Solving the AI Buy or Build Dilemma for Tech Leaders with Sergii Gorpynich, CTO at Star

Inviting technology solutions stakeholders to experiment as much as possible. Make mistakes, experiment, make some quick integrations, some quick prototypes. It certainly makes sense. But then again, if and when you are at scale, do very systemic consideration of this Build versus Buy. Hello everyone. Welcome to another limited podcast series from a STAR on Artificial Intelligence. My name is Cherry, am the head of communications here. Today we are diving into a topic that is on the mind of many C suite leaders and that is around the critical buy versus build dilemma on AI solutions. With this fast evolving technology, should business leaders develop AI solutions in house or should they leverage existing market offerings? I'm very excited to unpack all this with star's Chief Technology Officer and co founder Sergey Kripinich. Sergey, thank you for your time. Please give us a quick introduction about yourself. My name is Sergei Gurpinish, I'm a Chief Technology Officer at STAR and part of founding team here. One of my key and probably the most exciting responsibilities to set up and develop technology practices at star. At this moment, obviously the strong focus of our team is the AI capabilities and the AI technologies. So before we begin, if you can quickly explain for those who don't know what buy versus Build dilemma is and why is it such an important business topic? So in the modern technology landscape, buy versus build is very important question which all of technology leaders are asking, right? When they take responsibility over providing equipping their business with the technology solutions. And the landscape is usually very broad. So on one side of this landscape, on one side of this spectrum is 100% buy solutions, right? Oftentimes on the other side of the spectrum might be something like a hundred percent build solutions, so many kind of variants in between. So what might be the optimal choice? Not so easy question to answer, right? And there are a number of factors which should be taken into account when making these buy versus build decisions. But again, it's a complex question to answer, but then also a critically important question to answer for technology leaders. So let's jump right into it then explain this to us. Explain the buy versus Build dilemma or decision making in the context of AI. Why is it important today? What are the implications from a both business perspective and maybe end user? So explain the whole landscape to us please. So speaking specifically about AI, right? And again we take a horizontal point of view, so to speak, in the sense that we are looking across industries. So each industry have got their own specifics, but looking across industries. So basically there are several main choices within AI. So first choice, I would call it Big three and here I point to the Leading loss source, so to speak solution providers. And here I talk about OpenAI with their GPT line of models and line of solutions. And then I talk about Google with their Gemini Power solutions. And I would also add here Anthropic with their cloud family of models and family of solutions. So these are like big three big and their solutions are extremely powerful, they're also extremely general and then they're also quite expensive. Obviously the price at the cost of using the solution, the cost of running the solution is very important. So very powerful, very general, but then also quite expensive. So the solutions built on top of one of these big three solution providers is one side of the spectrum here. And I would call a choice of one of these three types of solution a buy type of choice. But then also what we see on the other side of the spectrum is a very significant development, very active development of open source based solutions. And here companies like Matter take the lead with their Llama family of models. Like very recently Llama 3 family of models have been released. They have released a llama 8B model. Meta have also released llama 70B model. Those are very powerful and those are kind of leading the overall movement of open source solutions. There's another provider company named Mistral and they open source, they release very interesting, very powerful models as well. And then there's plenty of choice in this direction. And these open source models, they're very powerful as well and they are certainly a good candidate to consider. So when I talk about solutions based on open source models, I consider this as a build type of solution ultimately. And then again there are many, many opportunities in between. And those opportunities in between might be based on further fine tuning open source models. So I would say when we talk about applied AI, AI solutions for specific businesses, I would not consider building neural nets, building language models and video models from scratch. That continues to be a very expensive undertaking. Even if we talk about so called small language models, they're still quite expensive to build and quite difficult to build. But on the other side, the option of fine tuning open source model, that option got significantly democratized during even last year. That option becomes available for a broad category of engineering teams. That is very interesting option to consider. Okay, so what we've got kind of again as a complete summary, so we've got big three, three closed source solution providers, that's Gemini GPT and then Cloud from Anthropic. Then we've got a family of open source models and companies like Meta and Mistral. But then again Google with their GEMMA models and Microsoft have been actively releasing into open source as well. So they take the lead and then they provide very interesting foundation of these open source models which are available to just be integrated into applied AI solution. And then we've got this category of and