Discover AI Foundations for Business Leaders

0:00My name is Simona. That's me here. And welcome to this content for the Microsoft AB731 AI Transformational Leader exam. Transformational leader, huh? It sounds so grand, doesn't it? Now, our first task in this video is to understand what generative AI is and how it fits into the broader AI landscape. And with that in mind, let's start with just the AI bit. And then we'll see where the generative AI part fits in.

0:31So you need to think of AI, artificial intelligence, as a collection of capabilities. So AI isn't just one thing, but what they do all have in common is that these capabilities allow computers to perform tasks which previously an actual person would have to have done with their human judgment and their human intelligence and their human experience and human learnings and whatnot. Hence, the idea of artificial intelligence as opposed to real-life human intelligence. And the kind of tasks that we're talking about here will include things like spotting patterns and recognizing images and understanding language and even understanding the sentiment of language and things like creating summaries and, I don't know, providing recommendations and more. We could go on.

1:19But of course, because it's a computer doing that rather than a mere mortal, one of the most obvious advantages we see is that we get faster results because, well, computers can process more information and they can do it faster than, you know, my brain or your brain could ever do without exploding. And for you, as a business leader, that's the interesting bit. We'll start to see as we progress through this course that these capabilities do deliver real business outcomes. And that, for sure, is going to be more interesting to you, I think, than the, you know, the technology and the algorithms and whatnot that make it all happen.

1:58So, therefore, keeping in mind that AI is this bigger picture of a collection of capabilities, well, generative AI is a subset of that bigger picture of AI. And you can see a good definition of it here. So, in a nutshell, we can describe generative AI as a subdomain or a branch of artificial intelligence, which is designed to produce or to generate. Did you see what I did there? It's designed to produce brand new content such as text or images or even video and audio from just a simple description.

2:32So, no programming required. Well, by that, I mean that there's no programming required by the mere mortal human being using the generative AI-powered tool. But, of course, there is an awful lot going on behind the scenes. And, in fact, under the hood, we actually are programming when we use these tools. It's just that we're programming in plain language, you know, natural language as if you were talking to another human. And we will pick up on that idea again another time when we learn how and why we need to put some thought into that simple description that we type rather than just type a thought and hope for the best.

3:08Oh, and I should say that this simple description would be known as a prompt. But as I said, more on that another time. So, I think this is probably a good definition for us to hold on to as to what generative AI is. And it is quite possible that you're already using a generative AI-powered tool like Copilot to produce summaries of, I don't know, Teams meetings or to give you insights on your Excel spreadsheet. And so, as a practical example here, in your Excel spreadsheet, you might launch, for example, Copilot Chat and then type in something like this.

3:43So, this would be the prompt. So, something like, analyze this workbook and give me interesting data insights. So, that's the prompt. That's our natural language request. And Copilot is going to review your big, horrible data set or whatever it is you've got in Excel. And then, as you can see here, it'll give you an overview of what your data actually looks like. And then it starts to identify some trends or whatever insights and correlations, et cetera, it can find in that data. Now, we will learn more about using Copilot and its capabilities later on.

4:17But I just wanted to show you a practical example of what I mean by generative AI. So, let me just go back to that previous slide. Yep, there it is. And because generative AI tends to be that branch of artificial intelligence which suddenly puts the power of AI into the hands of a regular end user, you know, looking at their Excel spreadsheet or whatever, and they neither know nor care what technical wizardry is going on behind the scenes, then it is tempting to think that generative AI is the sum total of what AI can do.

4:48But I really do need you to understand that, well, yes, generative AI generates new content from a natural language request. Marvelous. But also, I need you to appreciate that it's just a small part of the wider field of AI. It's just a branch or a subdomain of artificial intelligence. And in fact, let me show you this slide here, which I have stolen from my esteemed colleague, Jonathan Barrios, who has created loads of AI training here at CBT Nuggets. And I always say this, but he really is quite possibly one of the most interesting people you could ever hope to meet.

5:22But on his slide here, you can see that he says, if we think of AI as a box of goodies just here, well, yes, generative AI is in the box, but it is not the whole box. So AI, artificial intelligence, is the broad field, including things like machine learning and deep learning. And then, yep, there we go, generative AI. And this is where those generative AI powered tools like chatbots sit, which for many of us is our first experience of seeing generative AI in action with, I think it must have been chat GPT being one of the first on the scene.

