How Capital One built production multi-agent AI workflows to power enterprise use cases

How to balance risk management and safety with innovation in agents – and how do you wrestle with basic considerations about data and selection of models? In this VB conversion session, Milind Naphade, SVP, Technology, provided the foundations of artificial intelligence in Capital One, best practices and lessons learned from experiments and applications in the real world to spread and expand the workforce of an agent.

Capital One, committed to staying at the forefront of emerging technologies, recently launched a multi -factor production system, to enhance the experience of car purchase. In this system, multiple Amnesty International agents work together not only to provide information to auto buyers, but to take specific measures based on customer preferences and needs. For example, one agent communicates with the customer. Another creates an action plan based on the work rules and tools that are allowed to use. The third agent evaluates the accuracy of the first two, explains the fourth agent and is achieved from the validity of the action plan with the user. With more than 100 million customers who use a wide range of other possible status applications, the agent system for the scale and complexity is designed.

“When we think about improving the customer’s experience, and we admire the customer, we think, what are the ways that can happen?” Navdy said. “Whether you open an account or want to know your balance or try to conduct a reservation for a vehicle test, there is a group of things that customers want to do. At the core of this, simply, how do you understand what the customer wants? How do you understand the mechanisms of loyalty, otherwise?

He said it is clear that Agency AI is the next step, in cases of internal use and confrontation for customers.

Working process design

Financial institutions have special strict requirements when designing any workflow that supports customer trips. Capital One applications include a number of complex operations as customers raise problems and information that benefit from conversation tools. Make these two factors the design process in particular, which requires a comprehensive vision of the entire trip – including how customers and human agents respond, respond and the reason for each step.

“When we looked at how people think, we had some prominent facts,” Navdy said. “We have seen that if we designed it using multiple logical factors, we will be able to imitate human thinking well. But then you ask yourself, what do different agents do? Why do you have four? Why not three? Why not 20?”

They have studied customer experiences in historical data: these conversations are properly going, as they make mistakes, the duration they should take and other prominent facts. They have learned that it often requires multiple turns of conversation with an agent to understand what the customer wants, and any agent’s workflow that needs to be planned, but also on the basis of the institution’s systems, available tools, application programming interface, and regulatory policy handles.

“The main penetration for us was aware that this should be dynamic and repeated,” said Naphade. “If you look at how many people use LLMS, they slap LLMS as a front end for the same mechanism that existed. They only use LLMS to classify intention. But we realized from the beginning that this was not developmental.”

Take the sermon from the current workflow tasks

Based on their intuition on how human agents caused the clients ’response, researchers at Capital One developed a framework in which a team of artificial intelligence agents experts, each of them has different experience, and solved a problem.

In addition, Capital One included strong risks in the development of the agents. As an organized institution, Naphade noticed that in addition to a group of internal dilution protocols and frameworks, “in the capital, to manage risk, other independent entities monitor you, evaluate you, ask you, and check you,” said Naphade. “We thought this is a good idea for us, that we have the agent of Amnesty International as his entire job to assess what the first agents do based on the policies and rules of capital.”

The evaluation determines whether the former agents are successful, and if not, he rejects the plan and asks the planning agent to correct his results based on his ruling on the place where the problem was. This occurs in a repetitive process until the appropriate plan is reached. It has also been proven that it is a huge blessing of the AI’s approach to the agent.

“The evaluation agent is … where we bring a global model. This is where we mimic what is happening if a series of procedures are already implemented. This type of rigor, which we need because we are an organized institution – I think this is already putting us on a sustainable and sustainable path.

Technical challenges of the artificial intelligence agent

The agents systems need to work with loyalty systems throughout the institution, all with a variety of permissions. Calling tools and a application programming interface in a variety of contexts while maintaining high accuracy was also difficult – from angry user intention to create and implement a reliable plan.

“We have multiple repetitions of experiments, testing, evaluation, and human in the episode and all the correct handrails that must occur before we can already reach the market with such something.” “But one of the biggest challenges we had any precedent. We could not go and say, oh, someone else did this in this way. How did that succeed? There was an element of the grandmother. We were doing it for the first time.”

Choose the form and partnership with NVIDIA

Regarding models, Capital One tracks academic and industrial research strongly, progresses at conferences and remains aware of what is the latest. In the case of current use, use open weight models, instead of closing, as this allowed them to be a great customization. This is very important for them, as Naphade asserts, because the competitive advantage in the strategy of artificial intelligence depends on royal data.

In the same technology staple, they use a set of tools, including internal technology, open source tools, and NVIDIA’s conclusion. Work closely with NVIDIA Capital One helped get the performance they need, cooperating in industry opportunities in the NVIDIA Library, and giving priority to TRION Server and Tensort LLM features.

Artificial Intelligence Customer: Look forward

Capital One continues to publish and expand its work agents. The first multi -factor workflow was the concrete chat, which was published through the company’s automatic business. It is designed to support both car dealers and customers in the car purchase process. And with the rich customer data, customers determine dangerous threads, which greatly improving their customer participation measures – up to 55 % in some cases.

“They are able to generate much better strands through this natural and easier agent, around the clock throughout the week.” “We would like to make this ability [more] One of the links to our customers is. But we want to do it well management. It is a trip.