Episode 10: Building Business Cases for AI

Maciej Kłodaś (MK): Hello, my name is Maciej. I’m the leader of Analytics Experience Competency Group at C&F and this is C&F Talks, a place where experts and professionals discuss ideas, challenges and observations from the perspective of an IT vendor. My guest today is Marcin. Hello, Marcin.  

Marcin Ludzia (ML): Hello, Maciek.  

MK: Marcin is the head of AI Competency Group at C&F and today’s topic is one of my favorites. I’ve been waiting for this topic quite a bit and we’ll be talking about the implementation of AI in different industries.  

ML: The topic that everyone is talking about currently.  

MK: This is the hype nowadays. And when we were at Gartner’s conference, Data Analytics Summit in March in Orlando, about 90% of all topics discussed during this conference were related with AI. Everybody’s talking about AI nowadays. 

And there are a lot of expressions and names and abbreviations and I don’t really understand them all, frankly speaking. But I remember that you’ve been writing an article about AI lingo. Is that right?  

ML: Right, there is an article published on our website where I try to explain in simple words a number of terms related to artificial intelligence to help managers better navigate this area of technology and understand what to expect and how to discuss with technical teams. 

MK: Okay. So tell me, I think everybody tends to think they know what AI is, but I’m pretty sure that reality is a little more complex. Can you explain what AI is for dummies?  

ML: This might be a little bit of a tricky question. So there is a very simple definition of artificial intelligence, which says that artificial intelligence system is a kind of a system that behaves in a way that we would expect from an intelligent being, similar to intelligence as possessed by humans. And this definition is very simple. 

However, it is not very sharp and doesn’t precisely define what artificial intelligence is. We can go into finer details. And when I was working on my PhD, I even went to psychology books to try to understand the intelligence itself, what is intelligence, especially in humans. 

Psychologists, they don’t have like a single definition as well. But you can generalize all those research and definitions to several aspects. The first aspect is the aspect of learning. You can say that an intelligent person, intelligent creature is a creature that can learn based on some input, on some data, on some things that it perceives.  

The second aspect is adaptability to environment and especially to changes in the environment. And if you want to transfer those things into the field of AI systems, then we can say that artificial intelligence systems are those that can be trained and can learn from data, from the examples of behaviors that we provide to this kind of system. 

And based on those examples, the system learns how to behave, how to act. And also this artificial intelligence system can adapt to changes in the environment that it operates in. So those two aspects, I would say that are most important to differentiate AI systems, from regular software. 

So, in regular software we have to define exactly step by step how it should behave. Artificial intelligence systems are not explicitly implementing the behavior, but instead we are implementing an algorithm that can learn from the examples provided to the system. So instead of explicitly saying how to behave, we provide examples of how to behave. 

And based on those examples, the system learns how exactly it should behave, how to make decisions, how to take actions.  

MK: Okay, so by this definition, ChatGPT is AI.  

ML: Yes, ChatGPT is an example of AI system. So the system that actually was trained on a vast amount of text data. So the examples of natural language, how it is used in conversations, in articles, in any communication between people. And based on those examples, on this text data, system learns how to respond to user questions. 

MK: Okay. You know, everybody’s scared that AI will take over our planet like Skynet did in the Terminator, right? Because there is something called the Turing test made by a famous mathematician?  

ML: Yes. I would say that ChatGPT as a system or the current generation of large language models would actually pass the Turing test. 

The Turing test is a kind of a test to identify whether or not we already have this artificial intelligence or not. And it was defined by Alan Turing. The test itself is based on the principle that a human communicates with a different kind of agent. 

And if this human cannot differentiate whether they’re talking with an artificial system or real human, then we can say that we achieved the actual artificial intelligence. And by all means, we can say that ChatGPT actually passes the Turing test as it was defined.  

MK: Okay. So the future is here. I’m a bit scared now.  

ML: Now we have another test, that was defined by Steve Wozniak from Apple. Right now we are in an era where we are talking about narrow artificial intelligence. So artificial intelligence systems that can do very simple and, let’s say, narrowly defined tasks. But we are targeting to something that is called artificial general intelligence. So a kind of a system that can do multiple things. 

