Drive personalized CX using generative AI and Amazon Personalize

Tia White: Hello, everyone. If you're anything like me, you are so busy jamming out to that music. I don't know if you want to hear me talk. Um so my name is Tia White. I am the GM of AI and ML at Tech and Mar Tech here at AWS. I actually just celebrated my two year, two year anniversary.

We are going to discuss the power of generative AI and Amazon Personalize, unlocking the value you all can create for your end users.

Today I am joined, or I will be joined on stage with Leah Morrison, Principal Solutions Architect, overseeing a portfolio of AI services as well as Darryl Bowden from Fox, who is the Executive Vice President of Technology Platforms.

Our agenda is jam packed with goodness. First, we will talk about why personalization is so important, which I'm sure I don't need to remind any of you why it's now a necessity. Then we'll discuss some of the challenges and why Amazon Personalize is poised to handle that for you. Now, we will share the latest features that we've launched throughout the year as well as features we just launched on Sunday night and then we'll get into some generative AI capabilities. We'll go deep with the demo from um Liam and then we'll end with Darryl sharing the the Fox story, how they're leveraging Amazon Personalized, how they will merge generative AI and Amazon Personalized to transform their end user experience.

Every touch point is personal. Whether it's the categories you show to your end users or the artwork that's displayed the titles being promoted, the promotions themselves, the order of the content, the carousels in which they show up for each individual user, even the order of the outbound communication and what's sent to whom you have the ability to hyper personalize at an individual level.

Your end user experience in 2023 we find ourselves in the midst of a transformative era underscored by the incredible and fast advancements that generative AI is unfolding before our eyes. We are actively listening to our customers keeping a pulse check on the industry and we fully understand how to create personalized experiences by fusing generative AI and Amazon. Personalize the opportunities are endless. You can do so much with the power of the two.

Now, I'll be dating myself here a bit, but I'm a part of the Gen Y demographic. So I'm between the age of 36 and 54. Now you try to figure out what my age is. We truly care about individualized experiences. My generation values, loyalty, we value brand, and more importantly, we value the brands that understand us as individuals and create a loyal experience for us to want to return our buying power exceeds $350 million.

Now we have that other generation that's gonna soon take over the world. Gen Z, they range between the age of 25 and 35. They represent a new phenomenon that we must all get ready for. Because by the year 2030 they will possess a lion's share of the currency power in the market. This generation appreciates u uniqueness, authenticity and self expression. This generation comprises more than 320 million in currency in the market. They themselves typically don't return to brands if it's not personalized.

Together, these two generations will take over the market. And so we must be ready to meet them where they are and exceed their expectations. In today's dynamic marketplace. It's a mandate for businesses to offer personalized experiences, dive into the data and you'll find a staggering 76% of consumers leaning towards brands that offer just a touch of personalization. And guess what those that personalized experiences typically see a 10 to 15% at least lift in revenue.

Amazon has more than or almost three decades of personalization experience. Some would say that we pioneered this at skill based on our learnings. We launched Amazon Personalized a fully managed AI service in 2019. It enables you to easily, accurately and quickly build recommendation systems. Note, I said systems and not models or machine learning solutions but a system saving you a significant amount of time and resources.

Amazon Personalized, fully automates the machine learning development life cycle building training and deploying your models, reducing your time to deploy personalized recommendations. And more importantly in today's market, optimizing your total cost of ownership, we offer real time recommendations and that's so critical in today's market because what someone browser was interested in yesterday does not equate to an interest tomorrow.

Again, this is a holistic solution. So we go beyond machine learning. We understand that being able to do this at scale, you need features such as promoting important um items, filtering out things that maybe were just recently purchased but would have been surfaced, wrangling data. How many of you have tried to on board or build a model and just the uphill battle with understanding your data, making sure you had a full data set that was fully proficient and accurate, took tons of time. So we solved that in January, it's easy to integrate into your existing tech stack with the simple API call and lastly with AWS, something that I always admired as a former customer. Before joining the company, we have a comprehensive suite of security and compliance controls. We set that bar extremely high. That means that leveraging a service like Amazon Personalized, you get that rigor and discipline and it helps you stay compliant and ready when the auditors or someone is coming to make sure you meet or exceed that bar.

All models are private. We never share your data, they're only used for your recommendations. And again, we do not even share it with uh outside um entities beyond AWS.

So since launch, we have continually continually evolved to cater to our ever changing customer needs and demands. Today, we provide a comprehensive suite of six foundational use cases.

First user personalization, personalized products or content recommendations tailored to a specific user profile are more likely to hi to drive higher conversion rates related items. Also known as similar items, recommendations should not fall victim to a popularity trap. Similar item recommendations help user discover related or comparable existing items in the catalog to help with cross sell upsell and conversion personalized ranking. Business priorities require you require you to promote content or products such as seasonal merchandise or things that are time bound. Amazon personalized enables you to um the ability to re rank your product catalog or highlight something that users would find highly beneficial all while balancing what's important to you.

