How do you balance risk management and safety with innovation in the agent system – and how do you grip the basic reservations around the data and model selection? In that VB Transform AI Foundations in Sessions, Milind Napdea, SVP, Technology, Capital One, offered the excellent methods and lessons learned from real -world experiences and requests for the deployment and measurement of an agent workflower.
In order to stay at the forefront of emerging technologies, Permaid Capital One recently launched a production grade, the latest multi -agent AI system to enhance the car buying experience. In this system, many AI agents work together to not only provide information to the car buyer, but also to take specific measures based on user preferences and requirements. For example, an agent interacts with the customer. The second develops a business rules and a tool -based action plan that is allowed to be used. A third agent evaluates the accuracy of the first two, and the fourth agent describes and verifies the action plan with the user. More than 100 million users are built for agent system scale and complexity, using a wide range of other potential capital One use case applications.
“When we think about making customer experience, making the customer happy, we think, what are the ways that can be?” Napde said. “Whether you are opening an account or you want to know your balance or you are trying to make a reservation to test a car, there are many things that users want to do. In his heart, very straightforward, how do you understand what the customer wants? How do you understand all the procedures of completion, all the way you can do all the policies? , How will you bring all policies, all policies, all policies, all policies, all policies, all policies, all policies, like all policies, how you will regulate and otherwise.
He said, Agent AI was clearly the next step, for the use of internal and customer use.
To design the agent workflow
When designing any workflow, financial institutions have particularly strict needs that support consumers’ travel. And Capital One applications include several complex processes as customers raise issues and questions that benefit the conversation tools. These two factors make the design process especially complicated, which requires a comprehensive theory about the entire journey – which includes how both consumers and human agents respond, react and argue at every step.
“When we saw how humans argue, we were impressed by some of the significant facts.” “We saw that if we designed it using numerous logical agents, we would be able to copy human reasoning very well. But then you ask yourself, what do you do to different agents? Why do you have four? Why not? Why not? 20 Why not?”
He studied consumer experiences in historical data: where that conversation is right, where they go wrong, how much time and other prominent facts should they take. They learned that it often takes a number of conversations with an agent to understand what the user wants, and any agent workflow needs to be planned for it, but also to be fully founded in an organization’s system, available tools, APIs, and organizational policy guards.
Naphada said, “The important development for us was realizing that it would have to be dynamic and repeated.” “If you see how many people are using LLM, they are slapping the LLM as the front of the same mechanism that existed. They are using LLM just for the intention of the intention. But we realized from the beginning that it was not an extension.”
Take indicators from existing workflows
On the basis of their intuition due to human agents when responding to consumers, Capital One researchers developed a framework in which a team of expert AI agents, each, with different skills, solves a problem.
In addition, Capital One added a strong risk framework in the development of the agent system. As a regular organization, Nafid noted that in addition to its internal risk reduction protocol and framework range, “within Capital One, to manage the risk, other organizations that observe you, see you, ask you, audit you.” “We thought it was a good idea for us, having an AI agent whose whole job was to assess what the first two agents do based on Capital One’s policies and rules.”
The diameter determines whether the first agents were successful, and if not, it rejects the project and requests the planning agent to correct its results on the basis of its decision where the problem was. This happens in a remedy until the proper plan is arrived. It has also proved to be a great honor for the company’s agent AI approach.
“The diagnosis is an agent … where we bring a global model. At the same place, we imitate that if a series is really started, we need hardships, because we are a regular enterprise – I think it is in fact a great and strong place.
Agent AI’s technical challenges
The agent system needs to work with the entire organization’s entire system, all of which are accompanied by different permits. It was also difficult to request tools and APIs within numerous contexts while maintaining high accuracy – from the consumer’s intention to develop a reliable plan and to implement it.
“We have experiences, tests, examinations, diagnosis, human loops, all the correct repetitions of all the right guards that we really need to be before we come into the market,” Napa said. “But one of the biggest challenges was that we had no ideology. We couldn’t go and say, oh, someone else did it like this. How did it work? It was that element of novelty. We were doing this for the first time.”
Model selection and partnership with NVIDIA
In terms of models, Capital One is deeply detecting educational and industry research, being presented at conferences and the state of art is to keep the state of art. In terms of current use, they used open -weight models instead of closing, as it allowed them to specialize. This is very important to them, Napad claimed, as the AI ​​strategy relies on competitive advantage of proprietary data.
In the technology stack, they use a combination of tools, including internal technology, open source tool chains, and NVIDIA in conference stack. Working together with NVIDIA has helped Capital One achieve his or her own performance, and NVIDIA’s library has been supported on industry opportunities, and features for Trainin server and their tensoor LLM have been preferred.
Agentk AI: Looking forward
Capital One continues to deploy, scale and improve AI agents in its business. His first multi -agent was a workflow consequence, deployed through the company’s auto business. This car was designed to help both auto dealers and users from the purchase process. And with a rich data from consumers, dealers are identifying serious leads, which significantly improved their consumer engagement measurements – in some cases up to 55 %.
“They have been able to create a very better edge to work for them working for them through this natural, easy, 24/7 agent,” said Naphada. “We want to bring this ability (more) to the engagement you receive from our customers. But we want to do it well. This is a journey.”