And a test that was defined by Steve Wozniak is that we’ll have this AGI when this kind of system or robot can go into the kitchen, where this is a new environment for this system, and it can make a coffee. It knows how to orient in the kitchen, how to find a, let’s say, coffee machine and so on and put all the ingredients to make a coffee. That’s the definition. 

MK: So we have a lot of new fun toys right now to play with. From your perspective, because you’re working closely with our clients and you can observe how they are implementing AI, what is the state of AI implementation right now with our clients?  

ML: Artificial intelligence is not very new. The history of artificial intelligence is actually dated back to the 50s. And we had like multiple waves of interest of artificial intelligence and something that we called artificial intelligence winter. Right now, we are living in the era of, I would say, third wave of interest that actually started somewhere around 2012 with the revolution in computer vision. 

And, of course, the last two years is the new era of large language models. The systems that can actually communicate with a human using a natural language and can do different things for us based on language.  

MK: This is the chatbot era. 

ML: This is the chatbot era, exactly. However, this is actually a third generation of chatbots because there were previous generations. But this one brings new capabilities of generating responses. 

The previous generation of chatbot could understand, for example, the user question, extract some intent, but the actual responses of the chatbot were predefined by a developer. Right now, the model itself can generate the answers. And the state, especially with LLMs and generative AI, we are at the moment when we already know the capabilities of those kinds of systems. 

Multiple POCs were implemented within the last two years by companies. They understand capabilities, they understand strengths, and they have a vision of how to use it in real business. And we are at the moment when companies will start building longer-term visions of how to utilize especially generative AI and implement them in day-to-day operations. 

MK: Okay. What kind of challenges do you see right now? Because I’ve heard that our clients are sometimes fed up with building those multiple POCs, searching for different use cases, testing hypotheses. They want to see the bigger picture to start generating value from AI. 

ML: All right, challenges. So when we are talking with our customers, we hear multiple things. Companies did a lot of POCs for last two years of implementing gen AI systems to learn their capabilities, to learn how they can use them in operations. 

But managers are feeling a lot of push for implementing AI in the company operations. And according to some research, 80% of managers actually feel the urgency of using AI systems in their company. 

And they don’t want to do another POC. They don’t want to do another test, another evaluation. They want to have like a broader view of a vision, what they would like to achieve, what would be the business goal of implementing the AI systems, and how the company will actually get a business value out of the AI systems.  

So, they want to move from the POC era to actually implementing production systems that are generating business value, that are optimizing the efficiency of the business processes, enhancing the products that those companies offer for their customers, or lower the cost of operation for those companies. 

MK: Okay. So what I hear is that our clients seek for value or return on investment. They want to implement AI as quickly as possible, but they need to see the balance between the cost of implementation and the value that this AI will bring to the table. So what are the benefits of implementing AI in companies?  

ML: What we hear from our customers, and this is actually the exact phrase that I heard. And it actually really precisely defines what happened in the last two years, that there were a proliferation of chatbots. So for the last two years, a lot of managers and business stakeholders actually perceived AI as just implementing a chatbot. But a chatbot is just one example of an AI system that can be used in a company. And the capabilities for providing business values are much broader.  

The first example of business value is providing additional information. And this information can be used for making better decisions, data-driven decisions. So instead of thinking about, for example, how to plan a production schedule for the next week in my manufacturing site, I can ask a predictive model to provide me estimations of demand. Then based on that, organize a supply chain of raw materials. And based on that, also plan manufacturing plans.  

So this was the first example of providing business value. The second one is automation. And this is something that is on the rise. We are starting the discussion about AI agents built around LLM models, AI agents are systems that can act on their own and do a certain task for us. And those automations can be used in a business process. 

So when we have a business process defined, then we can take a look at our business process and identify any kind of bottlenecks or optimization opportunities and see how AI models, AI systems can actually automate the tasks that we are doing within those business processes.  

The third way to bring business value to the company, is by transforming the business process and transforming even the business model. This is also the most complicated way to bring value. 