5:59And then, of course, since then, we've seen others pop up with Gemini being Google's generative AI tool and, of course, Copilot being Microsoft's. And there are others, too. And leading us on to the next video in this skill, there is a term that you might hear in the context of AI and generative AI, and that's LLM, standing for large language model. And LLMs are key players in generative AI because they provide the foundation for your generative AI powered tools to be able to understand your natural language requests, you know, that prompt that you type in in natural language rather than programming.

6:38Now, under this video in the skill, I have given you a link to one of Jonathan's AI courses, the first skill of which does a really fantastic job of exploring AI and its different types and therefore how generative AI fits into the bigger picture. And I urge you to take a look if you would like to, but this is optional. You do not need that extra background for the AB731 exam. And if you really do want to get more technical, then Jonathan's got loads of other AI courses as well. And I urge you to take a look if that is the direction you want to go in.

7:09But anyway, having got our heads around what we mean by generative AI in the next video, I'd like us to jump in and learn about some of the different generative AI models.

0:00This is one of the objectives from the skills measured document for the AB731 exam. Describe the differences between AI models. And I think a good place to start is, well, what do we mean by an AI model? And we can kind of think of it as sort of like the sum total of all of the background maths or math as my American co-workers would say. So all of the background maths and probabilities and statistics and whatnot, which have been formed from learned patterns that have come out of vast amounts of data.

0:31We're talking internet scale data here. And these patterns and algorithms and whatnot are used to generate new content. And so at the simplest level, the model is using a kind of like a sophisticated network, if you like, of predictions to guide it as to what is the most likely thing it should produce based on the context. And the context would be, for example, the prompt that you provide it with. But another term I need you to be aware of in the context of generative AI models is training.

1:03And training is how these learned patterns in the model emerge. So, yes, a model is trained on huge data sets, massive internet scale data. And this idea of training means that its parameters or its own internal thinking, if you like, is adjusted in order to produce more accurate results. And the actual process of training is a fascinating and complex subject, the details of which we don't need to concern ourselves with right now. But what happens is it involves the model making a prediction, comparing what it comes up with, with the correct answer.

1:38And then it changes itself in order to be more accurate next time. So it's making changes to its parameters. And it's making mathematical corrections to all of this background wizardry in order to improve. Now, I realize this idea, if it's new to you, will take a little bit of processing. So go easy on yourself. All of this will start to feel more comfortable as we learn more. But with this idea of training in mind, be aware that traditional artificial intelligence solutions focused on building custom models for predictions.

2:13And I'm going to remind you of this when we come back to talk about machine learning later on in this skill. And so, yes, artificial intelligence solutions of a few years ago would all be based on custom models. But generative AI is the next layer of evolution and usability, if you like, of AI solutions, because it uses these models. And for that reason, we often describe generative AI as using pre-trained models. And that means that we can implement a generative AI solution without having to build our own custom model as per those traditional AI solutions.

2:52And that's great. But in addition, a pre-trained model could be further refined and customized for improved accuracy and relevance for your industry or your organization by a process of fine tuning. So that's a term I need you to be aware of. So a fine tuned model has undergone further training in order to, as I said, refine and customize the model so it performs better for a particular scenario. And sure, that will give you better results, you know, more accurate results. But the more fine tuning that you do, unsurprisingly, the more cost and complexity you add to the deployment of your AI solution.

3:31Now, you might also hear the term instruction tuned, which means the model has gone through another process of enhancement to learn how best to follow human commands. So this means they're trained to better interpret and execute human instructions and to respond in a conversational and helpful way that humans would expect. Now, some of the Microsoft learning content that you might have come across in your journey for preparing for this exam, well, some of that content refers to the term pre-tuned, which I'm not going to lie, isn't, I don't think, a general term that we see that often in the context of AI models.

4:07We normally just see the idea of various types of training, such as pre-training and fine tuning and instruction tuning. But if you do happen to see the term pre-tuned in the AB731 exam, then assume that the idea is that a pre-tuned AI model is ready to use and therefore offers a less expensive solution compared to a fine-tuned model that you've taken the trouble to further refine for your organisation. OK, so my goal on this slide was to explain the idea of pre-trained and fine-tuned.