this possibility of fine tuning, optimizing open source models which again provides a very interesting option as well. So plenty of options are available across this build versus buy spectrum. And you mentioned when you were talking about the big three, that is very powerful and also very general. Explain what do you mean by general? Well, obviously as we understand they are trained on whatever knowledge is available for training. We understand that especially recently, getting this knowledge available for training becomes a little bit more difficult for companies creating these models. But still they get access to overall Internet knowledge. And because of this, these models indeed provide a very general and very specific knowledge on very different topics. On the other side, when we talk about building a specialized optimized applied AI solutions for specific business needs, in so many cases, I would even say in majority of cases that general type of knowledge is not required. Say if you are building a ticket processing chatbot for IT service management solution, of course you do not need your chatbot to be able to answer questions about who has been the first President of United States of America. On the other side, you do want your chatbot to have a very specialized knowledge, a very specific answers to very specific questions. Then basically this leads us into a cost benefit type of question and type of analysis where you might realize that when you are utilizing this large scale model like GPT4 GPT4O or Gemini or Clouder, you do have access to this extremely broad knowledge. But at the same time, this access comes with the price. At the same time, this super general model does not necessarily provide you with a very specific functionality which you are looking for. It's so interesting when you explain this and it makes so much sense. And of course there's still that narrative, at least in the business environment and also in the press, that we are running out of data to train these models. So it's almost bigger is better. So if I understood it correctly, in a business context, AI solutions are more powerful if it's trained based on the company's data and needs. Did I understand that correctly? Right. For businesses, this is critically important that whatever AI solutions they build, those AI solutions are specialized and those AI solutions should be aware of the business context. So these AI solutions should be specialized towards the industry where the business operates. I would call it first level of specialization. But then second level of specialization is a Specific business context, a context of my specific company. So this is second level and third level of specialization might be optimizing these solutions towards the needs of my specific end user. So this is like a multi level specialization and large language models coming from this big three providers, they do have possibilities to specialize specific technologies like retrial augmented generation which enable this specialization. But again that type of specialization comes with the price. So you still are making call to this huge model and the so called inference cost is very high. So the actual cost of running that inference is very high. And this leads to a high price of using this general model. Again we get into the need, into the desire to consider some alternative options. A very specific and maybe simplistic example is tasks like, let's say document summarization. Obviously these huge models like Gemini and GPT4, they are great. At the same time, do you really need to make inference through that huge model? For such a simple task, you might decide to use a small language model which is available as open source, or you might decide to use maybe some, I would call it classical NLP algorithm which does not even require a call to a language model or a call to a neural. This case, this comparatively simple task of document summarization might be solved with a very simple instrument where the cost of the summarization is close to zero. So basically I would say that in AI as well as in other domains, the aspect of cost versus benefit is a critical aspect to make a model choice, right? We have been discussing one of the architectures for AI solution with starting just basically a few minutes ago before our conversation, and we looked heavily at the cost and pricing and related benefits of different types of solutions. So this cost versus benefit is extremely important factor. But then when we talk about the cost of the solution, we mean the total cost of ownership, right? So this is not just an inference cost which I'm mentioning right now, but then also cost of acquisition of the solution, cost of integration or cost of development. And then this also includes cost of operations, including inference cost. And then also ultimately includes a cost of transitioning of my current solution into upgraded solution. Right? So this idea and this notion of total cost of ownership is critically important to consider. Here you touch upon, based on my understanding, so many key elements. Let's unpack one by one more in details. Just on the infrastructure side around AI solutions, we can look at it from a sector perspective or from a client perspective, but talk us through the importance of a business existing infrastructure and capability to onboard AI solutions. But also talk about the Overall sector. Because it also sounds like the more business embraces technology, the faster it will also develop and hopefully become more democratized like you said. So there seems to be a, almost like a chicken and egg, there is like a interconnected dynamics right there. But of course you know, we service our clients and maybe we start with the client perspective. How should they even look at the infrastructure part? Because many clients sounds like this is new. They never really had to till recently build the infrastructure. Even