Harness AI/ML to drive innovation and unlock new opportunities

Hey, thank you all for coming to today's session about harnessing the power of AI/ML to drive innovation and unlocking new opportunities in your organization.

Before we begin our session, let me ask you a first question - how many of you believe AI/ML is going to be the future? But your organization is still struggling to realize the full value of AI/ML. Great. I'm not glad to hear that, but I'm not surprised as well. And this is the real challenge and that's why we want to share in today's session, how important it is for you to focus on certain priorities in your organization.

So first of all, with so much potential, I can tell you, we have just begun to scratch the surface of AI/ML and what is possible with this technology. And there are only a few organizations who are successful in actually taking advantage of this technology and actually aligning that with their business goals.

Whether you just started, began your journey or are already in the process of expanding your efforts on AI/ML, in today's session, we are going to share some of the best practices and give you some actionable insights. So that after this session, when you go back, you can apply that to your organization.

With that, I'm Mihir Patel, I'm a Principal Solutions Architect at AWS. And with me is Rana Gupta, I'm also a Principal Solutions Architect at AWS. And I have something important to tell you. At the end of this talk, I'm going to give out a secret word and that secret word is gonna enable you to play a treasure hunt game. And why do you want to play the treasure hunt game? Because that's gonna give you an opportunity to win some cool prizes. Sound good? So just stay till the end of the talk and I'll give out that secret word. Thanks and back to Yates.

Thank you, Donna. So at a very high level, this is the agenda we are going to cover. We'll start with some of the key challenges and the approach you want to take for an AI/ML strategy. Second would be driving innovation and using AI/ML to generate AI. In today's world, we cannot complete our talk about AI without talking about generative AI. And we will cover some of the most common use cases so that you can start thinking about how you can apply this back to your organization.

In an overview of AI/ML, I want to cover different categories - subsets of AI - whether it is traditional ML, deep learning or generative AI. After my talk, my colleague Rana will actually talk about some of the real world use cases used by some of the leading organizations, and how many of you have heard about Amazon CodeWhisperer service?

Great, nice. So that's the generative AI service that we offer and run. I will talk about CodeWhisperer but also how you can apply CodeWhisperer to some of the use cases and we'll have some key takeaways.

So talking about some of the key challenges in implementing AI/ML in your organization, let me ask this - how many of you think skills and technology are the top barriers in implementing AI/ML in your organization? Great, not many hands raised.

So this is based on the Gartner study, actually skills and technology are not the top barriers, but actually it is business and data. The reason being is if you, if I had asked this question to you before the cloud era, that makes sense because after the cloud, so many things have been simplified, there are so many frameworks and services, managed services available so that you can easily apply that into your organizations.

So talking about business, why is that the top barrier? There are a couple of reasons. The first reason is that there are very few organizations who actually understand how it can be applied and what are the benefits of AI/ML, right? That's one of the biggest challenges - they don't know where and how to apply it in different business units.

Second is the data. So there are very few organizations where they have democratized the data in their organization. So very few organizations who have actually made the data accessible to everybody in a secured manner. I don't know, but let me check - how many of you can easily identify the data in your organizations? You can quickly search and know that all this data is available, right? So that's a challenge. This is where the data strategy comes into the picture.

But because of these challenges, not many companies are able to actually take advantage of that. And data quality is also one of the biggest challenges that we have heard so far. When we talk about data quality, there could be data available, but they could be sitting in different silos in your organization. It could be maybe in the databases or maybe in some form of documents or maybe they are not easily available for you to use that to actually run the machine learning or build the AI solution on top of that.

So because of these challenges, not many companies can take advantage of that. But how can you address these challenges? As you look at this slide - many organizations actually start from the right. They start with the technology. And if you ask me, that's not a wrong approach because when any new technology comes, you want to make sure that you can basically learn about that technology, the team is able to experiment with the technology so that they understand the full capabilities of that technology, right?

