#05 - Data to Decision: Leadership in the Age of AI with Sravan Kasarla

So here is what my tagline, right? For myself and how I see the data world and the impact. So really, empowering experiences and business outcomes with actionable data. That’s what I believe in, and that’s how I approach whether it is data strategy that I need to create, architecture I need to create, enablement of capabilities, or management around it.

Yeah, hopefully we’ll get a chance to double click on a few of these. Yeah, that’s wonderful to know you’ve put it together very well, Sravan. I appreciate that. As a CDO how do you determine which data strategy to use when we speak of setting up one for a traditional data use case, where you’re typically wanting to do some integration between applications and systems, as opposed to when AI is the focal area for you. So how do you determine that data strategy, what to use and what not to use?

I know AI has got a real good attention to data. Data and also a lot of a classical AI existed, by the way, I think it is exactly 24 months ago, nearly 24 months ago that the chat GPT got released and suddenly AI has become a mainstream word for even from the youngest to the oldest.

And everybody started using it, which is great, right? It’s a really defining moment in the data and the AI and the technology world. Again, AI is not limited to data side. I think that’s another one we will probably visit. But the thing is, while there is a different technology and technique we are applying with the generative AI or even just simply AI, but foundation of that still lies in the data.

So from a data strategy perspective, at highest level, when I talk about what do you do in data strategy, you’re trying to really provide much better actionable insights so you can light up the outcomes or create some new experiences. I’m bundling many different things, smaller business cases and everything under these two categories.

You’re either doing that meaning business outcomes is you’re now trying to create your sales organization to be more productive or increase their revenue per book. In the financial service industry, it’s the most common thing is book revenue to be increased or number of appointments or conversion to be increased.

Yeah. All of these are set of outcomes you’re going after and you’re trying to create your data strategy to enable and address that. So it’ll aid to the advisors or whoever that is doing it on the other side, on the experiences. Thanks to the whole the iPhone moment that now we talked about for generative AI.

There was a real iPhone moment a long time ago which fundamentally changed. Insights have become not just executives. Insights is everybody. We are looking at how many steps did I walk? How is my average pace? And again, what is my resting heart rate and everything? All of us are consuming a lot of insights and these kind of experience we expect from our enterprises as well.

So the question of what experiences are you enabling? So outcomes and experiences become the basis of any strategy, I would say, whether it’s a business strategy you’re doing, and as a result, you’re creating technology strategy or you’re addressing your data and AI strategy. End of the day to enable something to create experiences, but there are key differences in the elements of the data strategy when you do it for call it traditional data strategy versus the data strategy for AI.

I think the first and foremost is data volume and variety. Clearly, there is a bigger difference. We keep talking about, I think, large language models started talking about. Now we are talking about half a trillion parameters from a 4 billion to 10 billion to 20 to 200 billion to now we are talking half a trillion and if they did these parameters and nothing but label data sets that getting supplied to the model, right?

So the volume is more important to create the accuracy and create the ability for the models and the AI to learn. And generally, let’s call it regardless of type of model. So the data volume needs are very high and the variety of the data, not just structured unstructured data that we used to talk when the big data thing started 15, 20 years ago.

Now we’re talking multimodal inputs. I think that’s another different thing. So you’ll start identifying, okay, how do I enable? What are the elements of the data strategy from a data volume and the variety perspective. Number 2, data quality and data preparation? Quality is absolutely important, whether you’re doing traditional strategies to provide reporting or simple insights or simple predictive models to more advanced experiences and the creation through the large language models, but

preparation. How do you label your data? How do you prepare your data so that your models can actually understand it? And you’re contextualizing the power of the really large language models in this case to really apply to your organization. Lot of that I think data preparation suddenly becomes a significant part of the strategy, which was never part of the traditional one.

Data preparation was a step in the data science process. That’s it. That’s all it was. You never talked about data preparation per se, data labeling, preparation, identification of data, lineage of data becomes a lot more important now than ever, not just for controlling this, but actually it will, the output of that can be wildly different or hallucinations might happen if it didn’t have the right labeling, even the right data, but not right labeling.