So once we already have optimized and automated business processes, we can think how we can actually use the new technology to change the way we operate, to reshape the business processes that we have in the company. Make them smaller, more efficient, more cost effective. Or we can even think about new business models that can emerge from the capabilities of the new technology.  

MK: So this is the group of use cases that you’ve seen. From my field, I’ve also seen some use cases of implementing AI to BI solutions to identify patterns or generate insights. You can also implement gen AI solutions to enable users to talk to the data instead of just to digest what is visually seen on the screen. But it’s not so straightforward to generate use cases to find out how AI can help in my field.  

How to identify those use cases? Being a manager who has to implement AI, because this is the pressure from the company generates. We need to implement AI in every division of the company. This is the hype, we need to have this value. So how to generate those use cases, how to identify where we can implement AI?  

ML: There are two ways that those use cases can be defined. The first way is from the technology. We know the technology, we know the capabilities of different kind of AI models, different kind of AI algorithms, how to use them. And based on those capabilities, then we take a look at the business and think how those capabilities can be used to optimize a business process or to automate certain tasks and so on and so on.  

The second approach is from the business side. We take a look at a business process. We identify some bottlenecks or challenges within certain business process or area of our business. We also observe some opportunities, for example, from market research. 

What are our customers looking for? What additional value they need? And if we will provide this value, we’ll get additional advantage over our competition. And based on those opportunities and bottlenecks in business processes, then we can brainstorm how AI can actually address those identified areas. And the use case is good when from this technology point of view and the business point of view are aligned. 

In the perfect situation, we have mapped the capabilities of AI, or another technology in general, and we mapped our challenges or bottlenecks in the business processes or business opportunities that we can provide to our customers. And then we address them with the capabilities of AI. Based on that, we have good candidates for good use cases. We need to evaluate them. The next step would be to assess the actual economic value.  

MK: Exactly. That’s what I wanted to ask. You need to see whether the cost covers the actual benefit from it.  

ML: Exactly, To leverage certain use case, we need to do some implementation. We need to train models or we utilize already trained models. The foundation models, but inference also costs. 

We’ll need to evaluate whether the productivity gains or additional revenue will actually cover the cost of building and running this use case. If we the value s positive, then this is a good candidate for a use case.  

But there are some additional factors that need to be taken into account. Like for example, if you want to train a model on your own data, do you have this data? Is the data of the right quality? Because if the data is biased or very poor quality, the resulting model will also provide a poor quality of responses.  

MK: By the way, Marcin has prepared a very interesting white paper about generating use cases for AI implementation. You can find the link to the document in the description of the episode. 

About the availability of data to learn from, there are cases, like the example of our client you mentioned, where you have generated some use cases, you have evaluated those use cases that will bring value to the company, but you had to do some pre-work to be able to implement this particular use case. What was the example?  

ML: As we said at the beginning, artificial intelligence is based on data and we teach those systems how to act based on examples that we provide to the model. And those examples are the actual data that we have. 

And the example that you’re referring to is a situation when we actually identified opportunities for the some use cases that would be beneficial for the company, but then we needed to build a roadmap. Set the priorities and based on that, build a roadmap when to build certain use cases. And some use cases, especially those that will utilize the company data were pushed a little bit back in time because the company needs to first build a data platform to collect high quality data, to have some history of the data. And based on this history, then we can train models that will be then implemented in the production.  

And a good example of such situation is building predictive maintenance solutions for manufacturing sites. In general, those kinds of systems are trained on history of data, of failures, of maintenance that were provided to the manufacturing machinery. 

And based on the history of failures, of how machinery works in the regular conditions, the model is trained to predict what is the probability of failure, within two weeks, one month and so on. And based on that, we can then organize and optimize the maintenance schedule for the machines in the manufacturing site. But the feasibility of this use case is mostly based on availability of data before building such a model. We need to have the history first.  

MK: So we know the value, we know how to implement it, we know what we need to have in order to implement it. But what are the traps? What are the risks of implementing AI?  

ML: When we are talking about risks or challenges related to AI, I would actually put those challenges into two different buckets. The first bucket is, are challenges perceived by managers. So some kind of challenges from the business perspective. The second bucket is a technology related challenges and risks.  

