The Product Mindset Meets AI

0:00Really spend the majority of your time understanding the problem that you are looking to solve and then look at the options of tools that are available, Genitive AI being one of them, and say, hey, how can this tool solve my problem in the most effective, cost-efficient, secure, safe, and compelling way? Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. Gen.ai promises to completely transform our world, how we work, how we live, everything, it seems like nowadays.

0:40But the question is, do the core first principles also change, especially in the realm of product management and strategy? And that's exactly what we're going to dig into today with Yuday Kumar, a product strategy and management leader with experience at Capital One, PwC, Volkswagen, to name a few. And Yuday focuses on data decisioning, and AI really striving to bridge that gap between complex technology and practical user-centric solutions. But welcome to the show, Yuday. Thank you, Matt. Happy to be here. Yeah, excited to chat about this. I love the strategy side of things. I love the product side of things.

1:18And to the initial point we just hit on, I think that's a big question. With Gen.AI coming in, it's new, it's novel, it promises this amazing future. But what's your take in terms of from a product mindset approach strategy do things start to change there as well what's the same what changes if anything yeah the questions you just posed are really fundamental questions and when i think as a product practitioner as a product person the fundamentals never change end of day when we think of why are we in the business of products is really to solve problems, to understand customer needs, whether they are explicit or whether they are latent, to help solve those needs in meaningful ways.

2:14And when we start talking about the solution part, the solutioning part, that is when we start looking at what are the tools and technologies that are available to us. And in that sense, generative AI, which is a subset of the larger AI category umbrella, is another tool set which has been made available to us, which has been offered to us over the past couple of years. whether you are in the healthcare business or whether you are in the fintech business or the education business, we all need to understand that domain. We all need to be passionate about the domain. We need to understand what good and great looks like to the different customer segments in each of those domains that we are passionate about.

3:11And then we go about understanding what their needs are and go about meeting them. Traditionally, if you roll back, it was classical software development. We were building and rather we were solving problems by first understanding the needs and then bringing the traditional tools back in my days. And I'm going to date myself a little bit. It was Visual Basic, Power Builder, those kind of things, right? The appliance server technology, right? And then Oracle and ERP, SAP, and all that stuff came along. We started to use those more elaborate enterprise-wide sort of tools and technologies to go solve problems alongside the internet came.

3:54And then we have the explosion of cloud and then AI came along and that offered us even more slick and scalable and secure and reliable tools and technologies. And now we have Genetive AI, which is offering another really interesting way to solve those problems. So I think end of day, it's going to be the fundamentals will still continue to be that we need to really, like they say, fall in love with the problem. Right. Really spend the majority of your time understanding the problem that you are looking to solve and then look at the options of tools that are available to you. Generally, AI being one of them, I say, hey, how can this tool solve my problem in the most effective, cost efficient, secure, safe and compelling way?

4:47Yeah, I love that. Fall in love with the problem. And with any hype cycle, you could call what we're in a hype cycle right now, people search for solutions and they forget about the problem a lot of times. I'm curious, how would you rate overall the world right now in our ability to solve for problems versus getting a little too excited and trying to search for problems on a scale of one to 10? What would you give it right now? I think we're still maybe at a five, maybe a six on a scale of 10 because I kind of see folks both when I'm talking to prospect clients or customers or even like I spend some time at college campuses talking to students who are looking to solve problems.

5:39And yes, across the spectrum, people do get caught up in the shiny object syndrome, which is like, oh, that's cool. Let me go do that. I'll do it that way. As opposed to like when I or we start having them clarify the problem and why is that a problem and who is the customer and do they really need a solution? Are they just happy and content with what they have? quick break in the pod if you're listening to this podcast chances are you've been thinking about how to actually use ai inside your business and that's exactly why we built the ai opportunity finder it's a free tool that helps you uncover high impact tailored ai use cases based on your business your goals your pain points and your industry no fluff no generic use cases just real ideas that fit your business and the rank by roi potential it takes about three minutes to run and it's like having your own personal AI strategist for free.