The problem comes when you launch the product from the technology - what happens with that approach is that when you launch the product from technology, there will be a huge challenge with the adoption of that solution, right? So if you are not able to measure the impact, then you will not actually take advantage of that technology.

And there are very few organizations who start from the left. And these are the most successful organizations who start from the strategy. Once they understand the full potential of that technology, they want to make sure that they align it with the business strategies and focusing on the customer needs. And with that, they will be able to measure what would be the impact of using the solution.

So this is what we always say at Amazon - fall in love with the problem or opportunity, not the solution. Basically make sure that you are not over-obsessed with the technology. It's important to note that while technology is a great enabler, it is just one of the tools that enables your organization to make a difference.

So how can you go about it? This is what we always believe at Amazon - you want to focus on your customer first, whether it's an internal customer or external customer. You want to first make sure who is going to leverage this solution that you are building using AI/ML.

Now, when you understand the customer, then what is the most important benefit that customer would get out of that solution? And why is that an important problem to solve in your organization, right? And how does that align with some of your top level business goals? And those top level business goals could be either increasing revenue or reducing cost or maybe capturing more market share, right?

Once you start answering these questions, it will help you understand - is it worth putting your effort and money into that initiative?

So we talked about the business challenges, we also talked about data challenges. Now, once you know your customer, you are able to frame the problem statement and you know how that aligns with your top level business goal, this is the most important next step that comes into the picture - identifying the data that you want to feed into that machine learning model.

And this is still true with generative AI - your data is always going to be the differentiator for any solution that you are going to build. So once you know that this is the outcome that you want to generate or some value that you want to predict, for example maybe you want to predict sales or maybe you want to forecast something in your organization, right? Once you know that outcome, then now for that outcome what data do you need to use? You need to identify that data in your organizations.

And sometimes you may not have all the data available in your organization. For example, if you want to predict sales, in that case the sales data could be dependent on some external factors such as weather data, right? So the whole point is you have to make sure that you have all the data available so that you can feed it into the machine learning model training.

And once you train the model, you want to test and validate before you use it in a real world use case. Now when we talk about generative AI, this is an optional step - training the model - and we will cover in detail why that is an optional step. But there will be a few use cases where you want to fine tune an existing, large language model, and sometimes we call it a foundation model as well.

So let's talk about some of the common use cases used in the industry. This is based on the IDC study - for many organizations who embark on their journey into AI/ML, based on the study, and also when we talked to so many customers so far, we found that either improving operational efficiency, improving customer experience, or improving employee productivity - these are the top business objectives and objectives that many organizations want to achieve using AI.

The last one is increasing innovation, which is a generic statement because with innovation, somehow you want to align that. As I mentioned previously, how would that align with your top level business goals, right? As I said, it could be anything like increasing revenue or anything.

So let's talk about some of the use cases for each of the categories. Enhancing customer experience - talking about the use cases here - these are some of the common use cases, not the only use cases, but the most common use case that we have seen in the industry is personalization.

Because talking to many customers, what we found is they want to increase their user engagement with their product or solution. If you have a product which generates revenue, you definitely want to increase user engagement with it. And how can you increase that user engagement is through personalization.

What are the chances that you will open or click on some of the links if that email or message is personalized for you, right? If you are searching for some product and if it is available at a discounted price, there are high probabilities that you want to look into that email campaign or message, right?

Second is virtual assistants. Let me ask this - how many of you have made a customer service call and got frustrated with that experience? I'm one of them. The reason being is I have to wait for at least 30 minutes to get a hold of any real agent. And once I get a hold of that agent, I will explain my challenges and issues. And the worst part is if my call gets transferred to another department, I have to start that whole conversation from the beginning.

Now, if you look at that problem statement, you will also find the opportunity to build a solution around it. So think about a solution here - when an agent takes my call, what if that agent knows my intention - why I'm calling them? And what if all the information is available to the agent when that agent is talking to me? That can significantly reduce the time I spend with that agent. It improves my experience but not only that, it could be a flywheel - think about that, that same agent probably can support four times more calls, right? That would reduce the cost for that organization. It also enhances my experience, right? So there are multiple advantages.