User segmentation breaking through users means targeting the right message to the right user at the right time. Managing segmentation with rules can be hard brittle and a lot of upkeep with machine learning. We dynamically evolve over time and we help you avoid over messaging trending. Now recommending items gaining popularity in a fast paced among your users. Again, something very important for the gen y and gen z demographic and popular items, recommending the items that are most popular based on interaction data.

Now, as I already shared, we go beyond just basic machine learning. We have, we have provided our customers with a comprehensive set of features that helps you out of the box build a highly personalized solution starting with real time recommendations. I won't belabor this topic but it's hard to do. I remember prior to joining AWS, I spent about 14 months building an omni channel personalization solution. My predecessor had already spent a year and a half and we were able to do it for a small population, a successful POC and it was offline. Now try to inject building a real time recommendation solution that has the proper latency latency mitigation. And most often companies can only get to near real time. It is frigging hard, cold start items. The challenge of delivering relevant recommendations, especially new items in a rapidly changing catalog can be hard and we call that the cold start problem. We address this challenge by balancing new con content exposure and relevance through an exploration, weight, ensuring new items find their way to the to your end user quickly and effectively business rules and metrics optimization.

Amazon personalized balances, optimum recommendations with your business goals. We allow you to work back backwards from what is also important to you. Take click through rate revenue, uh watch time. You can set those parameters and optimize your recommendations based on that. Again, we also allow for promotions, textual information extraction, important details could inform recommendations and this was often trapped in unstructured texts such as product descriptions or product reviews. So a few years ago, we launched a feature that uses natural language processing to extract a vital textual information to generate more precise and relevant recommendations.

Already mentioned, data insights, data wrangling. In January of this year, we launched a feature that allows us to assess your data after you upload it or stream it and determine if it's comparable or viable. And if it's not, if we see opportunities or anomalies, then we give you gentle nudges of things that you can work to improve um improve the data set. This helps with time to value, it helps with speed to market. It helps you resolve that troubleshooting much faster and quicker. And it's honestly one of our most widely used features, believe it or not.

Domain recommendor, we understand that certain do um industries and domains have specific needs. And so while we are in industry agnostic service, we invested two years ago in specific domain recommender for the retail industry as well as the media and entertainment industry to again accelerate time to value rapid experimentation. Again, we are very obsessed with making sure our customers can derive the most value at a high effective rate very quickly. And um this past summer, we launched a feature that allows you to accelerate your experimentation by adding new columns to your data set without having to start over or reproduce a resource. And that has also been a game changer for our customers. And then last but certainly not least impact measurement. It's all about how do you assess your return on investment. Amazon personal personalized allows you to uh input an event into our system or multiple recommendations to assess the impact that we could have on your uh output. It allows you to have a data driven strategy and make sure you are able to get a return on your investment.

Our commitment to innovation does not stop. We are relentlessly customer obsessed. In 2023 based on customer feedback and industry trends, we introduced three exciting new use cases that take personalization to the next level.

One. We launched in October which is personalized search. Next best action just launched on Sunday night and we just launched three new generative AI capabilities that I'm so excited to share with you today.

Now let's dive a bit into each of these and help you understand the power that we are bringing to you to create hyper personalized end user experiences.

Search is a crucial component in engaging site and app users as it brings high intent traffic from individuals seeking specific products or categories. Studies have shown that over 30% plus of website visitors utilize the search bar and they do that to locate specific products and content. We also see a 5 to 6 times higher conversion rate for an effective search experiences versus those customers that don't engage in search.

Enterprises want to improve the relevance of their search results. Previously, customers found it very difficult to provide relevant search results to their users due to lack of machine learning expertise or infrastructure scaling limitations. In October, we launched a new integration between Amazon Personalized and Amazon OpenSearch. OpenSearch is a service used by customers to power their search experience. It's also a fully managed service. This allows enterprises to increase their conversion and engagement from their site search.

This integration allows Personalized to offer a highly proficient machine learning based search experience. That's hyper personalized. Enterprises can boost relevant items in their search results based on different types of use cases that account for an individual's past interactions or contexts. Thus improving accuracy.

For example, let's take an automobile e-commerce use case, let's say for um start for starters, a customer goes on a site and puts in a very specific search criteria. Mazda CX 30. In that situation, the search is pretty easy. It's going to be able to return because it's so specific by mazda cx 30.

Now, how often are search queries or search inputs that specific? Not a lot? So now let's say you go on to that same ecommerce site and you simply put in toyota and putting in toyota, you could get a slew of responses. But with something like personalized search, we take into account your preferences, your interactions, we take in to account metadata and we are able to return hyper personalized search results drastically improving the conversion and the engagement of end users ultimately impacting revenue.

Next best action. Something that like I am so excited about because I think it's a game changer. It goes well beyond basic recommendations. With the proliferation of digital channels and technology advances that make hyper personalization more accessible. It can become overwhelming for brands to determine what marketing materials such as messages or actions will maximize engagement for each individual customer.

Traditional customer segmentation approaches are no longer sufficient as every customer expects a unique experience and when they don't have that, they typically abandon brands. Just sunday. We launch next best action. I'll use NBA from here on out. A new model that allows you to determine the best action that you should offer your individual end users that will help with engagement and increase likelihood of conversion and deepen loyalty.