6:37If you want to try it for free, check out the link in the show notes or go to hatchworks.com backslash AI dash opportunity dash finder. That's when things start falling apart. I think, again, going back to this phrase about falling into the 11th problem, it's another way to think about is that the more diligent and disciplined we are in the earlier part of that product journey or discovery, if you will, the better it'll serve us in terms of what we're going to solve for and how we're going to solve for it. It's a bit like pay me now, pay me later. And so you want to just do that upfront due diligence and really take your time, have those conversations.

7:19And then as you feel really start feeling good about it. Now that's not, I'm not suggesting that you have to really check all the boxes. This is the opposite of that is like this analysis paralysis. I'm just in constant state of analyzing, researching the market that solutioning will keep getting punted. And guess what? The problem doesn't exist anymore or somebody else has done some problem like you. So there's the time to market, the timing to market are equally important. So if you bring all that together, there is a way to thread a needle and orchestrate those activities along your product journey so that you make sure that you're optimizing for understanding, assessing, timing, solutioning, and testing and launching.

8:10Yeah, and I want to get into some of that here in a second. You mentioned like the product discovery side. I want to get into the piece where we're bringing it into the organization in a minute as well. But before we go there, I'm curious, do you have any kind of litmus test you use right now? I'm thinking when some solutions are pitched or provided with generative AI, I find myself like, okay, let me do it the more traditional way. And sometimes it takes longer. It's about the same with generative AI. And I think change is hard, especially with something new. Like it needs to be markedly better in a lot of ways for people to actually adopt it, which ChatGPT at the foundation level.

8:51Yeah. Yes, it's doing that. But when you get into more point solutions and process enhancements and things like that, but any kind of like litmus test that you've seen past that one that's been interesting. I'm going to take a stab at that, but I think a couple of different things, a couple of different ways to think about it. One is that there's this growing concern amongst academia, like everything from K-12 to college, right? Like how where students are using ChatGPT or Generative VI to like blatantly write essays or do their homework or things like that. And that's where I feel like, yes, Generative VI can solve those problems.

9:38that I think we need to be very careful introducing those tools and techniques to that audience, which is still a foundational stage of their learning, if you will. So for students, and I have a couple of students, one in college, one in high school, I feel like it's important for them to really put pen to paper and really push the mental faculties and do things the old way, the hard way, right? So that you are getting the different, understanding different pieces and parts and pieces of how this puzzle comes together. Now, once you start, once you have satisfied and met those competencies, if you will, depending what grade you are in K-12 or in college, then yes, in order to accelerate and amplify your outcomes, certainly go and arm yourself with those tools.

10:40So I think that's one thing that we need to be careful about. And it's a way for me to, what do you think about it? Is this, is generate something that will improve the outcomes of this audience, right? I think you need to think about that. And the audience comes in different shapes and sizes. Now, so on the flip side, if I go to, I start thinking about the baby boomers. I start thinking about, I'm a bit, I'm not sure if I'm a baby boomer yet or not. But anyway, but 50 plus, he started getting that audience. These are folks who have lived through a number of life experiences. And now they're a point where their time is super stretched.

11:27and how do I help this audience maximize and get more out of their investments and time. So whether that's they're looking for travel, itinerary suggestions, whether they're looking for something related to healthcare or something related to their financial retirement planning. And I think JerryVai can be huge over there. And those are the kind of problems that really time for plucking and solving. So that's two bookends, if you will, on the spectrum of the audiences, and then there's everything in between. So I think everything starts with, just my last point on this one is, everything starts with understanding who is the audience, how mature are they, and what they need, and then you solve.

12:19And then you see whether Genova at and is worth considering to solve this or should it be done a different way? Yeah, no, I love that. And just for the audience to unpack a bit, the one thing that I really keyed on is how is it impacting the outcome? Like the example you gave with the student, yes, they're aceing their test, they're getting all their homework done, but you're throwing the baby out with the bathwater and they haven't learned anything in the process. So it's negatively impacted the goal or the outcome they're trying to succeed with. I don't know if you saw some of the ChatGPT 4.0 demos, which that'll be old news by the time this airs because everything's changing daily.

13:01But the Khan Academy one, that was interesting because it was the AI helping the student along. It was more of a kind of teaching mode versus just blatantly giving them the answers. I think that was a super kind of interesting area there. But if we look at bringing Gen.AI into the organization, large or small, you work with a lot of clients that are trying to do this now. We do as well at Hatchworks. Where do you start? Let's frame it up. You're brought in by CIOs, chief product officer. And they're like, okay, Gen.AI. And a lot of the times they just want to get it in there. They want to sprinkle on the Gen.AI.

