Generative AI in Finance – Dwarika Patro Discusses AaDi’s Impact on Analytics and Decision-Making

In today’s fast-paced business environment, finance functions face increasing pressure to provide accurate and timely insights across multiple departments. Dwarika Patro, Co-Founder and COO of Aays, addresses these challenges head-on in our conversation. With 20 years of experience in professional services and a strong background in digital transformation, Dwarika previously served as a Management Consultant at EY-Parthenon. He holds a B.Tech from IIT-Dhanbad and an MBA from Imperial College London, where he honed his skills as a technology strategist. Dwarika discusses the challenges finance leaders face in seeking real-time insights and how generative AI tools like AaDi are reshaping the way organizations interact with data and analytics. He highlights the pressing need for a unified view of data and introduces AaDi, a groundbreaking generative AI tool designed to revolutionize financial analytics.

AaDi enhances decision-making by offering real-time insights and simulation capabilities, acting as a virtual analyst that integrates seamlessly with various data sources. This innovation drastically reduces the time to insight, transforming what once took days into mere hours. With its explainability features, AaDi ensures transparency in the analysis process, while its multi-layered analysis capabilities tackle complex financial queries. Early deployments of AaDi have demonstrated remarkable productivity improvements, enabling users to uncover new optimization opportunities and significantly reduce manual efforts.

Key Takeaways:

Transforming Finance Analytics: Generative AI tools like AaDi are revolutionizing finance by enabling real-time insights and proactive decision-making.

Productivity Gains: AaDi can enhance productivity by 55% by streamlining analysis and reducing the time to insight.

Integration with Existing Models: AaDi works seamlessly with traditional machine learning models, making advanced analytics accessible without extensive technical knowledge.

Importance of Explainability: Emphasizing trust, AaDi’s evaluation system helps users understand AI-generated answers through a nine-parameter assessment.

Multi-Layered Analysis: AaDi’s “analyst mind map” feature allows finance professionals to explore complex questions, enhancing analytical efficiency and insight generation.

Kashyap: Hi, Dwarika, thank you for making the time. So today we’re talking about financial analytics and the role of generative AI in this area. I understand Aays Analytics has been making strides here, particularly with your new product, AaDi, which is gaining traction with large enterprises. You’ve been helping these companies leverage AaDi to address some key challenges. To start, what would you say are the main pain points finance teams face in large enterprises today, and what inspired you to develop AaDi and other solutions to address these issues?

Dwarika: I think you know, Aays has a very deep connection with finance. We started as an organization, and, contrary to many others, we began with finance analytics—that’s where we germinated. Our entire journey began there, as that’s the background me, and Anshuman, the other founder came from. So, we’ve been working in this space, focusing on finance and analytics.

As you understand, finance is a very horizontal function. It cuts across multiple functions within any organization—it touches marketing, sales, supply chain, and more. The biggest challenge for finance at this stage is achieving a single pane of control over all the information flowing across multiple functions. This challenge is especially amplified in large organizations, large enterprises, which we generally work with. Understanding and gathering this information, then making sense of it—that’s where the complexity arises.

There are also a lot of macroeconomic and microeconomic challenges. From a business perspective, there are various obstacles, but from a data analytics and finance perspective, leaders are focused on building solutions that provide real-time insights about their business, enabling smarter, preemptive actions. Additionally, these solutions offer simulation capabilities so they can plan different scenarios and build resilience in their business. So these are the key top items, I would say what we observe in the market today.

Kashyap: In relation to that, could you shed some light on AaDi? First, I’m curious—how did the name AaDi come about? And, building on the points you mentioned earlier, could you explain how AaDi addresses the specific pain points and use cases finance functions and large enterprises are currently facing?

Dwarika: I’ll come to how the name AaDi came. But before that, I would just like to spend a minute on how generative AI is helping here. If you think about it, the finance function in a large organization does multiple types of analysis, which has been very traditional in nature and very agent-driven, meaning human-driven. A lot of reports are run, information is analyzed and absorbed, synthesized, and then reports are created that are consumed by multiple stakeholders.

Since we have been working in finance analytics for a long time, we have developed multiple tools that assist in decision intelligence for the finance function. However, with generative AI, what has changed is how people interact with that information and these capabilities. Generative AI is very effective in creating content and personalizing interactions with these machines.

AaDi acts as an orchestration layer on top of all the traditional capabilities built in the past, plus more, and helps you access those tools in a human-interactive way. Essentially, how it works is this: let’s say you are the CFO of your company, and you want insight about your business. What would you do? Typically, you would go to one of your analysts and say, “This is my question or hypothesis. Can you check the facts around it?” The analyst would go and search, depending on the organization’s maturity, traditionally finding that information in transaction systems, using Excel, or consulting Power BI or Tableau reports. They would analyze these reports and come back to you.

With the advent of generative AI, that’s where AaDi comes in. We have deployed generative AI, where you don’t need to go to an analyst for any of your ad hoc reporting. You simply ask your question to the AaDi interface, and AaDi will act as an analyst for you. It will plug into multiple data sources from your perspective, get the right insight, and present it to you. You can iterate over it if you are not satisfied with the answer. This approach cuts down your time to insight, as it acts as an independent agent to do it automatically.