Actions are any task that a user can complete to engage with the brand with NBA. You can recommend actions such as adding on a service, downloading an app. um enrolling in a loyalty program. NBA empowers you to simultaneously reduce marketing clutter and noise by tailoring the right action to the right user maximizing your revenue potential. And one important thing did i mention this real time? We believe this is a game changer. It will deepen the relationship again. with this demographic group i keep talking about because they will change the buying habits. they've already done it. gen y and gen z.

Now, what we've all been waiting for, i'm sure this thing that has taken over our lives. I was joking earlier today at lunch and i said i have an aunt. she's probably in her sixties and she had no clue what i did at work. she had no clue what AI was...

And now she wants to have a full conversation at Thanksgiving about what AI is. She even asked me, "Oh, do you think that was an artificial intelligent response?"

So generative AI is a form of artificial intelligence that creates new content such as text or images by leveraging something we call a foundation model which is, which is trained on vast amounts of data and the data allows you and that model allows you to adapt that foundation model or FM to multiple use cases.

The reason why generative AI has become so popular is because of the scale, the magnitude and the endless opportunities you can do with it in a very short amount of time. We are all excited about the evolution of generative AI. This is a transformative technology and it has not only reshaped the way we work, but it has also revolutionized how we engage with customers.

But generative AI alone cannot create hyper personalized experiences. On Sunday, we announced three generative AI launches that are tailored to different personas because we want to meet you all where you are.

First, we launched Amazon Personalized Content Generator, a fully managed capability that uses large language models to generate themes for a set of recommendations. These large language models or LLMs can identify the thematic similarity of a recommended item to use natural language to create a simple but highly engaging description which can be used for carousel titles, marketing, email, subject lines or push notification headings.

With Amazon Personalized Content Generator, our customers can augment their recommendations to improve content, discovery experience, product promotion and marketing outcomes.

So let's take an example. Let's say you've been shopping online at your favorite retailer for a black dress to wear to your best friend's bachelorette party and you 30 minutes later receive a push notification or email that says "We think you may like..."

Now let's rewind. Let's say you had been taking those same actions on that retailer searching for a little black dress that instead the subject line comes through and says "A toast to the little black dress, which one will prompt you to want to open that email versus simply delete it, which one will prompt you to want to engage with that brand versus simply ignore it. I think the second one would and that is the power of Amazon Personalized Content Generator.

A lot of customers also want the flexibility to build their own generative AI applications. As we showed earlier, Amazon Personalized can generate product and content recommendations, user segmentation. And we do that exceptionally well. The output of Amazon Personalized can be an ideal input into a large language model or an LLM to improve personalization in your generative AI solutions.

To facilitate this workflow, we now offer a streamline integration of Personalized with LLM frameworks such as LangChain. For those of you who are not familiar with LangChain, it's a framework designed to perform a variety of tasks that allow you to quickly and effectively build generative AI solutions into your end user experience such as text summarization, a chat experience, etc.

Builders can now get access to Amazon Personalized capabilities directly from LangChain. This will make it easier for you to add personalization to your generative AI projects while remaining in the same ecosystem.

Lastly, we have launched a return metadata capability to help you enrich recommendations with metadata such as pricing, rating and product descriptions. You can feed the enriched recommendation into LLMs to generate highly personalized content that will further boost your engagement for your end users.

To combine the last two generative AI capabilities we will release later in December, an upgraded LangChain integration that will combine LangChain and the metadata return for an even powerful experience across retail, media and entertainment, travel and hospitality, sports, profit, nonprofit businesses of all scale.

We can help you transform your end user experience. We are making advancements. We are providing real time recommendations and we help bring your vision to reality again. For every company small or large, we have over 1700 customers we work with and we are proud to work with you all to help bring your vision to life.

Equinox saw a 92% increase in engagement while using Personalized to recommend classes on their premier fitness app. PBS, a widely recognized public broadcaster in the US saw a 10 to 18 percent increase in weekly traffic across all of their properties. Chewy, our favorite pet store also used Amazon Personalized to re-rank items which results in a vast improvement in cart to check out, which means customers were spending more.

They said this was a substantial impact for a business like Chewy where relevant recommendations or sorry, where irrelevant recommendations could erode the trust with customers.

None of these advancements I talked about today would be possible without feedback from our valued customers. I'll now turn things over to Liam.

Liam: Yeah. So my name is Liam. I'm a senior manager here at Amazon Personalize.

So when customers engage, they expect personalization. So when they engage with the platform, they expect things to be personalized for them as Tia mentioned. So they're constantly also being bombarded with recommendations, those recommendations are coming from all sides.

So what many sites will do from streaming to retail is they'll provide a recommendation that again that's unique and it causes a customer to stop in their tracks. Look at that recommendation and engage with it.

We feel that Amazon Personalized combined with the power of AI can help our customers provide that type of hyper personalization for their customers.

So as Tia mentioned on Sunday, we launched the Amazon Personalized Content Generator which utilizes generative AI within the service to create unique themes based on a seed item.