13:46But where do you suggest organizations start to actually get real value out of it? Yeah, that's a good question. Excellent question. And I think that question is prime and pertinent for any disruptive tech that's coming along the way to a CI. Hey, how do I start? Where do I start? and I've been having some conversations with some leaders on the federal side, some on the commercial side, some even on startups. I think you need to understand the stakes that your organization, that your mission is all about, right? A lot of setups are really in the stakes app, right? Federal agencies and those leaders they are highly regulated sort of domains the stakes are super high you're there to serve the citizens you have to be very careful what you're bringing what you're offering right on the flip side startups a little while a bit like wild west yeah they can do some experiments so they can go to market real fast and fail fast and get the hands burned and maybe they're okay with that And then there are use cases in between.

15:05I think the key is first thing is assess your context and the space in which you operate, what your customers are, what your stakeholders are, what kind of regulatory and compliance needs are applied to your context. And number two is then always take baby steps. We talked about this classic crawl before you walk, before you run. So start by crawling. And the crawling can mean, for example, a highly regulated setup. Let me start with that. Let's talk about agency housing and urban development, right? I'm sitting on a lot of regulatory policies and those kind of documents. It runs a lot of programs for a variety of the citizen population out there.

16:00and there's a lot of text content that's sitting within its enterprise and a lot of the content is also public facing. If HUD wanted to take a small step and do a small experiment, say, hey, how do I demonstrate to my stakeholders that there's value in bringing Genitive AI into my context? I would look at, can we make it easy for my internal stakeholders, the HUD employees and partners and contractors to be able to search through these policy documents, through these regulation documents? Because I would venture, I would argue that there are hundreds and thousands of documents and pages that, right?

16:48We've been seeing, especially with Gemini coming out and now with 4.0, that the context window and the token size, that can use to blow up. I think we're talking about a million tokens and maybe trending 10 million. And so we could be feeding the typical model hundreds and thousands pages of documentation. It doesn't have a single doc. We can stitch all these things together. There's ways to orchestrate that. But now imagine somebody at HUD, it could be a policy analyst, it could be a pricing analyst, whatever. They could literally have a conversation with these hundreds and thousands of pages of policy documents that have been collected and aggregated over like the 10, 15, 20 years.

17:39and ask questions like, hey, what were the trends in Loudoun County in 2018 with regards to the housing benefits and funds that were given to the, I know, some community could be the Black and Latino, could be any of them. So imagine you can have that conversation with these documents, which today in the absence of generative AI would be really tedious and laborious. You have to really go to a subject matter expert who knows these documents intimately and then really rely on them and in a new block. But now we're doing something like this, which is a fairly low stakes experiment. It does a few different things.

18:28One, this is not very difficult to set up, right? Because your document's right there on your conference pages in your S3 buckets. It's very easy to wrap it with a chat GP like interface. Number two, when you build this, you can do a small pilot and just expose it to a small audience within the Hutt enterprise. So just in case something was to go south, you're limiting the blast radius. It's within that small context. And then you can gradually ramp it up to the bigger audience. Number three is it also allows you to demonstrate value and benefits and the outcomes that we were talking about earlier.

19:13Is this really compelling? Allows you to demonstrate all these things and more to your internal leaders who are maybe suspect about the benefits and merits. And because there's a lot of concern with regards to security and IP and iData getting exposed to folks who are not authorized for that. Yeah, you almost got to get some points on the board to get buy-in from others. But just to break it down, I love how you talked about it. And for listeners, I would almost visualize, and we use this a lot when we're doing a workshop, just take a simple matrix, right? Right. Value low to high. And then the ease of implementation, where you look at it on a risk vector as well and just start placing use cases there.

20:04And then you can find those that may be higher value, lower risk. That could be one way of looking at it or how easy is it to implement or how easy is it to access the data? But I love how you talked about it, limiting the blast radius. So if this thing goes sideways and starts fighting with itself, what's going to be the potential impact? I saw recently Morgan Stanley, they did a similar kind of use case where they provided this to their financial advisors. Right. So it took all the different reports and documents and things that they would use and spend hours weeding through to provide insights to their clients.