So these are one other big difference, right? There are a couple of more quick ones that I see data processing and how you approach the analytics. There is a difference in the way that you’re processing it. You’re processing it in a traditional data says strategies are traditional data uses.

You’re structuring it. You’re giving a lot more structure as opposed to not right. So there are those differences, data governance and management. There is a bigger difference. The goal of traditional data governance has been to control access. Sure, understanding and focusing on compliance and security and access control.

Now from there to, yes, you need to make the data available rapid experimentation. Compliance and privacy still becomes paramount. Now you’re talking about model efficacy. So now the management and governance takes a completely different turn about. It’s not that once and done. That previously was more static.

Now it is more active data governance and including model governance and risk management and responsible AI that these are going to be additional things that you need to do different. Obviously utilization. There are many areas and ethical considerations. If you had yes regulators always made sure that you need to know your data lineage, where data is coming from and all that in a traditional data strategies and banks and financial institutions were required to show that you knew your data.

But in this case the ethical considerations are so big that it can be life changing recommendations, life changing interactions can happen through the AI because data is really touching the end consumer directly than in the traditional data sense. So I think there are so many ways that this can be different but data is still the most important. Data foundation becomes the most important thing that trusted data that you have the right label that you have aligned this for the right outcomes and experiences. I think on that level, nothing has changed. Yeah, I think you’ve again put it really well. You’ve touched upon all aspects around the essential components around data management, I would say.

And more important I’m seeing is as we are putting a lot of AI use cases, AI solutions around those use cases. Data is becoming a essential aspect, it’s the center of those things where tagging that data, labeling that data, classifying that data, and making sure the quality of the data that the models are ingesting and using to make determination around that solution and are also able to generate things on top of it is becoming very essential. Yeah, that’s really good. So this leads to a question, Sravan, I’ve always been thinking about it. So who should own the AI strategy within a company?

What do you think a CDO is the best person to lead that AI strategy, or is it Chief Analytical Officer, or should it be the CIO? So I think there’s a overlap when we talk about who owns the AI strategy within a company. Yeah, and how should the AI strategy integrate with the overall corporate strategy?

Because that’s the most essential thing. Yeah, no, I think you touched up on a question which is currently in debate. And as I mentioned early on AI is not just data people, not the data world, just like the way that e-commerce are digital. Digital is a business model and a business channel.

E-commerce back in the days, I’m dating myself by calling it e-commerce, right? So e-commerce or digital, all of these were not owned by a technologist or even CIO. They’ve had a big role of enabling that one, no question about it, right? So to me, the need to be a business leader whose job is to really service the clients and generate the revenue, they should be the one who should ideally own the AI strategy for the company. I think it’s the CEO or one of the CEO direct reports are the ones who happen to have revenue generating or a growth responsibility for the company. They are the ones who need to be actually owning the AI strategy.

I’m not advocating for the AI execution, AI enablement and the strategy about how you enable it. No, that’s definitely in the technology realm on more importantly, CDOs are now turning to be CAIOs as well as CAO that are also can do that. When it becomes clear about how the company looks at it.

So to the 2nd part of your question about how should AI strategy align with the business strategy? AI strategy need to be part of the business strategy, it cannot be separate one. If you have a separate strategy, ownership and the adoption and the value generation is not going to be as much.

Why is it different? You could argue that technology strategies owned by a CIO or CTO, CDTO and business strategies owned by the strategy leader and CEO. Isn’t that like that? No, I think this is where AI differs, not just as an enabler, but as a way of doing business. As a way of optimizing and automating your processes, so it’s changing your business.

Fundamentally, it has ability to change your business fundamentally, unless the business leadership has really looked at this as one of those big strategies or the tools in their business strategy is not going to be as successful. Sure enough that you could do few automations and couple of fancy use cases, which are low hanging fruit, but you can’t really make this to be part of the DNA of the company.