As for the first one, according to research, 44% of managers ranks data privacy and security as the most important challenge is in AI systems. They are afraid that that data could be compromised or even used by different companies to train other models. And we’re talking about company-owned data.  

The second challenge is related with AI reliability and accuracy of the responses. 46% of managers also point to this challenge as one of the top ones. And 34% of managers also point to skills that they have in their employees, in their teams. And because this technology actually evolves very rapidly, gaining new skills and learning and adapting to how to use new tools is also a challenge for non-technical teams. And sometimes even for technical teams as well. 

So this was the first bucket. The second bucket is related with technology itself. The main challenge is data itself. Do we have right data? Do we have enough data to train our model? Because we need a lot of data to train models. 

Do we have non-biased data? Data that actually represents different aspects or different categories equally. It’s necessary if we want to train the models, for example, to make decisions. So all challenges related with the data.  

The second challenge is related with the cost of training models and running models. This is especially important for generative AI models. If we want just to have like a very narrow group of users that uses a generative AI-based chatbot, then the cost of inference might not be very high for our use case.  

MK: Because this is the cost of each prompt. 

ML: Yes. But if you want to make some kind of chatbot available for millions of users, then the costs are actually very, very high. We also need to have this in mind. 

And the third one, I would say, the explainability of models. And this is especially important here in Europe, where we have the AI Act that makes some rules that in certain cases where decisions made by a system can impact individual lives, the decisions need to be explainable. So in those situations, we rather tend to use simple approaches, simple algorithms, where it can be easily explained why certain decision was made, instead of building big neural networks that are more accurate than those simple models, but explaining why certain decision was made by the network is really difficult. 

MK: There are just too many parameters that are taken into account. So human brain can’t comprehend how the decision has been made, what kind of parameter has been taken into account, right?  

ML: Yes. You can think of neural networks as black boxes. So kind of models that were trained and they made decisions inside, but from the end user perspective, you don’t exactly know what factors actually were used to make the final decision by the model. And explaining exactly why and what was used for the decision making, that’s really challenging. 

MK: Like a credit score. 

ML: Yes, like a credit score. Exactly. Okay. 

MK: All right. So this is quite complex. I would say the plateau of different solutions is quite large. And I can imagine that there are a lot of managers and big companies where this  hype of AI and the pressure of implementing AI is very, very big and they need to do something with AI. They have budgets for that, even though budgets for traditional projects are being cut down. They do have new budgets for implementing AI. 

So being a manager like that, I’m kind of, you know, confused and a bit lost on how to start off with a project like that. So what is your advice? What are the, you know, high level steps that I need to take to start this kind of implementation?  

ML: That’s actually good. This is what we can also observe in market research that more than 80% of managers actually feel the push of implementing AI urgently in their organizations. So the steps that we would take, as we discussed earlier, would be first identify some business opportunities, and identify any kind of bottlenecks. If you want to optimize, for example, your internal business processes, then you map those business processes and identify bottlenecks in those processes. 

If you want to find opportunities within your products, then actually there’s like market research needed for asking actual customers what gains they are looking for. The second step would be to ideate how technical capabilities provided by AI models can actually address those identified opportunities or bottlenecks in business processes. Based on that, we’ll then have defined definition of different use cases. 

The first step would be then to evaluate each use case, whether or not it is feasible and whether or not it actually brings business value for the company. So the feasibility is mostly based on the availability of data. So something that we also discussed, that if we want to train our own model and use our own company data. We need to know whether or not we have this data, whether or not this is high quality data that we can use to build an AI system. 

And the business value, the actual economic value of the use case, is based on the evaluation of what kind of gains we’ll get from the use case, and what will be the cost of building and running such a solution for five or three years. What we want to get from this equation is a positive value. The expected gains should be higher, even much higher than the cost of building and running the solution.  

MK: Okay, sounds good. Thank you very much. This is for today. Be sure to check out the white paper that Marcin has prepared. And thanks for being here. 

ML Thank you, Maciek. It was nice talking to you.  

MK: I’ll see you in the next episode. Cheers.