have to consider this AI solutions integration as we talk to our clients, both in the startup domain and in the enterprise domain. The infrastructure aspect and the model hosting aspect more specifically is very important consideration as well. Critically important consideration generally. I think I mentioned that when we are considering buy versus build, it's not just about cost benefit, it's not just about total cost of ownership academically if you will. I would name maybe 10 to 12 critically important aspects to look into and calculate if you will to make a decision on when you are making this overall decision of buy versus build. But then infrastructure and then hosting is one of these critically important aspects. Just as some specific examples. Well generally when we talk about infrastructure and hosting again we talk about the cost of running solution, but then we talk about data privacy and data security as well. Many businesses are not very willing to channel their data and their customers data to some other data centers, right? And here when we talk about buy oftentimes we talk about transferring some elements of the enterprise customers data to some third party data centers. Many businesses are not willing to do that. But some other businesses are not able to do that. Many businesses actually they prefer and that's part of their policy to host and treat and process all of their customers data in their own data center. So that's one important situation which we happen to see oftentimes. But then some other businesses, their data centers are actually airlocked, which means that their data centers are simply not connected into open Internet if you will. And from that perspective, in those situations, those businesses and the underlying technology solutions, they're not even able to make calls to services providers like OpenAI or Google with Gemini or Anthropic. They just simply physically do not have that possibility. Does this mean that AI is not available for those types of solutions? No, this just means that whatever AI solutions we built for them, right, it has to be hosted in their own data centers. These are some examples of considerations about infrastructure. Again as another comment here, obviously if the business preference is to keep all of their data in their own data center, including processing of the data Then the option for these businesses is to host AI solutions in their own data center as well. And then again we talk more about build type of a solution for its businesses. But again you are not building from scratch, right? So still you might want to take some models which are already available in this open source format. You put those models in your own data centers and you either use these models or you optimize these models further and model optimizations technologies which are heavily used nowadays. First of all, this is a retrieval augmented generation, the one which I have mentioned. And then the second option is fine tuning the model. So if I'm understanding you correctly, then the overall cost benefit that you were talking about, that's the overall cost ownership, it's also tied to the client or the business existing business model. We can look at it financially, but we can also look at it as their existing and maybe evolving business model. So then my question is this technology is moving so fast, I'm assuming many business model will change because the technology is also changing their end users behaviors. So is it safe to say, and correct me if I'm wrong, in the short term the cost benefit is the right way to assess whether it's buy or build or a hybrid. But in the long run the competitive edge does come from having your own AI solutions. Is that a right? Yeah, I think this is a good reason. So again, we are very fortunate to talk to many clients in different industries. Many of them at this point experiment with the AI capabilities. They're searching and they're finding specific use cases within their current business model and they implement those use cases. So they bring this extra kind of additional functionality to their end users and that's their way to embrace AI. Some other clients though, they embrace the paradigm which we call AI first solution. And you might compare this to a situation maybe around 15 years ago where mobile front ends and mobile phones became so powerful with the introduction of iPhone and Android. And then many technology driven businesses realized at that point that their solutions need to become mobile first. That the key entrance point into their technology solutions, the key endpoint, the key user interface, is actually mobile phone. So I think that the whole industry is now beginning to realize that the future for them and the future for their technology solutions is really AI first. And this basically means that the mode of interaction, that the nature of interaction between these technology solutions and end users will be through AI powered interfaces, conversational bots, text chat bots, et cetera, et cetera. So again, some of our clients and the increasing number of our clients, they realize that they need to think forward. They realize that their end game might be AI first type of solution. And this is extremely interesting situation and extremely interesting challenge for our teams to embrace as well to find answer to this transformation from a current state where maybe some specific disconnected use cases are AI powered transitioning into AI first type of solution where the whole technology solution is really powered by AI and the whole user experience is transformed into being guided and being driven by some type of underlying AI interaction. So this transformation is really powerful. And then this AI first paradigm of user interaction, it drives the underlying technology, the underlying architecture of the solution. Right. Because with this AI first paradigm of user interaction, you