So to fade shortly, AI need to be infused into the business strategy. And AI strategy need to be owned at the highest levels of the company, meaning should be driven by and strategy should be defined at the ELT of the company and the board need to be the ones that who are well familiar with that. When it comes to actively enabling that significant role, I think a CDOs or the chief digital and technology officers are the ones who would enable that one, depending on the company and the structure, but this need to be part of how business thinks about changing and looking at this as innovation and disruption and how they will change the business model or how they will add on more services or products and offerings using AI as one of those differentiators, just like digital and just like the e-commerce of the past.

What you’re saying is so right Sravan. So the CDO and the CAIO should be able to work together hand in hand on a day to day basis so that they are able to set up an effective AI strategy for the company. But the ownership and driving it to be at the very top levels of the company.

Yeah, just be the functional, not the functional leaders, right? But it needs to be at the company leadership. Yeah, so we are talking about generative AI and many people are, they do not understand the difference, Sravan, between what generative AI is actually all about and what is classical AI, what is traditional AI versus now everybody’s talking about something that is producing a different output than what they have always seen.

So what do you think makes generative AI fundamentally different from the traditional classical AI? So again, the word classical AI or world of old AI. And some people try to use it as old AI, but I think there is a difference. End of the day both of them are a class of artificial intelligence category itself, right?

The model themselves, is it the transformer model which is trying to generate not just a decision or a particular inference, but rather creating a novel content creation, whether it is a content, a text or image or a video using the data supply to it. I think that is at a very high level.

I think that’s the difference, right? That’s why it’s called generative AI because it can generate from what is being supplied as opposed to just provide an inference to that one. So classical predictive models depend on the clustering algorithms or the random forest type of models to more of a neural networks which are constantly there.

They’re learning in stages and continue to learn from the next inference to next inference. But eventually they’re able to actually generate based on their understanding of it. So I think that’s it at a highest level. I think that’s the difference between them, but there is a big difference in the accuracy between just the neural networks that what they can produce, yes, they can handle a very large amounts of data and then continue to derive and increase the accuracy of them using large amounts of data, but at the same time the classical AI models which use more traditional techniques, whether now what they’re called traditional techniques are proven techniques of random forest or clustering algorithms, end of the day, are able to come to a conclusion or a inference that which is more reliable than what a generative AI model would do.

Again, I am not the technical expert when it comes down to how the transformer models work differently. How the GPDs work differently? But this is where the categorical name of the Generative AI came is because they’re able to generate a novel content based on their learning through the neural networks and other techniques.

So what you’re saying essentially generative AI is something that produces a dynamic output which keeps changing as the model learns from the behavior of the data and classical AI or traditional AI, of what we know as AI is more of a static response, static output based on a trained data set that has been trained for a very long time on a historical data set and also on a particular use case.

Okay, right? That’s where the difference is, right? That’s where the reliability of the classical AI is. Static is good because the same response it’s going to give based on the parameters, meaning, you can rely on if you’re trying to create a risk score, you’re trying to recreate a basically propensity model that you created to see who are the best people to do it.

You don’t want that to be randomly changing and you don’t want that to be variably dynamic that you’re not even going after the same clients that you actually want to go or your risk model is so far off every time that you can’t rely on that one to provide your underwriting that pricing that need to be better.

So that’s where if you look at the use of either Gemini or ChatGPT. Let’s take those two as examples. The same exact question. So if I put a question to say, hey, how is the data strategy different for AI versus the traditional and then two different windows, two completely different responses.

Few things are consistent, but everything else is because like every human you ask with the same input, they have a different interpretation, so that’s where the use of that generative AI, you have to be cognizant about what you’re applying this for. So this is where, when generative AI models came out and everybody wanted to go and adopt and CEO saying, Hey, what’s our AI strategy?