That’s where the name also comes from. AaDi stands for Aays Analyst Decision Intelligence. We see this as a tool that will replace many of these analyst-type roles in very mature organizations.

Kashyap: One question that arises in my mind is regarding the application you just mentioned, where you scan different databases across functions to generate insights. Why are these capabilities specifically better suited for the finance side of things? Similar use cases are being built for other areas, and these capabilities can also be extended beyond finance. What makes them more effective in the finance function?

Dwarika: I think this question has been asked multiple times to me, and it can definitely be trained to other areas. However, we just wanted to focus on our core area, finance, as we’ve been discussing. What we have done is take various models from the market, combine them, and also train them very heavily on finance data, finance terminology, and finance analysis. That’s where this finance focus comes in.

But this is a framework. Let’s say tomorrow someone wants to build something similar, or  if we decide to build something for the marketing or sales function, nothing stops us from doing it. It’s just that we have a deep understanding of the finance space, and we understand the key pain points of the finance stakeholder, which is why we have trained it well on finance. But it’s a framework that can be used for anyone else.

Kashyap: My next question is about AaDi’s claim of improving productivity by 55%. In your experience, what are some of the key factors or parameters considered to arrive at this figure? What analysis has been conducted to come up with the 55% improvement in productivity?

Dwarika: So this 55% that you see is a very conservative number, I’ll explain. There are multiple layers of effectiveness or productivity improvement we’re envisaging using this. First is, as I was also telling you, the time to insight is much faster. You don’t need to give a task to an individual, where that individual performs the task and comes to you, and then you take any action on purpose. In the areas where the model is well trained, it can significantly cut down those times. 

However, the other advantage is that if you look at a decently matured organization, you would have built so many reports, dashboards, or BI/ AI capabilities over the past. Many of these organizations or stakeholders don’t know which tools to use and what is built for what purpose. That is becoming a very big challenge for a lot of organizations. That’s what I understand from my discussions with a lot of senior stakeholders in this space. 

AaDi is very good at understanding because it’s all encoded into it. Let’s say, if you are doing financial planning, which database to go to, which report to look into, and what data to pick up – AaDi clearly understands that, so it gives you that effectiveness. 

The third part is AaDi is not only a generative AI engine, but it also sits on top of a lot of traditional machine learning models that we have built – It could be simulation, scenario planning, and other capabilities. It unlocks those tools to you in a very seamless manner. 

All these things combined, we have done certain scenarios with our customer base and  those piloting our product. By combining all this, we feel that a minimum of 55% productivity improvement is what we can provide when this tool is well deployed.

Kashyap: Can you provide more specifics about AaDi? I read into it a bit more detail and noticed that it mentions using an analyst mind graph. I’d like to understand the technology behind this in greater detail. Additionally, how does the analyst mind graph compare to traditional machine learning models, and how does it give an edge to the AaDi product in helping finance functions?

Dwarika: So it’s a very interesting feature, analyst mind graph that we’re talking about, and that’s where I think the core differentiation comes in. I’ll explain how. Many of these questions that we ask every day, on day-to-day business, are multi-layered in nature. For example, if I want to understand my sales number for this period or this month or this quarter, it’s not a unidirectional question. You want to understand how I fare this period, this month, or the current month versus last month, versus the same month last year versus my moving average? How am I doing from a statistical perspective? Whenever you ask a question, mostly in business, it is multi-layered. You want someone to analyze this particular question in multiple dimensions and then give you the right insight. AaDi, because this is where the analyst mind map comes in, has encoded this kind of thinking or this kind of chain of thoughts within this application. We know how finance people think and how finance people do analysis, and we have created those chains of thought capabilities, similar to what people are nowadays very excited about when they say, GPT-o1, which is really good in chain of thoughts. 

We have been doing this for the last one year in the finance space. That’s the analyst mind map thing, which is encoded there. It helps companies or stakeholders ask very deep questions, and the model understands these deeper questions and does multiple levels of analysis and comes back to them with whatever is the most significant answer for them, so that they don’t have to juggle through a lot of data. It becomes very seamless for them, and that’s where, again, even more efficiency comes into the picture.

Kashyap: Continuing on the understanding of the technology behind it, one feature, especially in finance functions, which is a very critical area, is the explainability aspect. It is important for someone working within finance to trust a co-pilot when it is answering questions and giving specific numbers. Explainability is vital, especially when generating insights and conducting simulations. You mentioned using something called step intelligence maps. What is your take on some of the features built through these intelligence maps, and what are some of the features included in Adi to ensure that this AI is explainable?