understand that you need to quite completely change the underlining architecture of your solution as well. Zurgy, I think you just brought up a really important point in this ever ending digital transformation and I would love for you to just further explain that Almost using the end user's AI driven solution to also drive business transformation, that means different departments. You can have the marketing, the customer facing department, then you have the IT team, the tech team and the finance team within a business who may not always talk. We know the cross functional silo within a business transformation. So you're saying with AI transformation that connect tissue has to happen because AI is fundamentally going to change the business from both employee the future of work perspective and also customer experience and the overall business model. It's not going to be the delta that brings all this together. Is that correct? Yes, indeed. So basically when you are transforming your underlining technology technology solution into AI first solution, ultimately this might lead to a transformation of your business model as well. We are talking about you as a, as a business owner. That's a possibility in many cases as well. So basically this is not just optimizing the functionality for end users. In many cases it's about upgrading, changing the business model quite completely as well. I have a few more questions, but let's address one really critical point here. Talk us through talent. We're still going through a. I almost feel like the first wave of digital transformation and a lot of employees feel left behind. You see business laying off a big chunk of their workforce because of AI. Talk us through the talent side, the organizational competency. How can business leaders continue to develop talent or maybe they outsource talent in this really fast moving technology landscape? This is extremely interesting question also because we at STAR, as a business or organization, we are undergoing the transformation which is very similar to the type of transformation which our client businesses are going through. Like specifically at sar, maybe I will Use our own example, at sar, we are very much a talent driven business. So our talent is our key asset, so to speak. And our talent enable the services which we provide for our clients. Right? And then right now we actually understand that the services which we are providing for our clients and this is really ideation, design, development, deployment and maintenance of technology solution digital solutions. So our services are going through the transformation which is very similar to our client businesses. And we understand that increasingly our own talent, our own teams become powered by AI solutions across different components of our services. When we are looking into strategy, part of our business strategy consulting or user experience development, we are now embracing a set of AI platforms which really amplify our capabilities. Again here we think about this and we see that this is really amplification. This is not a replacement of human talent and human effort. This really amplification of human effort. And the same transformation is happening within a software development domain where we use a set of copilots which really amplify the speed, improve the quality of development. Similar transformation is happening across QA quality assurance aspect where we increasingly combine the human talent with a certain AI platform capabilities, right? To amplify the value, to improve the speed, to compress time to market for our clients for their digital solution. So these transformations are happening really fast. Again, they're very significant. But what I do want to underline, to answer your question more specifically, here we talk more about collaboration between human teams and then AI platform. And here we talk about amplification of the effort which our human expert teams are providing. So it's exactly what they say, AI is not taking your job, but those who know how to use AI, right. It's not about the replacement, but it's more about amplification. But then also we acknowledge that the nature of our business is very high quality and very sophisticated services where students are playing a leading role and the type of effort which we provide, which we apply, helping our brands to build their digital solutions, this is a very high added value effort. It's very much about innovation, it's very much about invention. So this is a very unique effort, but indeed AI type of solution help us to amplify that effort, speed it up, provide a higher quality solution, provide for better product, market fit, et cetera, et cetera. Do you want clients or business leaders to also take a similar view when they, when they look at their internal competencies at least? Yeah, the right view, the right attitude to the problem, if you will. Right. Personally, I do like to compare the advent of the recent AI Capabilities. I do like to compare it from the impact perspective and the level of impact perspective. I do compare it to invention of electricity all the way back to 200 whatever years ago. At that point, at the very early stages of that technology, many people did not realize the possible impact. Okay, a bulb, it lights when I switch it on, but what's next, Right? So for many people it was not clear at that time. Right now we understand how transformative the impact electricity has brought to the whole industry and to the whole society, if you will. AI is bringing a similar changes, but at the same time, in the same way, electricity has not replaced humans, it just changes their ways of working in the same way. AI will not replace humans either, but it will significantly change the ways of working, the focus, the required skill sets, et cetera, et cetera. All of that will be transformed over the next five to 10 years, even during the short span. But certainly