Where do I use? Why don’t we apply generative AI now for almost everything, right? You can’t, right? It is a tool in a toolbox that can do few things, this generation of the content, being able to learn from it, being able to summarize what you have to more reliably create a response and predict something with a reasonable accuracy in a classical model.

So I think people need to be cognizant about what they’re applying this for, not use generative AI where a simple predictive model with a technique that which is proven is the right one to be applied. Yeah. So do you think the volume of data and quality of data has an impact on the accuracy and effectiveness of what the output is going to be for a Gen AI application?

Do you think they impact? Gen AI is a lot more sensitive to actually quality of the data. And that’s why we talked about originally, right? When it comes to data strategy, labeling of the data and tagging of the data, that becomes so very important because the way that the neural networks do process and send it to the next network, the layer into next layer.

And also, when the the programmable transformer is trying to finally set the input, this becomes very important and they become very sensitive. That’s why even wildly different hallucination came out because labeling of the data or the tagging of the data was way off or they’re out of context.

Actually, there is more resilience in the traditional models to even in the data drift. They are not fundamentally falling apart. So they can handle a little more of a data drift than what the GPDs can do or what the large language models can do, but large language models don’t tell that they don’t know this answer.

They’re confidently wrong. They’re confidently wrong with their answer. They’ll give you a completely different response that which is not even relevant, but they’re pretty confident about it, meaning that responding it, not saying that I don’t know answer. I don’t think you will ever get the I don’t know answer from these models.

So do you think if more data is being fed to these models that are constantly being wrong, the effectiveness of the output accuracy is going to increase? Not just more data, right? Just throwing more data, more bad data won’t help. Better label data, better tagging done, yes, it can learn the context once you have the right labeling and right tagging, then it’s able to, that’s where it can accelerate.

It can quickly relate to and then come back with even a novel generation of content or novel response or novel decision or the let’s call it novel response itself that it can come out with. Yes, the more data will help, but more data better labeled. And better tagged and that’s when the accuracy of the models or accuracy of the responses is going to be.

That’s a good discussion. When it comes to leadership in data and AI, how do you think data leaders should adapt their leadership style in the age of AI? And what pivots have you seen, have you personally made in your leadership approach recently with AI being the focal area of many enterprises today?

Two or three, I think, a very significant differences in the leadership, both for the any technology leader or the data leader. Now, it’s not just limited to the data leadership. I think technology leaders is the question of understanding how AI can be applied. So understanding your business and understanding the fit of this technology more than ever in a traditional world, even though technology and enterprise software and cloud and others have come up, it was not that you need to fundamentally understand the business differently.

You’re looking at, okay, how do I optimize this one? How do I implement something which makes the business process streamlined? But I think AI, the biggest skill leadership change and also the skill that the leaders need to learn, I think more so for the technology and data leaders and somewhat for the actual business leaders itself is where can I apply? How is AI going to be impacting your business and your business processes.

So understanding the business itself is going to be the skill. It’s always required. You always say that, hey, you need to understand business before you do technology, before you do data. But the level of understanding that now you need to get and the familiarity about what’s in the day of the life of the business processes, business leaders going to be is an important skill to learn, number one. Number two, you have to be very cognizant about the right tool for the right use case because for the first time, cost of experimentation is significantly higher. Yes, while experimentation is very easy through a gateway or just a quick Azure Open AI or on the AWS Bedrock or somewhere, right?

All of these places you can just go and create an account. But if you’re not well defined in what you’re trying to do and determine that it is the right tool for the right problem, your experimentation and training itself is significantly compute intensive. That’s the nature of it. That’s why Nvidia continues to grow the way that it’s growing.

Because why do you need, why is there such a big increase in the amount of the GPUs and the specialized chips need to be done because the compute is very high because they do deal with a very large corpus of data and the large language models themselves are built on significantly larger

set of parameters, right? So I think simple experimentation, if you’re not sure, and you continue to experiment, every experiment you’re running is running up your bills really high. So that’s another one that people do need to understand is how do you experiment? So you’re rapid experimentation and being able to really be conscious about the right use case and optimization.

