Satya Prakash Mohapatra

Good afternoon, everybody. Can you hear me properly? Good.

I know it is after lunch, and it is the last breakout session, I guess, so people must be excited for the networking sessions and all. I will quickly wrap around this session, and then we can all go and enjoy the networking, including myself.

A quick introduction: myself, Satya Prakash Mohapatra. I am working as a Senior Cloud Engineering Manager with Discover Financial Services. I think most of you must have heard about Discover, but just a disclaimer: recently it got acquired by Capital One. So whatever I am telling you about Discover, consider it as an integrated entity, which is Capital One / Discover.

I think I do not have to introduce Discover, but just to have a look around: Discover is like a complete digital banking company. It is everything at your fingertips. You can do everything on your mobile phone or on the laptop. You do not have to visit any physical bank or anything. You can get money also; we have the Pulse network ATM, so you can get money from that ATM. It is a completely digitalized bank; everything you can do through the digital apps.

The best part I can say about Discover is people and its culture. If you go and ask anybody in Discover, or if you have a Discover card and interact with customer service, the best part of working at Discover is their people and their culture. The people are so collaborative, they are so friendly, and the culture really helps us grow technically and in all other aspects. That company has that culture. That is why I see people who have been there for 30 or 35 years still working with Discover. Now, of course, it is Capital One, but yes, people are there for 35 years. When we ask why they are here for 35 years when they could go to another organization, the answer is the amazing culture the company has.

Its inception was 1986. I think Discover was the first financial institution that created the rewards-based credit card, where you spend on the credit card and get the reward after that. All the organizations copied that. That was the first company that introduced the reward-based credit card. Along with that, they have a reward-based debit card also. If you have a Discover debit card, you can get the reward just by swiping the card.

Along with that, the unique thing is its global payment network. Like Visa and MasterCard, Discover also has its own payment network. It is global because, while you might say Discover is only accepted in the USA, we have a partner called Diners Club International. By using that, our card can be utilized across the globe. That is why we have a worldwide presence.

These are the culture points I mentioned: small autonomous teams. Everybody has access to their leaders to discuss initiatives or innovations and get them approved. It is pretty seamless. We do not have to go through a lot of steps, procedures, or processes to get it approved. The focus is technical, because we are a technical company. Though it is a financial organization, it is a technical financial organization, and we focus on technical learning and innovations. Those are the primary factors. We also have academies for internal development, internal assessment, technical development, and all those things.

Now comes automation. Everybody is talking about automation, AI, and all those things. We are a financial organization, so we are a little bit behind. We are not a first runner like other organizations. They are already using different kinds of agentic AI tools. But we are a little bit behind because we have to think about governance, regulations, laws, and a lot of other things. That is why we have to be a little bit protective, because we have a lot of customer information and data. If we share that to AI or anything, people might feel reluctant. So we are a little slower. But yes, we are there; we are working on that. We also have some agent tools that we are using in our organization.

A disclaimer again: when we are using AI, it does not reduce any workforce. It just relaxes our workforce from day-to-day activities, day-to-day boring activities, and lets them get to productive work, innovative work, and creative work. That is why we use AI to automate things so that people can be freed from those works. That is why we are using GenAI.

The thing I mentioned is a Bill Gates quote: if we use automation in efficient operations or systems, it definitely magnifies the efficiency. But it has a negative path also. If we use automation in an inefficient system or inefficient operations, it will magnify the inefficiency also. So we have to take care of the efficiency of automation or whatever we are using in our operations or systems carefully.

At Discover, we follow some rules to identify whether a particular process or operation is automation-eligible. The factors are: if a system or process is repetitive; if it is logic-based; if it involves a significant amount of manual work; if it has standardized inputs and outputs; and if it has high volume and high runtime frequency.

Why do we go with all those steps? Because we want value. If we add automation in our system, or if we add AI capabilities or anything, we need some value. If it is a dollar value, that is good. But if it does not have a dollar value, it still has to be some value that we are getting, maybe man-hour relaxation, where we can reduce man-hours from the system and then people can utilize that time in other kinds of work. Any value that we can get from that automation, we need to identify that value. Then only can we implement the automation. These are the factors we keenly look into when choosing a candidate for automation.

After a process or system gets selected for automation, because we see some value in it, we follow a four-step process: ESSA. The first one is eliminate. Yesterday there was a session where they were saying our architectural diagram is full of entangled threads. The first and foremost thing is to eliminate complex steps, redundant steps, or unnecessary steps. We do not know what processes have been running since legacy days, but they are still running. Nobody is asking why they are running or what their significance is. We need to identify which steps are necessary, which are unnecessary, and which are complex things we can simplify.