So for instance, based on the title "Something from Tiffany's," which is a romantic comedy from Amazon Studios, we already had that capability to say if you like this title. Here are other titles you may also like or more titles like that.

As Tia mentioned, will that get you to engage? Possibly not, it has a good place on a site such as when you're done watching that one title, maybe here's more titles to watch.

However, with the Amazon Personalized Content Generator, you can take that seed item, generate a set of recommendations and give it a unique theme. So similar to the retail example, this theme that came up with from these five titles is "Love, Laughter and Hijinks." So it's designed to be witty and engaging.

And also worth noting is the Amazon Personalized Content Generator feature can reorder the titles based on their applicability to that theme. So that's designed to be engaging.

Normally, this is done very manually. So prior to this, you would have creatives that would take the the titles, create a theme, fill the rail or the carousel with that titles that match that theme and put it onto the site.

Now, what you can do is you can redirect that effort to review the large list of themes that are generated from Amazon Personalized, establish whether they make sense, check the titles to make sure they're relevant for the theme and really then take what was one rail and create dozens if not hundreds of micro genres.

So in media, these are often called micro genres. So everyone's familiar with "action comedy," micro genres are these really specifically unique sub themes. So the features we launch definitely can help generate these themes.

One caveat I didn't want to mention is all of the demos and slides I have today talk about media and entertainment use cases. As Tia mentioned, the features we launched are generalizable across many use cases from retail to travel and hospitality and more. So I'm just demoing in the media context.

So general AI makes it possible to interact with users in new and different ways, how we don't know exactly how you might want to use it. So it could be using chat bots that can then retrieve information from a knowledge base, create marketing content, be it copy or imagery that's now specific to your users and their needs or conversational search where you're asking questions about what you're looking for and then iterating based on that answer.

However, gene AI lacks the context of what the user preferences and behaviors are. You could pass that in as context into the prompt, but passing the entirety of what a user's interests are in a single prompt becomes challenging and also could be expensive.

That's where the power of Amazon Personalized comes in. So, Amazon Personalized provides a scalable low cost way to generate recommendations across an entire user base and item space. So you can create a large set of recommendations, get them ready, use them in real time in a large language model scenario.

So that could be from user personalization, which provides the most relevant items in real time to a user's behavior. What's trending now on a site to the best search result again, many more use cases, but this is just different ways that we feel can combine.

However, we're not sure how you might want to use it, which is why we launched the return metadata and LangChain feature. So let's dive a little bit more into this feature.

So Amazon Personalized is an API based service that provides a response based on an ask. So what we have here is the Amazon User Personalization API response. For simplicity's sake, I'm asking for just one recommendation. Generally you get 25 you can ask up to 500 recommendations.

What you previously got was you got the item ID and you got the relevant score. So this is how relevant is this item for this user? However, that's not a lot of information to pass to a large language model. There's nothing there for the large language model to grab onto.

So I'm gonna have to make a call out to an external service, get that item identification, get the description, get the plot, get the awards, it's won whatever I need to generate my output.

So what I've done here with the metadata return feature that we launched is I'm now saying I want the item, I get the item, I get the score which I've always been able to get. But give me the title, give me the description, give me the plot, give me everything I need to know to provide to a language model.

So how do you use that in an application? Let's go into an architecture. So let's start with the general personalized architecture.

So as mentioned, Personalize is an API based service that trains models on your behalf based on training data. So this data is item data. So what you know about your content or your item catalog? So this is the description, the plot, the awards, all the stuff you know about your content. This generally comes from a content management system and or an item database.

Then you have the interaction. So this is what you know, what your users have clicked on, what they've watched, what they've added to the cart, what they purchased, what they favorited. This is all of the interactions. This generally comes from a analytics platform or a customer data platform.

Then optionally, you also have user data. So this is your user demographics, what you know about your users. This could be subscription level user demographics. Any information that's all then used to train an Amazon Personalized model which generates a real time API or batch inference that can provide recommendations in real time as needed.

Using this new feature, you can now take these recommendations, pass them to the custom chain in LangChain or any other orchestration system, bring that information in which includes a prompt template. So from a generative AI perspective, this is what are you trying to achieve? Like what application are you building? And I'll show an example in a second that can also include what information you know about the items you're recommending, build that all into a combined prompt. Pass that to Amazon Bedrock for instance to generate an output.

So in this scenario, let's generate a 1 to 1 personalized email for a customer. So let's go into that and actually pick apart what that looks like from a prompt perspective.

So I have the prompt, I've simplified things you don't have to read everything here. I just want to give an example. So in the prompt, I'm instructing the model or the large language model. What I want, I want a marketing email that gets people to engage with my platform and watch the titles. I also, I'm gonna provide you the titles and some information about it.

So you combine the prompt with the returned metadata, which is the titles, the descriptions, everything you know about a title and the output is a 1 to 1 personalized email.

So now I've taken that and I've generated "What you want to watch" as an engaging title. It has the full body of it. This can be 1 to 1 personalized depending on your use case. Though you might not want to do 1 to 1 personalization. That's where the user segmentation feature in Amazon Personalized comes in where you can generate hundreds of segments and then customize to those segments using the large language model.