4:42But in addition to all of this, there are some specific model types that I really do need you to be aware of. And I have laid them out in this rather boring table here because I do need you to be clear on these, since you are likely to need to be able to distinguish between them for the AB731 exam. And the first one on the list, well, this is the one that, if any of these that you've heard of, it's going to be this first one, an LLM, a large language model. And I did vaguely mention this, didn't I, in the previous video? Because as I said, that tends to be the one, if anybody's heard of an AI model, that's the one they've heard of.

5:15So this is the model that generates and understands natural language text. It's what we first think about when we talk about generative AI. This is what understands our plain language instructions. And it's this type of model that's going to need to undergo instruction training or instruction tuning, I should say, to better be able to interact in a natural conversational way. And an example of a large language model would be GPT-5, which at the time of recording is the model behind the chat GPT chatbot and also behind Copilot.

5:48So this is a screenshot here of Microsoft 365 Copilot chat. And up here, when you're starting a new conversation, if you were to click the dropdown, that's where you can specify the model that you want to use. And yeah, it's using GPT-5 in the background there as the underlying model. But anyway, let me just go back to the table. There we go. And next on our list here, small language models. You won't hear about these as much as you will hear about their bigger brother, the large language models. But a small language model is similar to an LLM in that it is a language model.

6:22So yes, it generates and understands natural language, but it's characterized by having a much smaller size and scope with fewer parameters than a large language model. And that means that they require less memory and less computational power. And that might be useful for mobile apps or even for offline use, or maybe for very task-specific or very closely defined or targeted domain-specific applications. You know, something with a small, carefully designed scope. So yeah, small is the operative word there in that type of model.

6:56Next on my list, code models. Well, these are models which have been optimized for use in generative AI-powered tools designed to help developers. So it's going to be used in solutions which will assist with tasks such as, I don't know, completing code or debugging or automating tasks or running tests or proposing changes to code, you know, that kind of thing. You know, it's specialized to the world of coding. So therefore, I think that's probably an easy one to remember because it does do what it says on the tin.

7:26Diffusion models here. Now, these types of models are primarily for image creation and image processing tasks. But the applications of these types of models has been extended to other domains such as the creation of audio and video from, you know, user text-based prompts. And it is quite fascinating and mind-blowing how these models work. But hey, I will leave you to research that if you're interested. Next, you can see I've got a multimodal model, which is a bit of a tongue twister, I'm not going to lie.

7:59But these models can handle multiple input types such as it says here, text, image and audio. So they are designed to process and integrate information from multiple data types or modalities. So modalities being the posh term for different data types. And GPT-5 is described as a multimodal LLM. It's a multimodal large language model. So a model can be both an LLM and multimodal.

8:30And then lastly, again, hopefully an easy one to understand. A domain-specific model, as you might guess, is a model that has been fine-tuned. Aha, we know about fine-tuning. It has been fine-tuned for specific industries or tasks. So hopefully that's an easy one to remember. Now, as a business leader, it is unlikely that you need to understand the technical detail behind these models. But you do, I think, need to be able to hold your own in technical discussions where different models are being discussed and being assessed for use in the AI solution that you're deploying.

9:03And I really do expect you to be asking these types of questions when your AI solution is being developed. So, you know, how much cost and complexity will fine-tuning of the model add? That is a very relevant question to be asking. And what about something like this? Is there a model available which has been pre-tuned for our domain? So you'll see I've used that term pre-tuned there to mean something that's been further refined specific to our particular area of domain, whether it's healthcare or legal or something. Or maybe something like, is the model multimodal?

9:37So perhaps your solution is going to be used by marketing people who might be requiring something that can handle different data inputs, such as text and images and whatnot. But you get the idea. And later on, you are going to get the opportunity to match up the right AI model to a business objective. And that is absolutely going to be the kind of discussion you'll need to be a part of.

0:00Taking a look at the list of skills measured for the AB731 exam, there are two which are all about machine learning. Here they are. So number one, identify scenarios where machine learning adds value. And then number two, describe the life cycle of a machine learning solution. And surprise, surprise, that's what we're going to tackle in this video. I think we should probably start by making sure that we can define the term machine learning. And let's start though with what you already know. So you will remember that generative AI is that subdomain of artificial intelligence that generates stuff.

