Architecting for Trusted, Real-Time Decisions

THOMAS KOEP: Welcome everybody — everyone's settling into their food coma, get some coffee. Welcome. This is the first session after lunch and we're going to be focused on architecting trusted real-time decisions. A lot of what you all saw this morning is built on this framework, so we thought we'd do a technical deep dive and a demo — everyone loves demos — for Christian to cover actual features and functionalities. My name is Tom Koep, VP of Customer Strategy, from the morning session, joined by Christian here, Steven from AWS, and we're super excited to show you all these new features. With that — let's roll.

CHRISTIAN SAFFICI: Okay. We're going to hope that the video started. Just to preface: I'm not sure how I'm going to track to my own video, so here we go. What we've done is build out a way to storytell around real-time experiences and how they drive personalization for customers in their moment — whether that's on a website, in a client-facing app for someone checking in on a reservation, spa experiences, whatever it may be. The whole goal is to power that experience with the profile data we saw today. Everything you see here is actually live. I could demo it live afterward if you want to ask some questions. But right now, Richard is going to be logging in. Richard is presented with this beautiful landscape of the city and some luxury accommodations.

CHRISTIAN SAFFICI: But the reality is Richard is a longstanding customer. We've seen Richard time and time again, we've been able to track his behavior and understand what he does when he appears. So what I've done is reimagine the experience for Richard once we're able to determine who he is. Richard came back to the site — he didn't log in or anything — but we've seen Richard before. So we use that behavior to drive personalization from the Amperity profile. In this case, we have his cookie ID or any ID on his identity spine. We looked him up when he landed, and now we're able to change the graphics and drive his experience based on any attribute in the customer 360 that we have access to.

CHRISTIAN SAFFICI: And so now — there we go. We can pick any attributes. If you think about how you build your customer 360s and calculate those insights, any of those behaviors can be used to drive the experience: curate specific hotels on the featured property sites, specific offers if you want. Now Richard feels like he's been heard. He's seeing that actual property he's stayed at before — not a generic cityscape, because he doesn't go to the city, he goes to a specific destination. And this is a reflection that all of Richard's behavior as he clicks around — whether or not he's logged in — is now streaming into Amperity. This is new for Amperity. All of the payload data, all of his interactions — clicks, views, add to cart, end session, whatever it is — is coming to Amperity.

CHRISTIAN SAFFICI: And now we're able to do two things. First, we can compute a set of attributes: did he add a dining reservation, did he add a hotel to cart? And then we can aggregate how many times he's doing these things. Yes, some of these behaviors are available today in your analytics platform, but think about how you can marry all of these attributes and insights to Richard's real-time behavior — combined with the facts we already know about him, his preferred channels, his preferences — and you now have a powerful data asset to enhance Richard's experience. To make sure he feels heard, feels welcome, and gets the offers and insights he's expecting. So as I click around and add a couple of different things, it should scroll in a second.

CHRISTIAN SAFFICI: All of the data is landing in Amperity, Amperity is doing some work on the backend. And right there — a property view count of one. Just to be clear: this whole panel is an admin panel to tell the story. It's not part of the website. It's reflecting the data, the profile, and the actions we're actually taking. As I continue to click around, the payload data is received in real time and it's now marrying itself with all the batch data — reservation data, OTA data — that's already coming in. And now we're able to use all of this to drive what to do next. So I'm going to go book a hotel and put it in my cart.

CHRISTIAN SAFFICI: So we know Richard's lifetime preferred properties are Summit Properties. Even though he's browsing this particular hotel, I'm not using the last action signal to drive an action. Maybe I'm using it to figure out what the next action should be — but I realize his lifetime preferred brand is Summit Properties. If it's Summit Properties in a specific city and he's looking at a hotel in a different city, I shouldn't be reacting to that or offering against it, because I know what's going to happen. So I did add the stay. I also added a reservation for dining, which is something he usually does. Again, you have the option to monitor and track that behavior.

CHRISTIAN SAFFICI: And part of what was presented this morning was the idea of segments. At the bottom of this side panel you can see segments. What's also happening on the backend is that not only are we using the calculated and aggregated attributes to understand the next action or behavior in-site, we're also using that to drive how to engage him next. Our real-time segment behavior is tracking what his preferred behavior is compared to his actual behavior, and then allowing you to build segments and real-time journeys against that. In this case, if his preferred hotel matched the hotel he looked at but didn't book, that can signal a segment and trigger a journey action.

CHRISTIAN SAFFICI: So to recap: all of the data from Richard's viewing dining, viewing experiences, viewing different hotels, lands in Amperity and we're now augmenting the profile with all of those behaviors. So all the questions — what do I do next, does he belong in a real-time journey, does he belong in a slower journey, how do I suppress him from existing or future campaigns, how do I engage him in paid media — all of that data is now at your fingertips. It's allowing you to figure out the proper way to engage Richard from pre-booking and booking all the way through to when he actually shows up for his reservation.