So the end result, there was tailored recommendations via an email that gets people to engage with your platform. So let's go into a demo. I'll show the console how to build it and we'll talk to how to use it in an application.

So what I have here is the Amazon Personalized console and in the console, I have a set of data set groups, a data set group in Amazon Personalize is all of the data training to build a recommender system. So in that data set group, I have two different data sets that I've uploaded the data and I provided the input.

So I have a combination of two things which is one the items. This is everything I know about my content. So for this demo, I'm actually using the IMDb data, which is available on Amazon Data Exchange. Fantastic set of metadata about my titles, movies and TV shows.

I've then also trained a popular items model or I got I've got interactions, sorry as well. So the interactions of what people have clicked on and what they've built. So from that, I've trained models.

So in Personalized terms, a solution is the trained model and the surrounding capabilities. So I the similar items model is the similar items and also is the heavy lifter of the recommend of themed recommendations capability that we launched.

So I've also created a popularity model. So in my catalog, I have 18,000 movies. I don't want to generate themes for all 18,000 movies. So I've created a list of the top 100 popular items using that popular items recipe.

"And I'm gonna now use that as input into a batch inference using the new feature. So I'm gonna start a batch inference job and I'm gonna use the themed recommendations capability with the content generator.

I'm gonna give it a name again. It's the top 100. So I'm giving it the top 100 as the name. I'm gonna use this similar item solution that I've already trained and I'm gonna get 10 responses.

Uh just to make it easy to read, I'm gonna give it two bits of information, the item name and the description. The description is the plot and the item name is the title in media terms.

Um the plot is the unstructured text feature. So it's automatically detected and then I'm going to give it that list of 100 titles. So now I'm providing the input JSON, which is the 100 titles and an output location. So this will generate 100 themes.

Um one thing to note though is similar items may not always be similar, they're coin interacted. So I'm actually implementing a feature which is our current item filter to actually look and say does this have the same content, maturity rating? So PG 13 R and doesn't have the same genre. So this is a feature we launched to help really make sure the similar items match in similarity.

So I'm going to create the output. So let's pull up the output and take a look at what the output includes. So what you're gonna see here is the JSON output, which is the seed item. So this is my seed item that I generated the similar items from and then the the reordered list of similar items and the theme that, that those 10 items match.

So crime and mystery is the theme here. Um one other key point is you get the relevant scores. So this is how relevant are these items to the theme that I've generated? So I highly recommend using human reviewers to validate these titles and their applicability to the themes and the themes, applicability to your content and your needs.

So let's now take this into a fictional streaming site. So what you have here is, is unicorn flicks, which is our fictional streaming site, similar to one you'd see on any connected TV or Amazon fire TV device. It has a lot of things that you'd come to expect such as top picks for you.

So this is the Amazon personalized user personalization recipe at work here. So based on my recent activity, here's the most recent set of titles that are interesting to recommend to me, I have the trending now feature which we launched um which is just what's trending now on my platform in real time.

And then I have the type of things that you would come to expect from any site. So comedy, action drama, all the genre based carousels or shelves as they call them.

Now, I've created this custom um themed recommendation. So this is a micro genre rail. So this is classic, classic literature's tragic love story. So this was generated from the Amazon personalized content generator and it matches my my interest.

So I've been watching a lot of period pieces, a lot of historical dramas so this is a pretty good set of, of micro genre titles to recommend to me, I'm gonna change my behavior now and I'm gonna start clicking on other titles and really start engaging with different titles.

So let's watch a bunch of documentaries. So we'll go down and click on some documentaries. So every time I send a click in, I'm technically watching it, I'm sending the event. So man on wire.

So what you see here is the similar items that it's normally presented to a user. So if I like man on wire, here's, you know, king of kong spellbound other documentaries i may watch. So let's click on a bunch of documentaries.

So clicking and clicking and clicking, I'm watching documentaries. So now that I've watched four documentaries, let's do a refresh and see what set of micro genre titles.

So now I get captivating stories of life, art and music. So this is following my behavior. So it's changed. I now am interested in documentaries pretty good set of recommendations.

Now let's watch a bunch of big Hollywood sci-fi movies. So clicking on the most popular trending Hollywood sci-fi movies, watch four or five of those and then we'll do a refresh.

So I'm also using the re rank capability here. So I'm actually re ranking the results um which is how these rails are changing based on my behavior.

So now I get out of this world adventures. So again, this is designed to get me deeper into the catalog. I've watched everything the popular site has that's brand new, but they want to keep me engaged. So get me into these deeper catalog titles.

So let's use the return metadata feature now. So in my fictional site, i have this feature to chat with the site so i can ask for a set of recommendations. So let's say recommend me some drama movies.

So using lang chain, you can actually use this at to call out to Amazon personalize with the filter to say give me recommendations for this user specific to drama.

So I'm gonna get back a set of titles and their associated metadata that are drama. So i now can then engage with these titles. And i can say, well, i i know i asked for drama but i want a funny drama.