Now talking about boosting productivity of internal employees - this is something I do on a day to day basis when I go into any meeting. And if I know that the call is being recorded, I don't focus on taking notes in that meeting. Rather I focus my full attention on the conversation because at the end I can download that transcript and I can feed that data into a large language model and I can ask it to summarize - what are the highlights? What are the next steps that we discussed? Right?

That saves my 30 to 40 minutes of time. And believe me, I use gender TV, I for my personal productivity improvement, maybe 5 to 10 times a day. That makes a huge difference because now I can spend my time on other more productive activities. Not just that now code generation, what if your developers can deliver the product in maybe half of the time, right? Without uh too many bugs that significantly makes the difference in like, ok, how many features your organizations can deliver for your customers?

So based on our internal study, we have code whisper which is the service based on generative a i. So based on our internal study, it improved the productivity between 50 to 60% for our internal developers. So that could make a big difference, right?

So why I'm talking about all these use cases because we want you to think about like how to apply eiml on your day to day bus business, right? And we truly believe that every organizations, regardless of size of the organization and every business unit can benefit from. eiml

Last one is optimizing business pro uh processes and improving uh operational efficiency. This is the most common use case i hear from multiple customers, they have data into unstructured format and they want to extract some of the key information out of that, either they are in financial services, business or healthcare business, they want to get the data out of that so that they can take some of the business critical actions based on that. Or maybe they can automate that process, they can feed that data into their uh automation, right? So that reduces the manual effort.

Now talking about some of the uh machine learning use cases. So under the e i, there are different subsets of e i, this is the most common one traditional machine learning. So as you can see on the slide here we are talking about predicting something whether it is a forecast sales or maybe detecting anomalies from uh your probably user data, right? Or maybe you want to identify some fraudulent transaction. So here we are talking about predicting some value or maybe you want to classify some things such as whether that email was spam email or not.

So in this kind of a machine learning problem, you actually use your historical data and let me quickly check, do you understand these terms, labeled and unlabeled data? Great. Ok. Just wanted to confirm that label data is nothing but the data where you already know the outcome. Suppose in case of uh transactions, credit card transactions, if you know from the past credit card, credit card transactions that these are the fraudulent transactions, which means you know the outcome of those transactions so that you can feed that data into machine learning training model so that it can learn from it.

So in this case, traditional machine learning, what internally it does is it comes up, it comes up with math mathematical formula when it goes through the training process so that it can start predicting for future unseen data set.

When we talk about unlabeled data, there are few use cases, maybe you don't know what outcome to produce. So the classic example in that case is maybe you want to segment your customer base and based on that segmentation, maybe you want to run some kind of campaigning, right? So that's an example of unlabeled data set.

No, as you can see on this screen here, we are talking about even more com complex business uh complex problem. And as you can see here we are talking about unstructured data set. Maybe you want to identify something from image or video, you want to use speech recognition. And as i mentioned before, maybe you want to extract some key information from the document, right? Document processing autonomous vehicle is also an example which falls into this category.

So what is deep learning? It is a subset of a i and in this case, it is the artificial neuron network which works exactly like a human brains. So there are layers of artificial neuro neuron. So that information passes through this nonlinear layer of neurons and it can abstract the information out of that. So that it can start identifying some of the information such as you may want to identify some activities from the image or video, right.

So these kind of problems can be solved using deep learning, talking about generative v i which is against the subset of deep learning. But in case of generative a a i, these models are trained on internet of data, we are talking about not just billions trillions of parameters. And because of that reason, it can solve even more complex real world use cases and how that helps as you all probably heard about chatbot, right?