0:38So we saw this slide, didn't we, in the first video in this skill. And you might also remember this slide, which I stole from Jonathan Barrios, one of the other CBT Nuggets trainers. And you can see that, yes, aha, here we have machine learning as one of the other subdomains of AI. It's one of the other goodies inside the box. So let's get ourselves a definition of machine learning. And taking Microsoft's definition here seems as good a one as any. So what is machine learning? Well, you can see here it's a branch of AI that enables computers to learn from data and improve their performance over time without being explicitly programmed.

1:20And if this is starting to sound slightly familiar from the last video when we learned about generative AI models, then yes, generative AI is powered by models which have been trained using machine learning or should I say pre-trained. Remember, we talked about this idea of a model being pre-trained that we can plug into our generative AI solution. And you might remember that these models adjust themselves during training in order to become more accurate. They make changes themselves to their own parameters and the mathematical wizardry and whatnot in order to provide more accurate responses. And that's what we're saying here with this definition. They improve their performance over time without being explicitly programmed. That's the beauty of the idea of training a model. We're not hard coding.

2:12If this happens, then do that. We are training it to interpret the data correctly. So if that's our definition of machine learning that we will go with, now let's tackle our first objective for this video. And that is to identify scenarios where machine learning adds value. And I have to say, I am kind of stepping on the toes a little bit of what I want to say when we talk more later on about the business value of AI. But what we need to appreciate at this moment is that the point of using machine learning as part of your AI solution lies in scenarios where you are not explicitly focused on generating new content like we do with the generative AI subdomain of artificial intelligence. But with machine learning, you're doing these kinds of things here. And so yes, generative AI is great for generating new content. But machine learning is better suited for tasks like this, such as numerical prediction or classification and clustering. And what you see here would be some common scenarios.

3:15So predictive analysis would help us, for example, to forecast sales or understand customer demand or predict risk or predict customer churn, that kind of thing. While personalization could be more about classification. So perhaps grouping customers together who share similar traits or behaviors, grouping them into different market segments to help us understand our customer base better. And then automation might be things like detecting anomalies in our data. And that could be in the context of financial transactions in the case of detecting fraud, for example, or maybe even looking for faults and probable causes in a manufacturing process as another example. So do you see that these kind of solutions aren't about generating new and original content from a simple text-based description, which is what we see with generative AI. These are about predictions. And I really do think that predictions is the key word here.

4:11And in fact, you might also hear the term inferencing, which simply describes the scenario where a trained machine learning model is used to make, you guessed it, predictions. And we will be embedding this predictive functionality, if you like, into applications or services or whatever solution we're building. And I do hope that you can read sideways there. So that does say that we embed this type of stuff into an app or a service. Anyway, I'm going to be testing you on some of this shortly when we start to bring all of this together by looking at some of the business value of AI solutions. But for now, just know that machine learning is very much focused on predictions. But the other objective for this video is, as you might remember, to describe the life cycle of a machine learning solution. So that's what we'll do now. And you can see here that I've got the bare bones of a flowchart. And we are going to fill in the gaps together.

5:06So the first shape on the left hand side here, that is the data. Now, this could be your data, it could be just general internet data. But the point is, there's some kind of data source that you're going to be building your machine learning model with. And then you train the model. Well, I say you train the model, your technical team will get involved in training the model. But the point is that the model gets trained on that data, we then deploy the model, we will then be monitoring the model in terms of determining how it's getting on with its training and its learning. And then we can analyse its performance and retrain it. That will lead us to updating the data and the cycle starts again, you know, the process continues. Now, this cycle describes what we would call MLOps. That's machine learning operations. It's kind of like DevOps. But for machine learning. So it's the process for developing machine learning models for production.