So are any of these titles funny? You can again just have a conversation, dig into these titles, get information. So we can say are any of these funny? In this case, none of these dramas have comedic elements.

So I'm gonna go back to the, well, I'll ask again this making another Amazon personalized call, getting another set of recommendations. Now, i think maybe i want comedy movies.

So now that I've got these movies, i can ask about the movies, i can say, oh these are, you know, these look like a decent set of recommendations. Let's look grand budapest hotel. Let's ask about that who directed grand budapest.

No, it's not a keyword matching. I'm actually asking generally about the content, not spelling it correctly. Maybe shortening the title. It's able to tell me. Yes, wes anderson directed this movie. Good.

Did it win any awards is something i should invest my time in watching. So you can then go back to the metadata that was provided and query it and find out yes, it has won quite a few awards.

So that is the generative aI kind of integrations. There's many different ways from marketing emails, how you use it is really up to what integrations you build, why we build this type of integration with that.

I will hand it over to Darryl, right? All right. Uh my name is Darryl, but I'm the EVP of technology platforms for Fox. I oversee a number of different teams that of this conversation is our PNR or personalization and recommendations team. Make sure i have the right one there don't want to drink leaves.

Um their focus is really using personalization to drive usage and revenue across all of our platforms. Uh because that personalized experience is really key to our uh customer experience uh across fox, we have a number of different brands.

So Fox Sports, two B, Fox Entertainment, Uh Fox local stations, Fox News Media across those two B is the number one AVD platform in America. So the largest advertising supported V solution, obviously, Fox News, number one news channel.

So we uh we have a lot of scale. Uh you look at our monthly active users, about 300 million monthly active users. We do our largest events, things like the Super Bowl. We've seen uh well over 6 million concurrent streaming viewers at that point in time and pretty much every second of every day, we're collecting about 50,000 data points per second.

So uh what we do uh we do at scale and uh the reason we do all of this is what we call total view time, other companies might call it something else. So minutes spent whatever it may be, but really total view time is kind of the core metric in our business.

Uh it's essentially the total amount of time that people are watching your content. So no matter what your business' modernization strategy is, subscription video on demand at support of video on demand companion advertising, really, the more time you have people watching your content, obviously, the better it is you can sell more ads against that.

If it's a subscription service, the more you watch, the more likely to be happy, the less likely you are to churn out as part of it. So pretty much every revenue strategy comes back to TVT or total review time.

And like i said, we have a number of different business units and that proposes a little bit of a challenge for us because we don't have individual technology teams that do everything across all these businesses. Certainly some things are bespoke, but we've been working with Amazon for quite some time and you can see here we're using a ton of their infrastructure across that, what we do across much of our business.

I call it a few things here. Our media cloud platform is something that we use to all the video pipelines and coding, everything we do for both live and VOD assets across all of our businesses.

Spark our CMS and our PNR team. All these are built entirely on top of Amazon web services solutions and we're using them in different ways across sports entertainment station news, kind of all of our different business units.

We think about our journey with AWS as it pertains to personalization and we look about eight years ago, um freeze it around. This was uh you know, we were kind of starting to get in the streaming business.

We'd launched our Fox Sports go apps, Uh Fox news, Fox entertainment FX and everything. At that point in time, we were trying to bring personalization into that experience. At that point in time, it was pretty limited the offerings on the market.

Uh there were a few companies out there offering some pretty basic things, just Uh basic item similarity modeling type of things, Uh collaborative filtering models, nothing that was too advanced.

Um and there was some open source options out there. The Uh the office shelf ones are actually quite expensive. So we decided to go with an open source solution.

I can say what that is now is prediction IO now it's in the apache attic so i can say it out loud here. Um we went with that. The problem was we didn't know what we didn't know and that lack of experience led to a lot of bad decisions.

Um we did real time modeling for literally everything which gets really expensive in terms of the the service request and how long those take to go round trip, Uh the infrastructure we had to manage.

Obviously yourself on that side of things it wouldn't support EMR. So doing your elastic mapreduce to be able to get the data into a good form, wasn't something supported by that platform at that point in time.

Um and it turned out that that free open source software was actually quite expensive. It had a pretty negative ROI for us. So after a few years, we were looking for other options and other places we could go.

Uh 2018, I'm sure many of you in this room are familiar with SageMaker. We started exploring that and working with it for some of the item similarity models and some of the the easier things that we could pick off that point in time.

We luckily gained some chops in the data science area at that point in time, but it was really in 2019 at this very conference. Uh when Personalize was announced that we started getting on board with that and working with our partners there.

Uh so since that time, we've been using it pretty exclusively with some SageMaker Uh to power a lot of the use cases that are most popular and driving a lot of the total view time across many of our different properties.

Um why we made that decision? Obviously, there's still other options out there at that point in time. Like i said, we go back a long way with AWS, we've been cloud native for over a decade and with AWS that entire time, all of our interaction data for all of our users was already stored in Amazon.

So all the event data that we were capturing was already there. So everything we need to feed into those models already existed as well as all of our content metadata through the aforementioned media cloud platform that we had there.