That’s another skill that need to be learned. I think everybody need to become AI literate. I think that becomes almost given, right? While you understand your business while you understand what to apply but you need to also understand let’s call it the technology of the AI. I’m not saying everybody need to sit down and start coding. By the way, coding is not needed, right? I can ask ChatGPT here’s my problem, code it and Gemini to code and Copilot to code it. It’s already doing it. So it’s not a question of that, but I think being able to get to understand because it’s right now at an infliction point that it is rapidly changing and the pace of change is more, I think right now, a good technical understanding of the AI is also going to be very critical for the leaders.

Yeah, good to know about that Sravan. We are already in the last quarter of 2024. What do you see as the future for AI and data leadership? Where is it going beginning, 2025? What changes are you seeing in the leadership roles? One thing is very clear, right? AI or the generative AI, thanks to generative AI, that AI is here to stay, and this is the next wave of disruption and innovation for every business.

It doesn’t matter which industry you are in, it’s going to impact you some way or the other. Some industries are going to be disrupted more quickly, others might be a little more slowly, but I think that part has been proven and that’s clear. But what’s happening is the, if I can draw on my previous response, the skill of knowing where to apply and when to cut it, right?

I think it is going to be critical because the expectations are way too high. The boardrooms wanted to see that everybody wanted to put that in earnings call that they’re adopting AI and they’re doing something with AI or that introducing products and services that which are AI powered.

Depending on the industry and the line of the services that companies offered every company in the earnings calls have talked about AI. They didn’t want to be left behind. Yeah, and then there was a bit of experimentation. Everybody started experimentation. Significant amount of companies started.

They’re all AI projects are running and the problem is there were thousand use cases, a thousand flowers bloomed, but only there are a couple of roses in that, right? That’s all it got limited to. Then the reality came down to is that with the maturity right now, here are only a couple of things that you can do.

So now you see that there was a heightened expectation. Now there is a little bit of disillusionment, almost a hype cycle type of situation now that. I think the people understanding yes, it’s here to stay, but there is a excitement from the excitement to experimentation to now you’re coming to more pragmatically to see that what you would apply.

Looking at that, I think 2025 and this year, I think, hopefully, with the financial markets as well as the interest rate environment and the election. I think there are 3 factors, inflation which is economy is still jittery about it. Fundamentals are different. And then the interest rates are not helping that companies who are financially strong as well, they’re holding back on going on a full fledged innovation.

I think 2025 is probably going to be a year that companies will be more focused about where they’re applying AI, not just go and experiment for the sake of labeling themselves as an AI company or AI adoption or AI powered products and services. I think that kind of reality will start, but at the same time, the innovation is not slowing down.

The amount of let’s call it new type of models, which are now getting verticalized and now models, which are more smaller language models, smaller large language models coming out that which are more reliable and accurate. I think this is going to continue. Tech industry is not stopping.

there is anything they’re investing, they’re betting on more with the economic cycle, but I think I see that there will be adoption, but more value generation will start happening in 2025 is where I’m counting on. That’s interesting to know. Let’s see. We are hoping for the best all of us.

Yeah, I’m an optimist. I’m an optimist, Monika. So that’s who I am. And then I guess get excited about technology, but apply to the right use cases where my caution is. Yeah. Yeah. Oh Sravan, thank you so much for talking to me today. Thank you for your insights. I highly appreciate it.

We’ll talk again after sometime maybe we can have another podcast. Thank you so much for your time today. Yeah, thanks Monika. I think you’re calling it Beyond the Bottom Line, right? I think that’s exactly I think people need to look at. I wouldn’t say beyond the bottom line, beyond the hype look at the bottom line. That’s what I would talk about when it comes to AI adoption of any companies. Yep. Beyond the hype. Look at the bottom line. So what? How does this apply to me? How does this make our company better? So with that, thank you. And would love to be coming back.

Yeah. Thank you. Bye bye.