In recent weeks, several new large language model-based AI agents have been introduced. OpenAI launched Operator, an AI agent that can order groceries and book flights. Google released Gemini 2.0, which can "understand complex scenarios, plan multiple steps ahead and take actions on behalf of users," according to the company. Oracle announced AI agent services that automate business processes. Auquan deployed AI agents at private credit firms that research and draft deal memos.
This new generation of
"Unlike today's generative AI models, which respond to specific human prompts, agentic AI can independently perceive, reason, act and learn, without constant human guidance," states a World Economic Forum
"AI agents are differentiated from AI assistants and chatbots by their ability to autonomously plan and act to meet a user-provided goal," wrote Gartner analysts Gary Olliffe and Steve Deng in a report published in January. Gartner predicts that by 2028, 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024, and at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from zero percent in 2024.
The potential use cases in banking are almost endless – AI agents could be used in personalization, customer service, operations, risk assessment, underwriting, financial forecasting, Know Your Customer and anti-money laundering rule compliance, new customer onboarding and fraud prevention. They could save time and human effort and standardize sometimes haphazard processes.
There are also risks. Agentic AI can make the same kinds of mistakes human agents do, but more quickly and at scale. Glazed-eyed human reviewers looking at high volumes of AI agent activity might fail to notice errors. Criminals could tell AI agents to hack into a bank and steal information, and the AI agents could just go and do it, running many bots at the same time. All the risks of generative AI remain, including the potential for hallucinations and bias.
Banks like Capital One and Goldman Sachs, among other companies, see the potential efficiencies they could gain from agentic AI and they're laying the groundwork for it, including setting up risk and compliance controls.
Capital One offers AI agents to car buyers
Capital One recently launched an AI agent called Chat Concierge that helps consumers through the process of buying a car. Car buyers can engage with Chat Concierge on participating dealers' websites. It can compare vehicles to help car buyers decide on the best choice for them, schedule appointments with salespeople and schedule test drives.
Chat Concierge is a proprietary, multi-agentic conversational AI workflow that Capital One developed in-house, Prem Natarajan, chief scientist and head of enterprise AI at Capital One, told American Banker.
"It consists of multiple agents that work together to mimic human reasoning and not simply provide information to the car buyer, but take action based on consumers' needs and requests," he said. "All together, these agents work to understand natural language prompts, come up with an action plan to execute on those prompts, validate that plan to mitigate against hallucination and errors, and explain the information and the plan to the car buyer."
Capital One uses Meta's open source Llama model as a base that the bank customized with proprietary data to meet its performance, risk and governance thresholds.
The bank conducted extensive testing and implemented human-centered guardrails before introducing Chat Concierge, as it does with all AI systems, Natarajan said.
"Chat Concierge leverages both humans and technology to guard against bias and hallucination," he said. "Human-in-the-loop processes are our first and most critical line of defense. We also use technology tools like automated guardrails to ensure tone and language of responses conform to our approved guidelines and business rules."
The bank is always exploring how to enhance the customer experience with technology and AI, Natarajan said.
"While we believe the agentic framework we developed here has extensibility to other business processes and customer experiences, we are continuing to balance that potential with a thoughtful and responsible approach," he said.
At Goldman Sachs, agentic AI is coming
About two years ago, Goldman Sachs also built a platform, called GS AI, that hosts AI models for employees to use, and like Capital One it included risk and compliance controls to ensure safe and responsible AI, according to Chief Information Officer Marco Argenti. Some controls, for instance, implement data protection standards – they might search for personally identifiable information and scrub it so it doesn't get fed to an AI model, for instance. Other controls govern entitlements and data access control.
Today the platform is used to enable employees to use generative AI models. In the future, it will accommodate agentic AI, he said.
"One metaphor could be that we discovered this source of new energy, and we need to build a power plant around it to transform it into energy that can be used effectively and safely," Argenti told American Banker.
GS AI supports both large, proprietary models like OpenAI's GPT and Google's Gemini as well as open source models like Meta's Llama. It has a layer that developers use to create applications quickly and safely. To date, they have developed 35 AI applications that are in different stages of production; more are in the pipeline, Argenti said.
"They range from developer productivity to knowledge worker productivity to banker productivity to asset management and sentiment analysis," Argenti said.
One application is a virtual assistant, called the GS AI Assistant, that's been provided to 10,000 bankers. It's "essentially your daily companion that sits next to you, and you can ask questions," Argenti said.
But so far, the most substantive results have come from a generative AI copilot Goldman gave its 12,000 developers six months ago, Argenti said.
"Any hour of productivity that we get from a developer multiplies itself by 12,000 – it's easy to do the math," Argenti said. "There we've seen the first measurable evidence that AI can actually have a substantial impact. We've been measuring consistently efficiency improvements that range from 10% to 20% measured in terms of time."
But agentic AI has the potential to provide higher efficiency gains, above 100%, Argenti said.
"Imagine if I have a task that would normally involve 10 people, and I ask 10 agents to do that, then you just multiply that by 10," he said. "For certain tasks, for example, repeated actions, I think that is particularly valuable."
Argenti gave a hypothetical example: a company changes its logo and has to quickly swap in that new logo on 1,000 different pages across 10 websites.
"You could have people do that, or you could have scripts do that," he said. But people would take a long time and automated scripts would not be able to handle special cases, where the font has to be resized, for instance.
"That's where things are going to become very interesting from an efficiency standpoint, because it will almost be like having virtual employees," Argenti said.
