Deployment of agentic AI systems designed for banking are still few and far between. But this approach, whereby multiple AI agents are entitled to carry out actions on users’ behalf rather than interacting with them through language models, is already in motion at Capital One.
The US bank’s Chat Concierge tool is one of the global industry’s first examples of putting a proprietary agentic AI system into production – although CommBank’s disputes process is another notable case (see box).
Chat Concierge is now up and running on car-dealership websites in the US. It allows customers on participating dealer websites to compare cars and look at financing options. It will also give them an estimated price for a trade-in vehicle. It will even schedule a test drive by plugging into the dealers’ customer relationship systems.
This sounds auto-finance specific, but the tool’s deployment this year is a landmark for Capital One’s AI strategy, according to chief scientist and head of enterprise AI Prem Natarajan.
Speaking to Euromoney, Natarajan says Chat Concierge’s significance goes far beyond auto finance: important though that is for the firm.
“This is our first generative, AI-powered, agentic customer-facing experience,” he says. “You can imagine all kinds of other such experiences being built on this use case. The whole flexibility and customisability of this framework is what's exciting to us. This is our first beachhead. Internally, too, we think this will scale to a lot of use cases.”
Key factors
The ability to adapt individual use cases like this for other areas of the business has become one of the key factors for leading banks in deciding which ones to pursue in terms of investment. That stands in contrast to traditional AI, which was cheaper to run and tended to be designed for specific uses.
What is Chat Concierge?
Chat Concierge is a tool built by Capital One and presented to visitors to car-dealership websites in the US. It allows users on participating websites to compare cars and explore financing options. It also gives estimated prices for trade-in vehicles and allows users to schedule a test drive, plugging into the dealers’ customer relationship systems.
The tool is based on Llama, Meta’s family of open-source large language models (LLMs). The bank customised the Llama model using its own data, which Capital One’s Prem Natarajan says was the key stage in boosting the accuracy of tools such as this.
The agentic method at Chat Concierge involves the tool seeking to understand the customer’s needs based on natural language – what is the desired price range, colour or manufacturer. The tool then formulates a plan for meeting those needs, before checking that the plan confirms to certain policies and checking that plan with the customer. The aim is for the conversation to be as close as possible to one a customer might have with a human: checking that it has heard and understood correctly, using familiar and non-technical terms.
As this is an early instance of gen AI going in front of customers, there is still a human behind the scenes, checking key outputs. Those checks might include whether the AI is suggesting an appointment time, not on Sunday at midnight, for example, and that the tone and language is appropriate, in longer responses.
But Natarajan makes clear that the other key factor is where they stand on the risk spectrum.
The Chat Concierge use case is even more significant for the bank, and the industry at large, because it is customer-facing. It is not only allowing AI agents to make travel bookings for the bank’s staff, for example, which is the sort of thing many banks are thinking of in agentic AI. Beyond low-risk staff uses, customer-facing use cases for gen and especially agentic AI have the potential to transform the way that customers interact with banks, and perhaps the economics of banking, potentially in an even more radical way than the rise of mobile banking 15 years ago.
The actions Chat Concierge is entitled to take, however, are not especially consequential. And that is deliberate. Booking appointments probably do not have the ability to change anyone’s life.
Decisions around credit would be at the opposite end of an agentic AI risk spectrum – and at the highest point of the gen AI risk slope, as Natarajan conceives it. The bank, which is traditionally best known as a credit card firm, is consequently not rushing to deploy AI agents for credit decisions.
“We’re still using traditional AI and machine-learning analytics for credit, with humans in the loop,” he says. “Using gen AI in credit is more of a multiyear journey. It’s not that it wouldn’t be great to get it today, but we really need to understand these techniques properly. All the other things we’re doing allows us to understand these algorithms much better and gives us confidence, as we slope through, and say that we can predict and understand the performance of the technology.”
When the bank does deploy gen and agentic AI in credit, on the other hand, it will not just be an experiment done for the purpose of being able to tell us that it has done it.
Scaling
Natarajan’s view, on the contrary, is that modern technology is all about scaling. That is why the bank’s use of gen AI assistants for customer agents is important for him. The tools have essentially been deployed to all its customer agents, around 20,000 people. This has meant a much deeper impact on customer and employee experience than if it had been kept to a small test group and left at that.
Getting to scale, and deepest impact, requires not just progress along the slope from low- to high-risk use cases, but also a careful and steady process of advancement in terms of the number of users for those use cases. “Our approach has been very, very, very careful,” Natarajan says. “In some sense, you could almost accuse us of being slow. We wanted to test and learn, test and learn, test and learn.”
That is the approach Capital One has taken not just by starting off with the relatively low-risk use cases such as agent servicing, but also in terms of the low number of employees who were first allowed access to the tool, and who will first be allowed access to any new features.
“Our first pilot was with 10 people, for a couple of months,” he says. “You could say that, at that glacial pace, you’re never going to do everything. The next pilot was with a couple of hundred people for a few weeks. We got more feature requests from these people, so we expanded the solution. Then the next step is a couple of thousand people. By then, you’ve really proven it to yourself that this is scaling and you’re able to observe the risk performance of it. You’re building trust with the users and with the stakeholders.”
AI in banking is a “stairway to heaven” in Natarajan’s telling.
“The first step is getting a baseline solution to everyone,” he says. “Now, when we make improvements, and add new features, they’re automatically going to be available to everyone. Our ability to deploy and climb up the stairway to heaven is that much faster, because we have invested in this baseline. You test it out, learn rapidly, iterate, deploy.”
Natarajan is familiar with this approach from his past in big tech, namely Amazon, where he worked on Alexa. But he says the process of testing and learning is crucial in banking because of the presence of financial risk, adding: “If the belief is that the power of transformation that AI is going to unleash is an enduring and transformative one, why do you need to rush and take risks up front? Let’s make sure we’re doing this right.”
CommBank’s rival first step to agentic nirvana
Aside from Capital One’s Chat Concierge, another example of agentic AI is at Commonwealth Bank of Australia (CBA, also known as CommBank). In this case, it involves managing payment disputes, an area in which banks have an unusually prominent role in Australia. Banks categorise the dispute and interact with card schemes to determine whether it was a fraud, misrepresentation or some other problem, before seeking a resolution.
Gavin Munroe, CommBank’s chief information officer, tells Euromoney that CommBank customers could previously log a dispute on the bank’s app via a predetermined series of drop-down menus. The app would also ask for things such as an image of the item before processing the dispute. It might ask for more information or further documentation.
Now, Munroe’s team has rebuilt that process so that customers can inform the bank about the problem in a less defined way using a large language model (LLM), with an AI agent then assessing the issue in real time. That includes making judgements based on images that the customer sends and, where necessary, asking for more information – during the same interaction.
More complicated disputes such as fraud will still take longer and involve more human orchestration, Munroe admits. But the hope at the bank is that the new agentic system will allow the bank to process simpler claims in minutes, when they might have taken a day or two before – allowing the customer’s account to be credited almost immediately. After initially internal testing last year, it is rolling out the tool to customers for simple disputes.