So chatbot is just a user interface but under the hood that generate a model gp t four and new versions of a model that comes out. And it's not just like ok gp t is the only model available in the market, but there are multiple models from cloud or jurassic and stability, they are available, but it can help you to augment some of the human activities such as as i mentioned before, suppose there is a 2030 page of document and you want to summarize that document. And when you ask that question to summarize it for you, it can go through all that data and it can summarize in a form that you expect.

But and i will talk about some of the tips when you work with uh this la language models. So if you are wondering how generative a i completes the sentence, right? So basically generative a i is what it generates the content. But how does this it generates the content? As you can see on the screen, the sentence is sorry, sentences, the students open there. And after that possibilities of so many words, how does it choose the right word?

All right. So it's possible because of uh transformer architecture which is the evolution and its ability to uh its ability to learn the relevance and context of each word and not just next to each other, but context and reliance of each word to all the words, words in the sentence. So that now and then it can give some weight to the words and the word with the highest weight gets picked up and this is how it generates the content for you.

So let me talk about some of the tips when you work with uh large language models, which we call it as foundation models. So if you work with the foundation models, make sure you are aware about this concept called prompt engineering. This is very critical when you work with the foundation model because it can influence the behavior and tone and style of the content it can generate.

So when you work with it, make sure that you are very specific what you want foundation model to generate. Suppose say if you want to send automated email to your customer who is maybe not happy with some of your service, you wanted to write an email by providing some add additional context and maybe you want to provide some references and ask specific type of, ask it to generate spec specific type of email, right? So be more specific so that you can get a model to generate a type of uh like outcome that you expect from it.

Se second is don't think about this word for now retrieval augmented generation, but this is the technique it helps to reduce the hallucination. What do i mean by that? This model can confidently give you the inaccurate answer because founders and models are trained based on internet of data, they may not have any knowledge about the data for your organizations, right?

And when you want it to generate the outcome by looking at some of your internal data such as maybe past resolution or maybe it could be the stock information if you are in the financial industry or maybe some of the health benefit information if you are in health care industry. So suppose if you want to know what are the benefits of maybe xyz patient, right? If you are serving in healthcare industry, this foundation model may not have that information.

But what you can do, you can use this technique so that you can provide some of the content in the form of context while you ask that question to foundation model. And this technique allows you to go and look for that data extract that data and you can feed it into a large language model. And to do that, you need to first uh index the data into some of the uh storage which we call vector store so that it can go and find the right data set from the uh unstructured data set. And you can feed the data into this uh foundation model.

And finally, there are very few use cases where you may have to tune the model because as i said, these models are not trained for your specific domains. So in case of some maybe let's say uh insurance uh domain or maybe financial domain, you may want to fine tune that model. So what you can do is you can pick the existing foundation model, pass some of your internal data with the outcome that you expect and you feed into that existing foundation model so so that it can learn from it.

And with that, let me, yeah, let me hand it over to rana where he will talk about some of the industry use cases.

"Uh, here's the diet that you should be, uh you know, working on and things like that. But it's kind of a manual process to generate that discharge report after all that, that seven day hospital stay pages and pages of documentation. But now we have JA I systems looking through those pages and pages of clinical notes during that patient's hospital stay and able to actually generate and summarize a one page discharge report. Imagine how much labor that's going to save all these hospital systems. And if you're in the business of supplying software for clinics or hospitals, both of the two use cases that I just talked about are going to save a ton of hours and therefore, you know, gonna be very compelling new business opportunities, right?

Uh a third area by the way, uh is in drug discovery. So we're now seeing a whole lot of biopharmaceutical companies using gen A I to actually accelerate the development of new drugs and, and how can it do that? Well, that's because gen A I has been trained on molecular structures that have let that have led to effective medications in the past, right, effective drugs in the past. So therefore, the gen A I knows the chemistry knows the molecular structure. So it can kind of help biopharmaceuticals, uh you know, uh discover new drugs and accelerate that process because the candidates that they're, they're going to go through are more likely to be effective in treating uh disorders and things like that. Right. So some very, very exciting uh developments are happening now in the area of health care and gen A I now another uh industry where Jenny and I is adding a lot of positive value is financial services.