20:42And they were using generative AI to easily access and query their own data. And if you think about it, it's their internal people using it. So you have that barrier between the end user. So if something does go weird, they're there. So that was another kind of interesting use case there as well. I think the, where this is prime when it comes to leveraging and doing some experiment is where you have a lot of text context, content city. and that's what we call unstructure, right? That's words and paragraphs and essays and policy docs. But there's a way to also approach structured data and we can talk about that as well.

21:26But yeah, it's really, because that is the most tedious, right? As we as humans, as much as I enjoyed reading when I was growing up, I'm very selective about reading. I enjoy conversations. I enjoy watching. And reading has become a bit tedious, right? and there are a fair amount and number of adults who still enjoy winning, including my wife, but I'd rather have a conversation and watch something. So for me, if somebody was to come to me and say, Uday, you don't have to read all these security documents or these policy documents or you could just query and this is just like on your pace, when you need, what you need and you get that information right away, that'd be super compelling.

22:08That would make my life so much easier and then also outcome that I'm trying to drive would be so much more compelling. And I think one of the beautiful things of generative AI too, that some may say it's under valued or considered is just the pattern matching ability of generative AI. So it's not just that, okay, I have hundreds of documents and it can go find the right answer. It's matching patterns within the data set or what it's been trained on. And that really starts to become dynamic. It almost begins like it's mimicking logic in a sense, which you could argue that maybe that's what our whole just logic and brains, how we operate is based on anyways, but it starts to provide a lot of exponential value in that way.

22:52Yeah. One quick thing that just came to my mind as you were talking about logic and reasoning is, I remember you were calling for my Capital One days, one thing that I worked with my teams and for my customers up there was building these decision engines for credit underwriting and fraud detection. And most of credit underwriting and fraud detection back in the days had been rules-based and still continues to be rules-based. Imagine you've got a spreadsheet with rules learning, which is, hey, FICO score is this, approve if not, decline, right? And so on and so forth. During my time, we started experimenting and brought to market at AI models, both on the fraud side and as well the credit side.

23:42And then we started orchestrating your traditional decision engines, which were rules-based along with AI models. So there was a whole set of elevated and enhanced benefits we started seeing when we started running these decision engines in two parts. One, as a transaction comes in, it's hitting the rules spreadsheet and getting a set of decisions over there. And then it goes to the model and gets a decision or a score. And then we arbitrate to that and get a final decision. What's happening with your traditional credit and fraud models, which was based on training structured data. there was a lot of effort that was invested in labeling the data and fee changing and so on and so forth which required a lot of logic and reasoning what journeyvice is doing and is about to disrupt is that we don't there might be cases and i'm sure somebody's already solving this is where you don't have to really go and curate your data sets or your transaction in that way which was very expensive, right?

24:54Labeling data and feature engineering was very expensive. Now imagine I could just take transaction objects, right? That transaction object has the user's information, where they're purchasing, what they're buying, and also maybe their transaction history and just let a generative AI, a fine-tuned generative model, lose on that. I'm pretty sure I can solve that fraud problem or credit problem fairly easily. It may not be as precise and on point as an MLM model, but I believe that with an NLP model, which are traditionally MLM models, I can get very close to that. So I think that's where it's going to start lowering the barriers in terms of these use cases that were in the past considered very sophisticated and had to be done by scientists.

25:45scientists and that really specialized staff, I think that barrier is going to start getting lowered as we move forward. Yeah, that is a major change in how we operate. And you mentioned something around having to have these rules-based engines. I look at the whole world of business process, automation and optimization. Those are very rules-based systems. And you had to account for every edge case, which is very difficult to do. And now you don't necessarily need to do that. I think that's a huge shift at play as well. But on the data side, I think if you think, all right, we're bringing AI into the industry, focus on the problem, think through some of the things we talked about in terms of the value and the risk and all those different things.

26:33But in this new age, data is going to be one of your best mode builders and differentiators potentially because chadgbt may be trained on the entire corpus of the public internet but it's not unless there's some sketchy stuff they're doing it's not trained on your private data and that's proprietary to you as a company or even an individual if you think about it in some ways but that's proprietary to you and then you can unleash these generative a models and approaches in conjunction with that proprietary data. So that's another big consideration when you're thinking through use cases is how can we leverage our proprietary data in just interesting and novel ways, right?