6:00It's kind of the structure of processes for reliably building and monitoring and maintaining machine learning models. So it's a, you know, it's a set of practices for repeatable processes for the development of machine learning models. And while it will be your technical teams who are concerned with the detail of MLOps, that's why I kind of glossed over it quite briefly. It is really important for business leaders to have this overview and your input as a business leader will make sure that all of this MLOps work stays aligned with business goals. So for example, in the red here, you will be involved in exploring and preparing the data, especially if this is going to be data based from your organisation, then it's up to you as a business leader to be clarifying that this is appropriate and current and the right data to be using. And you will also be part of the process when it comes to evaluating the outcomes. Because when we say analyse the performance, well, what are we measuring it by? How are we determining whether performance is good or bad? What are we analysing for? And the answer is, we are checking for alignment with our business goals, aren't we? So if we take a big step back and look at the entire machine learning project as a whole, as outlined in purple here, then there's one more thing we need to add to complete

7:16the picture. And that's defining the goals of the solution that we're building this model for. What are we trying to achieve and why? You know, what are the tasks we want this solution to accomplish? So this is where it all starts. And that requires business leader input. So what I'd like you to take away from this slide is that the blue bits in the middle is a simple representation of an MLOps machine learning model lifecycle, the detail of which will be of concern to your technical teams. The red bits is where we require your input as a business leader. And then the overall diagram outlined in purple might be one way of looking at the lifecycle of a machine learning project.

0:00As a business leader, you might be thinking, sure, some of this AI stuff looks interesting and maybe you've got people working with you or for you who are already excited about it. But I'm hoping that you're thinking, where is the business value? And yes, you really should be asking this question. We can't just deploy stuff because we can. We have to understand where the potential returns on investment might be. So where is the business value of the AI solution? And what we're finding is that, yes, AI does have the potential to offer every kind of organisation new ways to streamline their internal processes and their wider operations and new ways to innovate, whether that's new ways to connect with customers or new insights into customer behaviours or, I don't know, new ways to design products.

0:45All of this means that we've got new ways to stay competitive. And therefore, if we aren't using AI and our competitors are, then, you know, maybe we're in danger of being left behind. And while adopting AI might seem complex and costly, the potential return on investment really is significant. And that's where you come in. So if we think about AI as being more than just a technological upgrade, the responsibility of which will fall to your technical teams, think of AI instead as a strategic asset.

1:18And then it becomes easier to see that the success of it depends not only on the tools, but also on strong leadership and a clear roadmap and a clear shared understanding on what we're trying to achieve and what problems we're solving. And as I said, that's where you come in. So a successful AI project relies on you clearly understanding what the business value is going to be, making sure that that vision is shared by the technical teams who will be implementing it, and then leading the organisation through the, you know, through the transition to get to the point where business value is realised.

1:51And truly, the possibilities are endless because AI powers everything from chatbots to automation tools to advanced analytics. So your role is going to be to identify the potential that AI offers for solving real world problems in your organisation. So let's start with some scenarios where we can apply AI to business to start giving you some ideas. And I have lifted these from the Microsoft documentation, just to be sure that we stay on track with the lines of thought that the exam will be testing us on. And these are just four examples to help us bring the possibilities to life.

2:25So firstly, you can see here generating marketing content. And what we're talking about here is using AI to create personalised or localised campaign copy or promotional emails or product descriptions or social media content, whatever it is, you know, give AI the task of creating this kind of content, which is customised for different market segments, or different audiences or different geographies or whatever. That legwork of creating this personalised copy can take minutes using AI rather than literally weeks using humans.

2:57So the marketing folks will still be reviewing and refining it, of course, but the time savings are tremendous. And the people who don't have to write all of that content from scratch now have the opportunity to apply their talents to other value add tasks instead. So in terms of business value, then yes, sure, using AI in this marketing context is going to accelerate launches, it's going to accelerate, you know, campaigns and product launches and whatnot. It will, of course, reduce production costs in terms of the hours spent creating it.

3:27And it is going to increase engagement by easily being able to have customised messaging for our different customer segments. Next on the list, though, in fact, let me change my pen colour to purple here because I can already see that this slide is going to get rather busy. But in a customer support scenario, we can use AI to automate responses to common enquiries, probably in the form of a chatbot, or by integrating AI into the processes of our customer service teams. I mean, maybe you could use AI to summarise previous interactions with that customer and then use it to draft emails in the tone of voice that you specify.

