that meet customers where they are, namely on WhatsApp. Banco Bradesco’s BIA tool and Itaú Intelligence are virtual assistants that let customers text or phone in questions related to their accounts or the products they offer. BIA also lets customers process real estate loans through WhatsApp, allowing customers to upload documents or check on their loan status from the app.

Zooming out: Latin America is hotter than you think, even outside Brazil, Evident’s Paraguay-based Sam Meeson points out. The region is historically underbanked – which led to a receptiveness to fintech when internet access exploded last decade. With that has come an openness to AI.

Brazil’s Nubank now has more than 100 million customers across the region and uses AI tools to handle customer service and speed up transactions. Incumbents like Colombia’s Bancolombia and Mexico’s Banorte have followed suit to keep up. “AI talent remains scarce across Latin America,” one finance executive said off-stage at Paraguay Tech Week, an industry conference held last week in Asunción. But the willingness of consumers in the region to adopt new tech means it could be a huge ground for AI experimentation as banks look to grab the regional AI lead.

Goldman’s taking Gen AI firm-wide. The 45,000-plus employees at the bank will now all have access to GS AI Assistant, up from 10,000, per a memo circulated by CIO Marco Argenti. As we told you earlier this year (see: “What DeepSeek really means for business,” The Brief, Feb. 6), the tool uses models from OpenAI, Meta and Google and toggles between them depending on the type of task users ask it to complete – whether that’s coding, document summarization or translation.

Keep your friends close and your models closer. In a new report, Anthropic researchers warn that the company’s flagship AI tool – along with those from OpenAI, DeepSeek, Meta, Google and xAI – all resorted to blackmail to get around irreconcilable conflicts with human counterparts. Anthropic calls it “agentic misalignment,” and, so far, it only shows up when they purposefully try to get the systems to take rash action. But it’s a reminder to anyone building AI systems that guardrails aren’t always foolproof.

Amazon, Google, Microsoft and Meta are joining forces to convince the Senate to pass Trump’s contentious 10-year moratorium on state-level AI regulation. The Senate parliamentarian, an unelected official who is kind of referee on rules, this week allowed the measure to stay in Trump’s spending bill, but an uphill battle remains: Prominent Republican Josh Hawley has vowed to join with Democrats to fight it.

HSBC is in talks with London-based software startup causaLens to begin deploying agents to automate back office tasks in its corporate and institutional banking team. The partnership comes as the bank aims to trim staff costs by 8% by the end of 2026.

NAB hired Lloyds’ Pete Steel to be its first digital, data and AI chief beginning next January. The Australian bank is following JPMC’s lead on having the AI head sit at the management table: Steel will report directly to CEO Andrew Irvine.

Danske Bank appointed Kasper Tjørntved Davidsen chief AI officer and head of GenAI. He joins from Topdanmark, a Danish insurance and pension company, where he served as CIO since the beginning of 2023.

Santander’s AI bench is getting deeper. The bank tapped José Manuel de la Chica to lead its AI lab, appointed Lucas Arangüena as the commercial and investment banking unit’s chief data and AI officer and brought in Silvia García Ledesma to be the firm’s global head of AI special projects and chief of staff of AI.

Citi’s AI shakeup creates a three-headed team. The giant U.S. bank appointed Gonzalo Luchetti, the head of U.S. personal banking, Tim Ryan, head of technology and business enablement and COO Anand Selva as executive co-sponsors of the bank’s AI initiatives. The trio is tasked with deploying generative and agentic AI tools to more employees.

Tues 8 - Weds 9 July
RAISE Summit, Paris

Weds 9 - Thurs 10 July
CIO Leadership Forum, Tokyo

Tues 15 - Weds 16 July
Momentum AI, San Jose