Today, the GS AI platform supports a form of agentic AI called "chain of thought," a technique that helps AI models solve problems by breaking them down into a series of logical steps. It uses LangChain as a workflow engine. It also has tools that can do things like retrieve documents and determine which documents are more relevant than others. Workflow tools can do things like generate a draft, get it approved and translate it to many languages.
A copilot in the works will help investment bankers put together M&A deals, using knowledge graphs to encode information about the deals, using a retrieval augmented generation tool to answer questions about the market and retrieve relevant past deals, and then using that information to create a summary for the banker, Argenti said.
"The bottom line is, I think this is a very natural evolution in AI, where you're starting to go from simple, 'give me the first answer that you have' to 'let me do some research in the background and then come back to you with something that you can review and give me some feedback,' and it becomes very interactive," Argenti said. "We think that this kind of ability is where AI is going."
But agentic AI will require controls and human oversight, Argenti said.
"The most dangerous thing would be if you let the agent act without human supervision," he said. "That's one of the fundamental tenets that we have: you're always helping a human who then makes the decision. Or you're informing a process that goes through several approval levels. So it kind of fits within a process. It never acts by itself. We're very far from a time where they're going to be replacing human judgment."
AI agents won't be allowed to do independent work at Goldman until their error rates are close to zero, Argenti said.
"The bar is pretty high," he said. "Even the best agents today are not entirely error proof. So we're going to continue to observe. It's a journey, and so we will continue to walk a safe path, and hopefully we're going to get benefits."
Creating dealbooks in private credit
"If you look across finance, there is a lot of work that requires people to look through very large amounts of messy, unstructured, noisy data to find useful information, to find interesting data points, which they then put into a report or a document or a presentation or a template, and that becomes a starting point of taking interesting actions, making interesting decisions," said Chandini Jain, CEO of Auquan. "In the absence of software or technology, you end up using people to just basically fight with data."
An AI agent connected to a company's data can do the work that a person would have done previously to produce a presentation or document based on that data. One of the most popular use cases for Auquan's agent, according to Jain, is in private credit – lending by non-bank financial institutions such as private equity firms and alternative asset managers, often to small and mid-sized businesses.
"Private credit has blown up as an asset class over the last few years," she noted. "There are lots of tailwinds in the market, with banks pulling out. When a borrower comes to you and you're a private credit firm, before you even decide to look at the deal, the first thing that you do is you look through all of the company's disclosed data, whatever documentation the bank that originated this deal has sent. You also conduct public data searches, and you prepare a five to 10 page memo that makes sure that this deal aligns a lot of the gating issues before you even start investigating the deal further."
In the absence of an AI agent, an analyst spends a lot of time looking through documents, often PDFs or Powerpoint files, searching for relevant information.
"There is no naming consistency, so you don't know what a PDF contains, so some analysts will open 20 to 25 files in the desktop and run Control Fs on 20 different terms across each of the files to find where information exists, because not every company names things similarly," Jain explained. "And then you copy and paste it all back into a document and it takes two to three days to prepare this report."
Auquan's deal screening agent that can find relevant data elements including a company overview, company management, revenue growth, earnings analyses, market analyses and sector analysis.
"You just drag all of the company documents into the platform, you enter the name of the company, you press the 'run' button, you go get a cup of coffee, you come back, and your report is ready for you," Jain said. "
Auquan recently launched an agent that goes through this process for sustainability. Soon Auquan plans to launch an anti-money laundering and financial crime detection compliance agent.
Jain acknowledges her company cannot guarantee 100% accuracy. "No one can with AI today," Jain said. "But what we can do is get to high enough levels of accuracy, like 90% plus, and then make it easy to verify the rest of the information."
Catching financial crime
Oracle recently rolled out an AI agent for anti-money laundering and financial crime detection. It also has AI agents for enterprise risk and finance.
The AML AI agent can collect evidence of suspicious activity and make recommendations about when to file a suspicious activity report.
"It's all about, how do we help ease the time of investigation for financial crime?" Jason Somrak, head of financial crime products for Oracle Financial Services, told American Banker. He previously worked in financial crime detection roles at KeyBank and PNC Bank.
In the financial crime space, most banks have anti-money laundering, know your customer and sanctions models that generate alerts for human review.
For example, the risk engine might identify that a customer has sent a high-risk wire to a high-risk country. An investigator needs to consider which country it was sent to, the source of the funds and where the money went.
"They're basically trying to figure out who, what, when, where and why," Somrak said. "In doing that, they have to go and look at all their internal systems. They have to follow their procedure. They type up a long narrative explaining what triggered the alert, what the evidence tells them about it, and then ultimately, they make a decision on it."
AI agents can automate much of that work, Somrak said. They can collect the evidence, recommend an investigative decision and write a narrative of what they've done and why they feel it's potentially normal or abnormal based on bank policy. A human investigator reviews what the AI agent did and can ask the AI agent to find more details about a transaction, then ultimately confirm or deny the AI agent's decision.
"How we think this is really going to change the game is the largest banks have sometimes 4,000 to 6,000 investigators doing this work," Somrak said. "We believe this agentic workflow can automate at least 80% of that work. It might take five to 10 years for banks to get there, but we think we have the ability to shift the industry from trying to basically throw a bunch of bodies at this work to try to make sense of which of these events are good, and you the humans into labelers for a decision engine and quality control officers."