4:03And maybe the solution is even configured to suggest what steps the customer service representative should take next. And in fact, we would call that prescriptive AI when it's configured to help with recommendations and optimisation, you know, when it helps with decisions and recommends specific actions. But again, where's the value? Well, the value is going to be in reducing resolution times for customer queries and reducing support costs and improving customer satisfaction, of course. And ultimately, being able to provide better and more complete customer service without having to increase your headcount.

4:39Next on the list here, document drafts and summaries. Now that might sound a bit lame. But look, if you've got weighty documentation to deal with, such as regulatory compliance documentation or legal documentation, or very heavy, detailed financial reports, then we can use AI to extract the essence of that documentation in the form of text-based summaries or even audio summaries, or to generate FAQs for the team, or for your customers, or whatever the audience is, or to pull out what's changed from the last version of that document.

5:12It's saving many, many hours of human review when the size of the documentation is significant and the content is complex. And the business value? Well, aside from the time saved by automating these tasks, we are reducing the risk of non-compliance in the context of regulations and whatnot by making sure that the content is easily available and consumable to the people who need to know it. And we're potentially reducing legal review time. And there's certainly going to be a cost saving there, isn't there? And then last on my list here in the green scribble, AI is going to help us accelerate product design and innovation too by helping to simulate new product designs, or helping product teams to brainstorm, you know, helping to explore creative options, or optimizing supply chains.

5:57The possibilities are endless. And the business value? Well, we're going to be able to have faster innovation times and faster times to market. Note, though, that these are all examples of scenarios where generative AI would be used. And I think the clue was in the first one, wasn't it? That we generate marketing content. So these are all examples of generative AI. And if we turn our minds to examples of where machine learning can add business value, well, happily, this is what we mentioned in the last video.

6:28So thinking of those machine learning models embedded as a solution into an app or a service, we did cover these same examples, in fact, in the last video. So yes, the first two on my list here, predicting customer churn or assessing risk forecasts, for example, that's all to do with the predictive analysis that we know machine learning models are so good at. Anything to do with predictions is going to be an indicator that we're talking about some kind of machine learning model. And then personalizing product suggestions here. So that's going to be all about personalization and classification, you know, grouping customers who share traits, grouping them together, and then uncovering the likely next product of interest.

7:07If, for example, this type of customer buys that type of product, that kind of thing. And we also mentioned, didn't we, how machine learning models are very useful for automatically detecting anomalies and outliers in data and detecting fraud in financial data, being a great example here, and maybe spotting problems in a manufacturing process by analyzing quality inspection results, that kind of thing. So yes, we already mentioned those, didn't we, in the last video as examples of where machine learning adds value. So having got some ideas as to where AI could be applied in business, now you need to start thinking about where it could be applied to your organization.

7:45And Microsoft recommends that we, and I quote, identify and elevate business value, which sounds rather grander than it is. But it is basically these really sensible and practical three steps here, which are kind of obvious when you think about them. But the point is, somebody somewhere has to think about them. So firstly, when we're talking about implementing some kind of AI solution, we have to ask ourselves, what problem is this solving? What are the pain points?

8:16And again, I put pain points in quotes there. That's a term that Microsoft loves to use to describe any areas of inefficiency or slow processes or high costs or, I don't know, bottlenecks in your systems or scenarios where you lose customers to a competitor or you've got too many customer complaints or returns, whatever it is, identify the pain points in your organization. Because if we don't know what problem we're trying to solve with our AI solution, then how do we know if we've been successful? We're not going to know if it's delivered any business value if we haven't properly identified what problem is it designed to solve.

8:55And then number two on the list here, we've got to make it measurable. We've got to tie all of our AI initiatives to clear data, you know, identifying those KPIs, key performance indicators, such as time to value or accuracy or speed or, I don't know, time to close a ticket or the number of returns or cost savings or the customer satisfaction ratings. Whatever it is, make sure it's measurable. Put a number on it. Let the data do the talking. Because otherwise, how can we determine whether you've got a return on the investment that you made in your AI solution?

9:30And then thirdly here, does it align with your organization's priorities? Does it fit in with your organization's bigger picture and mission? You know, does it support your compliance regulations? That kind of thing. Because otherwise, I mean, if what you're trying to do doesn't align with your organization's strategic goals, then you're not going to get funding or support for it. So why bother? And at the end of the day, you have to show that your AI initiative does have a part to play in moving the organization forwards in the direction that senior management want it to go.