Um obviously going to a managed service means that we don't have to spend our time managing the infrastructure and everything else. It turned out to be a challenge for us and our lack of experience in that area.

And obviously, we had the team supporting us, you know, they're building these recipes for the media and 10 and verticals"

Obviously, there's ones for retail, many other ones, but it allowed us to work with someone who'd really seen these at scale across different competitive companies for ours and it could recommend what was really going to be important for us. And so it really eliminated most of our weaknesses, you know, the real time modeling stuff that wasn't necessary at that point. We were bringing in way too many data points. They weren't actually driving anything in terms of predicates, uh high latency infrastructure, all those kind of things. So, um, working with AWS was an obvious choice for us at this point.

And so I'll talk about a couple of ways that we actually implemented it and the outcomes that you'll see here, you can see these giant numbers here. And I mean giant in every way. Um in 2019, the, uh the first time we launched personalized uh was in our Fox Nation app, which is a subscription video on demand platform, part of our news vertical, but it's really kind of a lifestyle app, do a lot of uh reality programming, true crime, uh faith based content, things like that.

Um and the most powerful piece of real estate that we have in those apps is what we call the end card. So if you've ever been watching Netflix or anything else, uh you come to the end of something and it says, hey, the next thing you're going to watch should be this and that's really easy if you're watching episodic drama, it's like I'm watching season two, episode four, like you should probably watch season two, episode five. That's a really easy thing to do. Uh but a lot of the content that we have, there's no kind of obvious. Next thing we don't do a lot of episodic dramas. So how are we going to use that real estate to keep people watching content and not having to manually search for things?

And so when we launched the models and moving away from prediction io and using personalized to do that, uh we recommend we, we saw an increase of 45% in terms of all minute stream coming from that end card. And even more importantly, we saw a 15% reduction in the bounce rate, which means that people were more satisfied with what they were seeing on that end card, which is obviously huge for us.

So I'll show you what this looks like here. This is uh Abby Hornacek show. That's her dad, Jeff Hornacek, uh a uh former great Utah Jazz point guard. I'll show you how it looks like here.

That was really fun. And it's, I can't wait to watch the rest of the documentary. Can I go? You can go clean your garage and I'm sorry for agonizing you about the loss in the NBA. I will walk in. Go ahead.

So you've watched this. He's very sad, obviously for never win the NBA finals, but you come to the end here, you see two recommendations. We're using personalized for both of these. The first is an item similarity type model. Uh because again, there is no kind of obvious next uh episode in that series. So the world of TikTok with Casey McDonald, so something really related to that particular piece of content you were watching just now.

And the second one is something that's actually using your user interaction context a lot more. So this one, I was just kind of clicking around to create something. This is a documentary about Hunter Biden. I've actually never watched it. I'm sure it's whatever. Um but it gives the options there essentially so that we have personalized experiences for you and more of what you're already watching there. That's what drives those 45% increases. Like that's a huge number for us because again, going back to TV T, keeping people watching video constantly and not going back to a street and searching for things is the best thing we can do to drive that forward.

I'll take you through a second example, that kind of shows the flexibility of it for us, which is our Fox Sports app. So Fox Sports really couldn't be much more different. Uh Fox Nation is something that's very much uh video on demand, long form content. Uh Fox Sports is almost all about the live event that is going on at that point in time and recommending content to a live streaming audience can be really challenging

um because everybody ends the game at the exact same time and so you're dealing with thundering her and you're dealing with a latency coming back from that and it can be a really unique challenge and that's where the scale of personalizing their services have been really powerful for us.

And so, uh we used to have an experience which I'll show you in a second here where basically you would come to the end of an event and it would just say thanks for watching. That's great. We'd like to thank people. That's nice. Um that's not doing a whole lot for us. Um by using personalized and bringing people into personalized recommended content. We were able to recognize this 400% increase in content starts.

We actually launched this during the women's World Cup this year and it hadn't immediate and huge impact here. And now those recommendations are number one in average minutes viewed coming out of the end of those events. So I'll show you what it used to look like here. I actually picked this for Kevin. So let's go ahead and watch this one. This is Tim Brando saying we'll see you with Texas Houston next Saturday on Fox. It's like the worst case scenario right here. I mean, great because I lost but the worst case scenario because at the end of the content here, you're stuck with a dead end, which is a place we never want our users to be. They're stuck here looking at something that's saying thanks for watching but there's no obvious next step for them to take no obvious next best action there.

And worse than that, I can't monetize that. And even worse than that, it cost me money just to keep delivering this late to you over the CD n. So there's really no upside for the, for us in this. So now by doing this and like I said, we did this during the women's World Cup when we launched it, you now go into a situation like this crew. This is Tim saying we'll see you with Texas Houston next Saturday on and for Marcus dere at full back, Jim Austin Jones in the backfield and here goes, Caleb Williams go deep young man touchdown on the right j do. That's one to watch all the way through. It's a great moment early season before things went horribly wrong.