How many of you are in the financial services industry? Like banks and mutual funds, asset management? Ok. Quite a few of you. So you can imagine that uh you know, insurance companies and banks and so on, they normally have a customer facing chatbot of some sort, right? It does uh customer care, people can interact and ask questions about their account and things like that. But those chat bots are pretty rigid. There's not a whole lot of questions that they can answer, right? But now with gen A I, we're actually looking at chat bots that are so intelligent that they can help take you through an entire mortgage loan application process where you can ask all kinds of tangential questions, right? And it'll still help you complete the mortgage application on its own right, with very minimal interaction with the loan officer.

So for example, um you know, with a traditional uh chatbot, you might be applying for a mortgage and you know, it'll go through, it'll be quite rigid, it'll ask you, ok, how much do you want income and things like that? Right. But it can't answer questions out of left field. Like uh you know, the user might suddenly say, uh can you explain what are arms? What is an arm? How does that work? Right? Or hey, um what, why do you need my credit score? You know, uh i they can pay, basically ask all kind of off the cuff questions while they're interacting with the chatbot, which a traditional rigid programmed chatbot cannot but a gen A I chatbot can. Why? Because the JA I chatbot has a partner which is a foundation model and the foundation model can answer questions like why do you need my credit score? It'll be very patient and they'll say, well, you know, the reason we need your credit score is such and such and oh, by the way, you can go get your credit score by talking to experian or one of the one of the credit bureaus, right?

So what I'm saying is that these JA I enabled chatbots enabled free ranging uh conversations, tangential questions that will come up in these processes and do a pretty good job of answering them, right? Uh you haven't programmed it to do that but it can answer those kinds of questions. Uh another area, retail e commerce with websites being powered by gen A I uh we now have virtual try-ons going on. Uh people might be reluctant to uh to buy a pair of sunglasses or to try on some shoes cause they're looking at a something on the, on the website. But JA I can take a picture of, of your face, superimpose the uh the sunglasses on it and show you how you would look with those shades and that removes a barrier to you. buying that, buying those sunglasses. This is not some pie in the sky stuff this exists right now. Uh the amazon retail store has has virtual try-ons with sunglasses, shoes, right? And things like that.

I'm i'm just scratching the surface here. Every industry is going to be impacted by gen A I. You talk about advertising and marketing, I'm gonna talk about a uses in advertising and marketing soon. Uh travel, right? Booking all your travel, things like that. You can imagine that any industry out there. This light is not big enough to capture every industry that's going to be positively impacted by gen A I uh I, you know I mentioned what uh what health stripe uh was doing and here's an example uh if you go into the health stripe console and you give it an audio recording of the doctor interacting with the patient, you got that conversation going on in the left hand side.

Hey john, how are you? Uh well, doc I don't feel so good. I have, you know, I think I have some chest pains and it's just a free flowing conversation here, right? That got transcribed. You, you give that to heri heri is gonna come up with the structured record, right that you see here. Um ok. The the the chief complaint is the chest pain, the history of the patient illness is such and such. Here's the assessment, here's the treatment plan, right? It, it basically took this free flowing conversation between doctor and patient and came up with a structured health record that can automatically update the ehr so if you're in the business of providing software for clinicians, i think this is this is kind of a game changer.

Ok. Uh here's a use case that i think is going to be broadly applicable to just about every industry that's out there. And this is a use case that we call conversational document search. What this means is that you have a repository, almost every company has a repository of hundreds or even thousands of documents that they're storing in their internal knowledge base, their wikis, their share points, their confluence, whatever right? These are documents that could be, you know, hey, these are uh here are our training materials, here's how we onboard new employees. Uh oh, here's, here's some assets for our sales teams to use. Uh here's our financial statements for the last few quarters, every company has these troves of documents that they store internally.