27:19Yeah, yeah. The one thing I want to really shine the light on is when it comes to the health and hygiene of the data, that is the case. It doesn't matter what technique, which tool you're using. your data is still needs to be treated as the first class citizen right the moment you drop the ball on the data is when you have no clue you have no faith no trust on what the outcomes are going to be regardless of what tool you're bringing in especially when it comes to generative i and like you said these models were trained on large corpus of data then if i'm just saying if i get that in it i just want to train this or fine tune this on my data that i've conference on you still need to make sure that it doesn't have any privileged information, doesn't have any personal information, doesn't have any of that stuff, right?

28:09You still need to do those scans and security checks and have those sort of controls and checks and balances. You still need to have humans in the loop to make sure that you can trust that data because you're about to unleash an LLM, which is just going to crawl everything. Right. And then on the flip side, think about all the right guardrails you need to put in place. Right. To make sure that even if the model is picking up certain things that it wasn't supposed to, that you have a control right before the response is put out. we're just checking through these things and catching them and flagging them if not blocking them from getting to the enemy and getting into especially in in the hands of bad actors which could do anything malicious with your data yeah and that's the guardrails piece is even more important when you have a system that can use logic and reasoning to do whatever it may be so that becomes a huge part of it so i want to shift to the strategy side of it in a book i love seven Powers by Hamilton Helbert talks through the different powers you can leverage for mobility and differentiation.

29:26I want to go through them and think of how are they applying in this age of generative AI? Do some become more important or less important? But if we're going through those, so just for reference, they are scale economies, network economies, which is like network effects, counter-positioning, switching costs, branding, ordered resources, and process power. But if we start with scale economies, so this is like the cost advantage you have when you start to have scale. And I don't know about you, but the one thing that comes to mind is just the foundation models. They're starting to pull way ahead and they're starting to gain scale.

30:04I don't know what it says, this effect at play where it's compounding in nature, right? So you have these foundation models taking place, but anything interesting on the scale economy component that comes to mind for you in this generative AI age we're living in right now? Yeah, I think this depends on where you're coming from and looking at this, right? So are you coming in as something like a chip manufacturer, right? Like an NVIDIA or AMD or Intel? Are you coming in as a consumer services or products like an app or Facebook or something like that. It just depends. I think the scale economy or the economies of scale will really, I think, grease the wheels in terms of the cost.

30:59We've seen the cost drop dramatically because there's more and more horsepower that's being generated. The chips are getting smarter and better and more efficient. So the unit cost of producing the same outcome is dropping dramatically, right? We always want to see the context window size improve or increase dramatically, right? The Moore's Law is just going out of the window. We are seeing this thing double every two months, forget 18 months. Yeah, there's almost this new law at play with AI. I've heard somebody referencing a few was, they said, this is like the fastest accelerating depreciating asset we've ever seen.

31:48When you think of how fast it's accelerating from the capability and then the reduction in cost that's happening over time. A hundred percent. And the thing is, you mentioned a few different concepts over there. Scale economy. You mentioned network effect. Yeah. There was another thing. But I think they all sort of play in tandem, right? There's a very interesting interplay between all of these concepts. And it's like you're trying to go for this flywheel effect, right? Exactly. At some point, you get to a point where the engine is just like feeding itself. And network effect, for example. I strongly believe that as companies, whether you're in the federal business organization, whether you're in the federal business or you're a commercial startup, as you build more meaningful use cases, more compelling use cases, people will just, they will latch on to it because the biggest sort of, the most expensive sort of asset that we all have, that cannot be recovered is time.

33:04And if you can save time, then that is a huge thing. My favorite example is like travel. I'll go back and travel as my favorite example. I and my family, we love to travel. We spend a bunch of time when we're trying to think of trips. We're about to head out towards North Carolina, drop my daughter off for a pre-college program. And then we're going to head out towards Asheville and Nashville. Those are two new towns I haven't been to. And I've already spent like almost eight hours, maybe 10 hours researching Airbnb. And what are the stops on the way? And where should I stay? And where can I? And imagine if there was a solution today that can help solve that problem.