Dwarika: Very good question. It is very close to my heart because with AI, the relationship between humans and machines is changing fundamentally. I think we are very used to trusting people, and that trust is won over time. When you meet someone, you probably don’t trust them so much, but over time, as the interaction grows and as we know their value systems and behavior, we start trusting them. But for that, it is important to build that  partnership with that individual. Same is with machines because these machines are emerging properties of creating new stuff, emerging properties of understanding your needs and so forth. I feel fundamentally it is no different. You need to understand what these machines are, how they are trained, and what values they carry. Can I interrogate this machine and understand how they have calculated this or how they are giving this information? When we built AaDi, it was the first principle we applied that whatever answer it presents to the user, it should have a complete trace of that path. Be it the mind map, what mind map it has used, what framework it has used, what database it has used, what kind of SQL queries or any other queries it has used. We give that as a lock to our users. Sometimes, some users are very technical; they might understand. Some users are not very technical; they might not understand. For business users, we have translated those technical terminologies into more business-friendly  answers. For example, if you want to find an answer by querying five different tables, we will give that information in business-friendly jargon so that they understand which table or which data was referred to and what filters were applied, what calculations were done. That is critical. We’ve also built a responsible AI framework where we evaluate each answer by nine different parameters. We also give a cumulative score, which tells the users whether they can trust this number right away or whether it needs a human in the loop to verify this answer. We also encourage them to build the relationship with the machine by giving it feedback on a continuous basis so that over time, they can trust the machine. It’s a journey; I won’t say it’s not like traditional deterministic software where you give A, and it will always give you B. It’s not like that because these are probabilistic models that we’re running. I think it’s a bit of a new way of working, a new way of trusting the machine or not trusting the machine. There’s a journey everyone has to take to really make it happen. What we’re trying to do is to make it as seamless as possible, as simple as possible, but I think there’s a long way to go.

Kashyap: How does AaDi effectively establish the cause-and-effect relationships that are crucial for generating insights? While numbers often reveal the “what,” understanding the “why” typically requires human intervention or experience. Is AaDi capable of learning to generate these “whys” on its own, or does it incorporate a human-in-the-loop approach? What is the thought process behind building these cause-and-effect relationships in AaDi, and how can customers leverage this feature to maximize the value of the solution?

Dwarika: I think all these ‘why’ questions or the root cause analysis type of questions, originate from the chain of thought and the analyst mindset that we have created or encoded in the application. For instance, I may ask why this number is higher this year or this month, which  needs a lot of chain of thoughts to analyze the numbers in multiple ways. What I’m trying to say is that every “why” question can be translated into multiple chains of “what” questions. That’s where the key feature that we’re trying to build lies: Every “what” question can again flip into “why” questions, so you can have multiple levels of chain of thought. It’s a multi-tiered chain of thought that we want to build. 

At this stage, in the current version of AaDi, we have  created  one tier or one level of chain of thoughts when a user is asking a ‘why’ question, AaDi effectively creates one layer of ‘what’ questions around it. Then it also constantly understands or absorbs what type of next questions are coming from the user. In a way, it is learning on the job to understand what “why” questions create which “what” questions, and then what is the next tier of the chain of thought. While we have incorporated just one layer of chain of thought, it is continuously learning the next layers of chain of thought. It is a journey. It has not completely solved every ‘why’, but it is learning very fast. Models like GPT-o1 are helping it, because some of this naturally comes from GPT-o1 as well. We will get better and better with these kinds of new things coming in. 

This is where I think domain expertise is very important. Without understanding the domain, you cannot do a root cause analysis; you will just be throwing stones in the dark. That’s why we are being very careful; we understand not only the domain, but even subdomain. For example, in finance,  you know a P&L person and  a treasury person looks into things very differently. We are also trying to train AaDi at a very subdomain level. We need to go very deep to really make it work effectively. That’s the challenge we are addressing.

Kashyap: Are there any case studies or success stories from your customers who have shared how AaDi has protected them? Given that it’s more of a bottom-line application aimed at reducing costs and improving productivity, has any feedback stood out since AaDi was implemented or scaled within their organizations?

Dwarika: We have already deployed AaDi at three different places. We generally always believe in starting small and then learning quickly and then expanding.  These are in different sub-domains within finance and one in the intersection of finance and supply chain. Out of these three, two have gone to beyond POC phase, into more like MVPs, where we are trying to implement certain capabilities for the users to start using it on a day-to-day basis.

The key responses that we have got so far are: A, it is unlocking a lot of ad-hoc analysis which users could not have done without a technical user. So, their dream state is that if we invest enough time and money on building the solid data layer, we would deploy some tool like this, so that we don’t have to create a lot of dashboards. AaDi will sit on this data and then build the BI reports or the other decision intelligence on top of it automatically as much as possible. So the feedback we have got so far is that time to insight has become much faster. Instead of days, it has become hours. That is the biggest takeaway for us. 

B, some of the insights that we are talking about, the users have never seen or they’ve never thought that will come. It is because of a lot of emergent properties that these insights are possible. 

For example, AaDi can reveal new areas where they can optimize their cash flows, areas where they can identify specific classes within  accounts payables or accounts receivables by performing clustering analysis. This is something users may not have considered doing manually.  But with AaDi, they just tried to experiment with it, and it really worked and gave some really cool insights. I’ll not take names, but there is some really interesting stuff that is really exciting and we are investing heavily on AaDi. We have invested a lot of R&D budget on this so that we can make it a much better product each day. 

Kashyap: On that note, thank you so much Dwarika for making the time. It was a lot of fun to learn about AaDi and all the very best for you.