Well, what if you could now build easily, build a question answer system that actually can carry on uh a humanlike conversation about all the information that you've got in your documents, right? Well, that's exactly what J E I enables. Uh and um I'm actually gonna show you that there are, there are examples here where you can quickly build a chatbot like this, right? That then goes into your documents. And here's an example where a company uh has stored all the sec filings of public companies that are out there, right? Uh and so any of their employees can ask, what was the unearned revenue for amazon in the year 2022? And the chatbot can come back and give you the answer that by itself is not that impressive. Ok? I'm i'm giving you a one word answer, but the open endedness of this is pretty amazing because they can actually ask, you know what on earth is un earned revenue anyway, right? It'll give you the answer. But more more than that, let's say now that i want to send an email to my colleague jane about this unearned revenue.

Well, all you gotta do is say, you know, draft me a concise email about this un earned revenue to jane and out comes this um this very nice email dear jane, i wanted to uh provide you with an update on amazon's unknown revenue, right? And then you click send and there you go, very rich chat experiences where it's also generating the text for you to take the next step, right? So i think this is going to be a very broad based use case and we can help you build these things pretty quickly.

Ok. Speaking of financial services companies, one of our biggest, uh well, one of our customers is one of, one of the biggest uh hedge funds in the world, i think right now, uh it's bridgewater associates and uh they have, last i checked, they have $100 billion or more under management in any case. Uh they have scores of investment analysts. Uh and what the company noticed is that these investment analysts were taking too long doing manual tasks. Ok? By the way, this is something that's going to be common throughout when you're thinking about what you can do to deliver value for your customers. Think first about all the manual tasks that they're doing, right. Think about automating those tasks. If you can automate those tasks, you've built something of aaa high degree of value, right? This is something that people are gonna want to pay you for, right? It looks easy to you. It looks hard to them, right? Something that looks, that looks hard to them. Yeah, they're gonna pay you for these kinds of things.

So, uh with uh with the problem that, that they saw was that their analysts were spending too much time on manual tasks. And what are some of these manual tasks? Uh these are things like, well, you know, if you're an investment analyst, you're often going to be asked to come up with a valuation model for a company. How do, how do i value a company like amazon and that model, you know, normally it uses something called discounted cash flow dc f analysis. And so you have to write some code right to actually get the output of a dc f. So they were, they were actually having to write a bunch of this code manually. Uh and they're also, you know, creating charts and graphs manually as well.

So what they did is they took amazon bedrock, which is our gen A I service and they took clad, which is the foundation model, uh one of the foundation models behind it and they built this virtual investment analyst assistant, right? So now what can this virtual assistant do? Well, one of their human analysts can just type in instructions in english, build me a financial model to value amazon, right? And it just goes ahead and generate some python code for that model. It doesn't, it doesn't matter what financial model you want, you can be trying to forecast earnings for a company for the next quarter. Well, that's another financial model, another set of python code again rather than writing that python code manually. The JA I system can actually produce that for you. Another area where JJ A I really excels is producing code as mesh said uh with code whisper, right? You can just type in a comment and like i generate the code for you, ok?

Um advertising and marketing, this industry is absolutely going to be impacted positively with gen A I, how many of you are working in marketing and sales and you are wanting to improve the open rate of your outbound emails for your marketing campaigns, right? If you're working in marketing, i mean, uh just about everybody has encountered a situation where the campaign isn't living up to, you know, what you want it to be in terms of how many people are engaging with you.