humans will remain in their leading role rather than robots. Let's jump into recommendations then. So for business leaders who are going through this buy versus build dilemma, your perspective, what's your recommendations or approach or process to help them make the right decision at the right stage of their business? Well, ultimately this has to be a well calculated decision, right? They mentioned there are 10 to 12 factors to be taken into account. All of these factors, the answer which is behind that factor, it might be quantified, it should be quantified. So all of these aspects have to be carefully explored. And this obviously includes cost benefit analysis, including total cost of ownership factor. It includes infrastructure aspect, it certainly includes data privacy and data security. It also includes specificity of the solution, right? And I've spoken about three levels of specificity. Industry specific business and then specific end user, who's the customer of that business. So that specificity factor has to be taken into account, et cetera, et cetera. So make that analysis across all of the factors, taking endgame into account. Where I want my technology solution to be in, let's say one year or three years, right? And then based on that type of analysis, based on those answers, carefully crafting the solution. In many cases this is not a pure buy or pure build, it will be somewhere in between. And again, the tool chain, so to speak, which is available enables very different kind of flavors of that solution, directly speaking in those situations, because one of the factors is actually available, team, expert, team to make that type of analysis and that type of decision. So in many of those situations, it makes sense for business owners to just get consulting and get help from specialized organizations. Who might help them to make this buy versus build solutions. This actually might be store's own team and we've got a team and initiative which we call AI Innovation hub. And this team is making this buy versus build type of solutions almost on a daily basis for different clients. Right. So this might be store own team or some other specialized consulting teams which are helping businesses to make this type of choice. That said I again I think about some specific scenarios like we as show we we actually work a lot with the startup enterprises. So for startups, especially for early stage startups, my recommendation would be they need to iterate extremely fast. So for these types of early stage flashlink businesses it might make sense to build their quick prototype based on why rather than build because build means certain investment for them because they need to validate their business idea quickly. They need to provide for this fast product market fit validation for those situations. Almost obvious choice is to integrate their end to end olution based on big three. Those big three which I mentioned especially as all of OpenAI, Google Anthropic, they provide for for this kind of toolchain for quick customization. But again that comes with the price. That price is okay when you need to make this quick product market fit validation. So quick prototyping this will probably be a buy type of a solution. But then if you are scaling or if you are already at scale, then you engage into proper analysis across those 10 to 12 factors to make a decision. The three specifications industry, business and end users. When a leader is assessing each of them, do they come in sequential order or is interconnected in parallel? Okay, understood. So the conclusion and the learning is do not blindly apply AI. Really look at your organization and really understand what you're trying to achieve. But also experiment early on. Right? I would rephrase it into inviting technology solutions stakeholders to experiment as much as possible. Make mistakes, experiment, make some quick integrations, some quick prototypes. It certainly makes sense. But then again if and when you are at scale do very systemic consideration of this build versus buy. By the way, when I'm inviting for experimentation, it's also important to understand what is actually possible. Again, what we find out at strong in our interactions with our clients. Well as oftentimes actually it's less important to understand the specifics of technology and specifics of implementation. It's way more important to understand the overall realm of possibilities. And actually within our own AI team, within our AI Innovation hub, we put a lot of focus on product management and user experience design expertise in AI. Because product managers, user experience designers business strategists, they need to understand what is possible in principle, and they think in terms of use cases. And that's a very important knowledge, actually critically important knowledge to build this AI. First solutions, as I mentioned earlier, but then experimentation. Go quickly, make some integrations, try some use cases. This is certainly welcome. That's the way to learn this realm of possibilities, which is expanding, by the way, on a daily basis. Yeah. The choice you make today definitely sounds like it will have significant impact on the business down the line. Is there anything that I didn't cover? Serhi, this has been a very, very interesting conversation. Jerry, thank you so much for asking these questions. Again, what I do want to underline, even right today when we talk about digital solutions, we already talk about AI powered solution. So AI is not coming. It's already here. So the electricity has already been invented and now it's a question of how to use these capabilities effectively for the business. For end users, it's already here. And the importance of these capabilities will just grow in the future. Not even on a yearly basis or not even on a monthly basis, on a weekly basis. So it's obvious that integrating these AI capabilities is certainly a priority.