10:04So I think these three steps here is a great place to start when it comes to identifying where AI could add business value to your organization. But having done that, we need to think about how we, and again, I quote here, how we turn capability into outcomes. And I think these steps here would be performed in conjunction with your technical teams. But the point is, for business leaders, the question we need answering is how do we translate the AI capabilities into business value?

10:36That's the outcomes here. And we need some kind of end result that's a working AI solution that is repeatable and reliable and doesn't expose us to additional risk. But anyway, as I said, having identified the business problem, as per the last slide, we now need to be sure to look at this, match the right capability with the need. So this is all about matching the right AI capability with the need. Is it a generative AI solution or a predictive AI solution as per the machine learning scenarios we talked about?

11:07How much customization or fine tuning do we need to do? So matching the right capability with the need. And then number two here, investing in data quality. And again, this is going to be something that will be performed in conjunction with your technical teams. So having clean, consistent, and well-labeled and well-classified data is going to be the foundation of good AI, because without it, even the best models aren't going to be able to deliver what you need it to deliver. And then planning for operations. So you need to be defining how you're going to be measuring performance and monitoring the output and how you're going to be detecting any drift in the results.

11:43Or when do we need to retrain models as conditions change in order to keep the outcomes that we need for our business to get the value that we're expecting? How do we keep that output stable as things change and time passes? And we must always maintain human oversight. So the goal is that we're using AI to add expertise, to augment and enhance what our humans are doing, not to replace humans, especially for high-stake decisions where oversight and context and human judgment really do matter.

12:14And then, of course, we need to, well, measure what matters. And that's just a catchy way of saying, sure, having, as per the last slide, identified the metrics that we're trying to achieve here, we need to make sure we are measuring those outcomes that the solution delivers so that we can see whether we have achieved that business value, have we achieved a return on our investment? The overall message being, if we don't have the data, if we can't measure the outcomes that we're after, then how can we determine business value that the AI solution has brought?

12:46We can't.

0:00OK, it's time to review what we've learned with a few extra quiz questions. And the first one was one of those matching quiz questions. So match up the AI model with the business need. So if we have a business need of needing to create visual content, then that's going to be a diffusion model type. That's what we call those types of AI models. Interaction with different input types requires a multimodal model. Programming assistance. Well, hopefully that was easy. I was trying to avoid using the word coding here because that would have been too easy.

0:32But yeah, obviously programming assistance is going to be a code model. A customer support chat bot is probably a good example of an LLM, a large language model for natural language input and responses. And then look here, a compact language model for a closely defined scope. So that was the large language model's little brother, the small language model. OK, on to the next question. It's another matching question. And quite a few matches to make here. So I know this might have been quite tricky, but I really do want you to think about this.

1:04But this is all about the machine learning project lifecycle. So these are the different stages in the correct order here. Let's match it up with the description. So firstly, we need to define the problem, which would be described as identify the business question the model should answer. Then we collect the data. So there we go. Ensure access to the right data sources. And then we prepare the data. Yeah, there we go. So the technical people will clean it up and then leaders are going to ensure that it's accurate and complete and relevant and compliant and whatnot. And then train the model.

1:37Oh, yeah, there we go. That's where the model learns from the data. We deploy the model. So the model gets integrated into the business process or the app, whatever the solution is. And then number six here, monitoring performance. That's got to be the only option left. We track how the model performs and ensure alignment with business needs. Good. Okay, the next question then, which three of the following are use cases for a generative AI solution? So yes, summarizing documents, creating marketing content.

2:08So we're looking for three and drafting reports. There we go. That'll do. And you might have spotted that the next question is the same set of possible answers, but this time it's three use cases for a machine learning solution. So this time we're looking for personalization of product suggestions and I've just spotted a typing error there. So I'll hopefully try and remember to correct that before you get to see this. But also detecting fraud. Yes. And predicting customer churn. Good. And then finally, what should you do to convince senior management that the AI solution is worth the investment?

2:42So hopefully you disregarded many of these as being complete and utter nonsense. It's this one we're after. We need to show them how the solution will align with their strategic goals. Well done.