So here, here's where, you know, let's, let's, let's imagine that you are a, a marketing manager in a company that sells athletic shoes, right? So, what i did is i went into bedrock and i went into a playground in bedrock with a playground. You can just type anything you want just to play with it, right? So i said, you know, um create me a draft for me, a marketing email that uh describes a limited time offer to buy our running shoes. Ok. Uh and um out comes this, uh this marketing email and it's pretty compelling. I mean, it says, you know, we're having a flash sale with a 50% discount and it's only for the next 24 hours. And, oh, by the way, these shoes have breathable mesh uppers. They have plush memory foam midol and they have durable rubber outs. Soes. Right. Right. I look at this and i go, whoa, hey, this might be something i wanna go by. Right. But the interesting thing about this is that nowhere in my prompt did i say that my sale was gonna be 24 hours or how long it was gonna be? Nowhere in my prom? Did i describe what my shoes properties were at all? And yet bedrock with this clawed model in the playground gave me all this text. And the and the reason why it can do this is because it's been trained on marketing emails that have elicited a good response in the past, right? So having elicited. So it's learned what marketing emails have have elicited a good response, has that knowledge and can generate things like this for you, right? So try, try something like this out"

Uh and see, you know what kinds of marketing and advertising copy you can uh you can generate. Ok. Uh back to the world of big pharma Mark. Nobody needs an introduction to that company. They're a research intensive biopharmaceutical. Uh they started using Bedrock for knowledge mining and market research and what is knowledge mining? It's just taking, you know, terabytes and terabytes of information that you've collected over the years and using machine learning to try to gain some insights into that data.

So in this particular case, they've collected all this data about how their prescription medications have been dispensed. And uh what was the point in time at which the patient started taking it and what was the journey and so on. So they collected all this data, they integrated it with their patient workflow. So, so essentially they got some knowledge, insights from Bedrock, they integrated it with their patient workflows. And uh we actually uh heard from their um director of data science who said that by integrating the patient insights that they got from Bedrock with their workflows, they're actually able to improve the lives of patients, right? That I think is a significant outcome to be able to, you know, make a difference in the lives of patients.

Uh this is a great example of how gen AI right is changing the world for the better. Let's go to Salesforce. Uh Salesforce, of course, is everyone knows is one of the biggest software as a service companies out there. Uh and what they have built using Bedrock and gen AI, they've built something called Salesforce Data Cloud. And let me tell you what this does. If you've got a website, an ecommerce site that's powered by Salesforce Data Cloud, then you can capture uh when customers are coming to the site and know whether or not this customer has a high propensity to buy or not.

So here's how it's working. You have this prospective buyer that's shown on the left hand side here and um you know something about this buyer, you know, you know, their previous buying patterns and things like that, right? So the Salesforce data uh in the Salesforce data cloud can calculate a propensity to buy score, right? Based on what it knows about the customer's past buying behavior. And then what it does is if the customer has a high propensity to buy, then that interaction goes into Amazon Bedrock and Bedrock then generates an email that a sales rep can send right back to the customer.

So what's going on here is that with the Salesforce data cloud and Bedrock and generative AI will actually enable customer outreach because we are now engaging with the customer at the exact point that they are expressing interest in our products and we're not wasting our time because we know that this customer has a high propensity to buy our machine learning model gave us that right.

So now because of the fact that the sales rep and the prospective buyer are, are in an email has already been sent. Now, we can have questions and answers, you know, what is this product and things like that? And you've got some engagement going on, right? Ok. Another compelling use case, I think for JAI, so, you know, I've talked about Bedrock uh quite a few times here. What is uh so I want to get a little, get a little bit bit, I wanna get a little bit bit deeper into Bedrock uh and talk to you about what, what's going on uh underneath the hood there.

So why do we build Bedrock? We built Bedrock because we wanted our customers uh to have the easiest possible way of interacting with our foundation models to build JAI applications. So that means that you do not have to be a machine learning expert by any means, right? We have playgrounds where you can just type stuff in and you can see what happens. Uh and even if you're, if you can even hear programming and you're trying to build something, you just need a few lines of code to actually send a prompt into a model and get the answer back.

So there's no need for you to launch machine learning instances, manage them, right? It's just foundation models that are accessible on demand, it's completely serverless and it's a pay as you go service, right? That's why we built Bedrock. So it's quite easy, how do you use it? Well, just choose from one of the leading foundation models as shown here. And then the next step is if you want, you can customize it with your own data, right? Sometimes you don't even have to, sometimes the model is so good that you're, you know, you're good to run with it. But oftentimes you wanna customize it with your own data beyond that.