Then simplify. If we can simplify a process or operation, then we do that by reducing complicated steps or introducing straightforward steps. Then standardize. We have to standardize the process so we can see the value, see the steps, visualize the steps, and then try to automate those processes. Without these three steps, it is very hard to introduce automation. If we introduce automation without them, it will improve the inefficiencies. That is why we first introduce these three steps, then with a simpler process we can do the automation. Those are the four steps we usually follow in our organization to make a process or operation automated.

We do not only use AI. We have a lot of different processes that we use in our organization. For example, in the financial industry we have statements, records, and different communications to our customers. A whole lot of records are there. If an audit comes and they want a 1996 document, we need to search a pile of records. That is why we use an IDP process. IDP can convert PDFs to an online, digitized format and then search the documentation.

Similarly, for high-volume bursts we use RPA, robotic process automation. If we can use RPA to automate some process or operation, why go to GenAI? GenAI has a cost to it. If we need infrastructure, maybe GPU or TPU, to stand up that framework, there is cost. But if we can simply write an API to make a process automated, why not? Why go to a complex process to use AI?

We also use process mining for transactional data: if we have to identify transactional data and visualize the whole business process, we can use process mining. The AI/ML-based things are newer; we introduced them around 2022 onward when ChatGPT and everything came out and everybody was going after GenAI. Then we introduced some of the features we could do. BPM can simplify and visualize the whole business process so that business people, who do not have the technical knowledge to go in and see the process, can simply visualize the workflow and understand the whole process. These are a few processes we use in our organization for automation. GenAI is one of them.

These are the generic GenAI capabilities, not only Discover-specific: chatbot, document search, speech to text, text to speech, image to text, topic modeling, and summarization. Most organizations across industry are using GenAI in one of these features. In Google search, for example, there is an option using AI. It provides options like chatbot, image-to-text if you upload an image, and speech-to-text if you talk with the chatbot. All of these capabilities are getting used across industries.

Now I will give some use cases. As I mentioned, Discover is a customer-based company. We keep our customers first, and that is why we put a lot of effort there. We always feel customer is God. Whatever they feel, that is our feedback. We always try to go to them, call them, and get their feedback. That is why we create generic capabilities for our customers.

The first capability is for our customers and agents. The problem statement is that somebody calls our customer service saying, "My credit card is not working, how can I fix it?" or "I have these statements and I got overcharged; how can I revert that charge?" In that case, our agent needs to go through a whole lot of documentation because there are policies and processes to answer those questions. Our agents have to go through the documentation, get the answer, and give it to the customer.

We simplified that by using chatbot capabilities, where our agents can ask the chatbot. It is not "Hey Google," but an internal Discover chatbot where they can type the question. It searches across a lot of documentation using document search and finds possible solutions. One question might have three or four solutions, so it gives options to the agents.

This is where human-in-the-loop comes in. The large language model can provide a lot of solutions, but with a human in the loop, he or she can decide whether it is the exact solution or not, because we cannot directly or blindly give the answers as they are. The agent decides which is the correct answer and answers the customer accordingly. If the agent does not like the answer the LLM gives, there is feedback to refine the answers or questions. There is a whole architectural flow for that. We introduced this back in 2023, and we got feedback that agent productivity was highly impacted. Average handle time of calls got reduced, customers were satisfied because they got answers pretty quickly, and agents were satisfied because they did not have to sort through a whole lot of documentation.

The second use case is similar. If you call customer service and they do not have an answer, they usually route you to a different agent. With that different agent, you have to explain again: my name is this, my account number is this, my assessment number is this, and so on. Now we introduced a generic capability where all the things you gave to the first agent go to the second agent as a summarization. The second agent gets all the context from your first interaction and then answers the question. It improves both customer and agent satisfaction, because the customer does not have to tell the same things again and again, and the agent does not have to spend time collecting those answers. We use the Vertex AI language model suite with customized prompts for that AI model so it can help solve our customer problem.

The third use case is very important because for that we received the CIO 100 award this year. From the U.S. banking industry, Bank of America and Discover were highly recommended for action and risk reduction. For this, we also got an award.

We have obligations to communicate to customers. If any policies change in their account or credit card, we have to inform customers through letters or calls. We chose calls because we can connect with the customer, get their feedback, and improve ourselves. But the problem is that we cannot call at any time because customers have their own obligations: meetings, office, hospital situations, and all those things. How can we identify the correct time for a customer and call accordingly, so the customer gives the answer properly and we also get the proper answer?

If we call them during office hours or while they are driving, they may just say yes, yes, yes, like we normally scroll down through policies and agree. But we want a proper valid response from our customers. So we implemented a model. We created LLM models where we get the customers, their contact information, and their preferred time. We call it a no-contact model, where we restrict any calls to customers during that time. If that time frame is there, do not call. If it is an allowed time frame, then the call can go from our agents and the customers can answer. Sometimes customers may say, "Call me later," and we ask what their suitable time is. That data is fed to the model again, so next time the call is initiated, the answer given by the customer will be taken into preference.