Um I will admit I edited that because I really wanted to do that to Kevin as part of this. There was an ad that existed between these things. So normally when you come out of when you come, at the end of event, there's actually a 15 2nd pre roll we give you and then we go into tailored highlights, but this was actually recommended to me because I watched a lot of USC contents where I went to college. And so I actually knew that I would be interested in that.

So number one, it's driving monetization, it's pushed me into pre role. It's actually able to monetize me after I finish the event there. And number two, it's a never ending loop. Once this goes, it goes on the next one to the next one to the next one. So it keeps you in that experience, which is always driving that total view time up. And like I said, still the most important thing we do.

And so what's coming up next for us? There's a few things, generative AI, I'll talk on the next slide here about a few of the things we're investing in and that we're working on right now to create hyper personalized experiences using generative AI lang chain integration as was mentioned earlier by I would say my esteem colleague, but we're not colleagues. We're just great partners to you the lang chain integration with personalized and the amazon personalized content generators. I'll talk a little bit about what we're going to do with that as well.

And obviously for us at scale, I talked about the nation example, the Fox Nation example, the Fox Sports example. We also do this for Fox.com and the Fox Now product. Obviously, there's other business you'd like to bring that out to, to really show the flexibility of what it can do for us.

So I'll talk through a couple of these next cases right here. The first one is the conversational recommender. So kind of similar to the demo that Liam showed earlier. We've been making some pretty big investments in generative AI internally. Um and the work that has been released this week actually allows us to kind of bring that together with the personalized services so that we can uh bring all together something that is unified for end users.

Um on the Fox Nation product, we often have kind of themed months. So it might be true crime month or a faith based month or, or whatever it might be. And so all the marketing goes around that everything you see the inbound communications, they all are kind of tied to that theme. And in the past, that was a number of people's jobs have to say, ok, well, it's faith, it's faith month. We got to make some editorially created collections that are going to be really tied to that thing that we're pushing out there. So they go to this library of thousands of assets and try to find things that really fit that based upon the metadata that we knew about everything. And certainly it works, but it's a pretty tedious job and we don't really think we're getting the best selection of assets.

So what we've already built actually for these editors and something we're gonna be uh we're gonna be expanding with some of the work coming for personalized is something that allows them to use a conversational interface or a chat like interface and say, create a collection of 20 of 30 whatever you think the right number of collection is of a true crime content with a faith based angle and it'll bring back to them, obviously, all the content that relates to that but not only bring that content, but bring back the content they think is going to be the most successful for that quantity of content. And in the order that they believe will be the most successful. So that when we put it in front of our users, it's most likely to have the biggest impact on them. So really excited about that, probably not as excited as our editors are, that's huge for them.

Uh the second thing is a marketing email generator. So Liam touched a little bit on this earlier, you know, when I talk about TV T total view time, the worst thing for us is when you turn off the television, you're not watching it anymore. My whole job is to get you watching it again. And so really what I'm trying to do is find any way I can to say, hey, there's great stuff for you here, you should come and watch it again. So using the capabilities, we're working right now on building that marketing email so that we know what content you're most likely to interact with so that we can create those emails and say, hey, here's some content we think you'd be interested in, we can decorate and make it very personalized as part of that. But also we can send it at a time. We think you're most likely to actually interact with it. So by using all the capabilities unleashed by what person has announced here, we're really excited about that's going to drive forward for us.

So, you know, I think it would be an understatement to say that this generation of AI is moving incredibly fast, obviously, uh very, very fast last week in particular. Um and I, who knows exactly where it's going, but obviously having a partner like AWS, uh to be with us and work through this with us, uh allows us to be nimble and maximize our efforts at Fox. So, uh a huge thank you to them for that.

And uh with that, I will welcome back to the stage to you.

Thank you, Darryl.

Ok. So a couple of things I was talking too much and going over my time. So one thing I wanted to highlight, um before I get into this close out slide is that for next next best action, we have a launch partner crea, it's a global boutique consulting company across all the continents, right? And they've already been trained on next best action. They've already been trained on personalized. So if you want to accelerate your adoption of next best action, please reach out to them. They are a premier partner with AWS and they can help you fast track your adoption of that.

In addition, we have a session on Wednesday um November 29th. And it's a workshop. It's a hands on workshop that brings some of our most brilliant AWSs to you to be able to enable and understand how you can personalize marketing content with generative AI and Amazon Personalize. You just heard Darryl, how important that is for Fox. So you can take that back to your end users and your teams.

We have our product detail page, we have our documentation, we have the QR codes for that. And again, we have our QR code intentionally for partners because I just mentioned and others are phenomenal partners that can help you accelerate your journey.

Thank you. Thank you. Thank you from Liam Darryl and I that's why you get three. Thank yous.

Um our LinkedIn QR codes, please please complete the survey and give us glowing scores because we've been awesome, but we really appreciate your time today. I know you could be a million other places almost 6pm this time you probably have dinners and events to go to, but hopefully you got some out of this. Hopefully we have enriched you and we have triggered your mind to think about the next best action you are going to take for your company and how you are going to leverage Amazon Personalize and generative AI, whether it's the managed capability Lchain or metadata to create profoundly enriched hyper personalized experience for your end users.

I hope you have a great day.