Um sometimes you actually want to automate some workflows. Now, we're going beyond question answering. Let's say you're in the travel business. Somebody asks you to uh generate a 10 day itinerary through Europe and, and you know, Jen is really good at that. It's gonna give you a whole travel plan with all the cities and the highlights and you know, but do you wanna stop there if you're, if you're in travel business? No, the next thing they wanna do is they wanna book all that travel, right? One of our customers is Booking.com, right? And they're actually using gen AI.

So if you want to go beyond chatting and doing workflows, you gotta have agents and I'm going to talk about agents for Bedrock. I think that's a very exciting new service. Uh that's going to automate a lot of manual workflows and have gen AI uh help, help that process. Ok. All right.

Um so these are some of the foundation models that underlie Bedrock. And um you can see that uh there's, uh there's actually six of them. So we at uh AWS believe in you having the broadest possible choice, right? In terms of being able to use uh AI models. So that's why there's six of them out there.

Uh if you want to generate images, you know, you've all seen the thing where, you know, there's an astronaut on a horse on the moon or something. And if you want to do something like that, you can use Stability, which is shown here, uh you know, in the, on the, on the, in the bottom. But if you uh want to just deal with text, then we've got this other five large language models and four of them are supplied by our partner Anthropic which with their Clawed model AI21 Labs, which is really good with foreign languages. By the way, if you dealing with foreign languages, go with the AI21 Labs. Uh Jurassic, we have Co here, we have Meta's Lama two and of course, we have our own uh large language model which is called Amazon Titan.

So how do you choose among these? I would say try them, try them all uh send your prompts into them. Uh and look at your business case, pick the one that solves your business case the best, right? They all have different cost models as well. Ok?

Um now I talked about uh agents. So here's what an agent can do. Let's say that you're in the insurance business and often times what happens is that someone submits a claim but they didn't submit the uh the photo, the damage, right? So you can quickly build an agent that actually reminds people, hey, you never sent your damage photos for your car accident claim, right? And you can say you can do that with one sentence and the agent comes back and says, ok, there's 22 claims and you know, uh these are the, these are the one and I have sent email reminders to all of them.

So this is powered by agents for Bedrock and to build an agent for Bedrock. It's not that hard. You select your foundation model, you just describe what the agent is doing. Ok? This is an insurance agent, right? It can find claims it can send reminder emails, uh write a few lambda functions that can, that can invoke those actions and describe what they are. Just describe everything in English. That's all you got to do. The gen AI the Bedrock, the agents for Bedrock can figure out what set of API calls to call in sequence to automate your workflow. You you're not writing all this code, right? To go through the workflow itself, you just describing what the API s do and the agent will figure out what to call in what sequence that's pretty powerful stuff. This is what we call low code, low code automation, right? Yeah, there is some code. It's not like there's a zero, but you don't need to hire an army of developers to do stuff like this. That's what I'm talking about.

Ok. So uh we've talked about, you know how, how companies are using AIML. Hopefully, i've been able to spark some ideas as to how you can use it in your business. Uh start with your customer and identify the opportunities, right?

Um use your data as a differentiator. You've got some proprietary knowledge base, use that to differentiate yourself from everybody else. And every one of your business units can benefit from gen AI, it could be your uh HR department, accounting department, marketing department doesn't matter, every one of them can benefit. And I would say talk to us, talk to your account team. We'll, uh, help you on your journey towards adopting AIML. We'll give you some demos, we'll, we'll help you go forward.

And the last thing is the treasure hunt, right? Ok. The secret word is innovation. Ok. The secret word is innovation. If you take a picture of this QR code, it will give you the instructions as how of how you can participate in this game and hopefully you win some prizes. All right, I'll give everybody a chance to click on this and capture this and thank you very much and don't, please don't forget to complete the uh, session survey. Uh, that's in your mobile app. Thank you very much for your time.