We got a lot of success in that and a lot of response from customers. They were really happy with that model implementation. The advantages were customer satisfaction, because we were not disturbing them; accuracy and correct information, because we were calling during their preferred time; and data set coverage from 20% to 100%, because we were filtering the customers during the right time range. This was introduced in 2024, and for that we got the award.

There is another use case for training our agents. When we train our agents, we have to give a lot of classroom training. Sometimes classroom trainings are not possible. During COVID, it was all virtual. We usually record the trainings or classes and give them to our agents to follow the recording. The caveat is that if something needs to change in that recording, we cannot change it because it is static. If there is a policy change in a state or anything else, the whole classroom or whole training needs to be recreated and recorded again.

GenAI provides a solution called text to speech. If we have a script, a text file, or a Word document, just change the portion you want to change in the document. Go to the page number, change that policy or whatever page you want to change, and then use the text-to-speech workflow, which can create a recording instantly. Once the new recording is created, give it to the agents to utilize. It is pretty simple. It also helps our coaches who train agents concentrate on other things, because they do not have to go and train classes every time. Once the recording is done, they can focus on other fruitful or impactful work. Our engineers can just change the scripts. If anything needs to be changed, the recording does not have to be re-recorded again. That is one use case we use, and most of the industry uses it because text-to-speech is a pretty generic solution workload.

AI has a lot of possibilities and capabilities, but since we are in the financial industry, and since all industries are regulated and finance is highly regulated, we have to be very skeptical when using any third-party tool or open-source software. We have to use it responsibly. That is our next thing: use GenAI responsibly.

A story came to my mind. Last week, I got a new phone that has Google Gemini. My daughter quickly took my phone and started asking, "What is this Gemini?" I told her it is an AI tool that can be used to give answers. She asked, "Can I ask any question?" I said yes. She asked, "What is three plus two?" I told her to ask a more complex question. She asked, "Who is the president of the United States?" and she got the answer. Then she said, "It can answer everything. That means AI knows everything." I told her, not everything; whatever we fed it, that is what it knows. She said, "My life will be so much easier when I am growing up." I told her yes, that is true, but she has to be responsible. GenAI is good, but only if you use it responsibly. If you share personal data, location data, card data, or anything, attackers can easily get your data. That is why you need to be very careful.

The same thing applies in our industry. If we are using GenAI, we have to use GenAI very responsibly. If our engineers are giving ideations or innovations, we have to go through a lot of steps to get them implemented. First, we need ideation and sponsorship. Then we explore and finalize, like a POC method. Then governance. This third step is very important. There is a GenAI council, which recently got merged with Capital One. The council reviews all ideations very carefully, including impact, outside effects, compliance, vulnerabilities, regulations, and data security. If any step fails, the idea is rejected, saying it cannot be implemented right now due to those things. If everything passes successfully, and auditors and regulators give a green signal, then only can we implement.

After implementation, we do monitoring and reporting. We have to continuously monitor the models we are using because we have to see how the models are behaving. Sometimes data corruption can happen. Hallucinations can happen, which can give wrong answers or wrong solutions to questions. We have to keep monitoring that, keep looking into data quality, and check threat assessments frequently so that our LLMs are safe and can give proper and valid answers. That is the kind of flow we follow to implement any GenAI ideas. Our motto is: if we want to use GenAI, we have to use it with responsibility. That is why we have not built agentic AI yet, because we are thinking agentic AI might have some data going away or something that we do not know. We are exploring multiple options, and definitely we will start using that, but right now we do not have it.

In conclusion, we are using an automation-first approach in our organization. We want to automate every process, but along with that, we want to see the value. If we implement automation, what is the value: dollar value, human-hour saving, or whatever else we are trying to do? It has to have some value.

Automation is becoming easier because there are a lot of tools available right now, like Codex, Perplexity, Gemini, and others. Yesterday we went through the Gemini CLI, and it is pretty straightforward: create a QR code generator, and it gave us a whole set of applications. It is pretty easy to use, but there should be guardrails on GenAI usage. If data is getting used, data privacy and data security need to be checked every time. We have to continuously do threat assessments. We have to continuously audit the models, and keep all those things in mind.

Yes, GenAI is very powerful. We definitely use it. I am not against any tools that we have in the market, but we have to use them very carefully so that we will not be in a problem later on. We can keep using that without knowing what is going on behind it, and then after certain months or years learn that our data is breached or corrupted. We need to be responsible. It is better to be preventive than sorry.

This is how the financial industry is currently focusing on GenAI. This is a Gartner review. There are different bank workflows now. Banks are working toward completely automating everything. You can create an account, close an account, and do everything at your fingertips. Those are use cases from a banking-industry perspective.

That is it from my end. Any questions?

If not, I think I am on time. Thank you. Thank you very much.