CHRISTIAN SAFFICI: Here's an example of real-time segments. Our abandoned cart segment has picked Richard up, knows that he didn't book but should have. And then you'll see our abandoned journey has done the same. This is going to load a simple journey, but in reality these can get as complex as we saw earlier. We can use the platform to figure out his preferred channel for engagement, preferred experiences, and reach Richard in the moment and in the way he likes to be engaged. I think I ended there — that was about 10 minutes. Spot on.

STEVEN M. ELINSON: Well done. Thank you, Christian.

CHRISTIAN SAFFICI: The takeaway is the demo booth — we can talk about this in more detail and you can ask a thousand questions. Feel free to stop by the demo area after this session or between sessions.

STEVEN M. ELINSON: Thank you. That was a very powerful setup, for two reasons. First, you really paint the vision of what we have for every single brand and everyone in this room today: by the end of 2026, you should be able to serve a bespoke website or mobile app for every single one of your guests who comes to your properties. There's no reason they should be seeing something generic that isn't personalized toward them. The second is that it's event-driven architectures that are going to drive us and ultimately allow us to deliver these kinds of experiences. Formally, I'm Steven Michael Elinson. I'm the Director of Amazon Web Services for Travel and Hospitality — a practice dedicated to the very specific needs of airlines, ground transportation companies, travel sellers, accommodations, lodgings, casinos, and cruise lines.

STEVEN M. ELINSON: One of the things I'm very fortunate to be able to do is explain overly complex subjects in ways that are easy to understand. I'll also pull the curtain back on a lot of what we do at Amazon as a whole, because many people are interested in learning how one of the world's largest organizations continues to innovate at incredible scale and speed. What you'll see today is that I'll present through the idea of mental models. We use them at Amazon as a way of more rapidly understanding a complex subject — ingesting that information and being able to use it, even if you don't have the depth to fully comprehend every detail. The second piece is that I'll talk a lot about durable truths.

STEVEN M. ELINSON: Think about amazon.com and three durable truths that really drive us. Our customers will always want a greater selection. They're always going to want lower prices. And they're always going to want faster delivery. Those three durable truths drive every single function or action we take inside of Amazon. The same can be true when I think about the travel and hospitality customers I work with. Their travelers and guests are always going to want their experience elevated to a higher level. At the same time, these are pretty low-profit-margin businesses. And so all of the C-suite leaders and boards I work with tell me they also need to improve their operating margins at the exact same time. So how do you do both? How do you elevate the guest experience while also improving operating margins? It's through things like event-driven architectures that give you the data and insights to inform both.

STEVEN M. ELINSON: A logical representation: we begin with signal detection, which is what Christian was talking about — how do you get all of those signals? This can be structured data, unstructured data, clickstream data. We've heard different terms used today: zero-party data, first-party data, second and third-party — things your guests or travelers give you directly. Behavioral observations you make through your transaction systems. Data you collaborate on with partners, or even third-party data you buy externally to augment the insights you already have.

STEVEN M. ELINSON: The key is being able to bring those signals in and ingest them at a rapid pace — that's where the term "streaming" comes from. Now, Amazon has more than 220 different services, and I'm only representing some of them here. But these are capabilities like Kinesis or Managed Streaming for Kafka — ways you can bring in incredible volumes and velocities of data. Essentially the same technology that powers the New York Stock Exchange and allows hundreds of millions of transactions to occur at extraordinarily low latencies. A key durable truth I'd share about this pillar: take a vertical slice when you're working to enhance the signals coming into your business. Let me translate that differently.

STEVEN M. ELINSON: We've seen some brands spend two years trying to get every single possible signal into their ecosystem before they can go and do personalization or revenue recovery — and that doesn't bring value to the business. So instead, take that vertical slice. Put together your hypothesis: what are the specific data points or signals you need to accomplish something for the business? Build out your pipeline, validate that you can actually be successful with those signals, then go back and rinse and repeat. It becomes very iterative and agile in nature. Once you have those signals, the key next step is to orchestrate all of your decisions.

STEVEN M. ELINSON: We have to take the data and turn it into insights. One of the most common things you do is standardize or normalize that data. Often you'll create a catalog so you understand all of the different attributes, where they are, and how your business can come to use them. Once you've cataloged that data, you also need to make it available to all of your resources and other tools inside your systems — this is where storage and databases come into play. A key durable truth I'll introduce here is what we call fit for purpose at Amazon. The entire world cannot be solved with a single relational database, though some vendors would like you to think otherwise. You need many different types of services — time series databases, graph databases, relational databases, key-value pairs — depending on the outcome you need to accomplish.

STEVEN M. ELINSON: So on the diagram: DynamoDB is for real-time personalization, allowing you to take a profile and do something with it immediately. S3 is our simple storage service — the most durable and economical way to store all of your data. This can become an expensive proposition over time if you're not storing data economically. And then Redshift, which you can think of as the enterprise data store, connecting essentially the entirety of the enterprise to perform all your insights, data visualizations, and business intelligence.

STEVEN M. ELINSON: A new durable truth we see emerging is that customers are re-propagating their information back into Apache Iceberg open table formats. A few reasons why. First, they're trying to avoid vendor lock-in. There are a lot of platforms out there that allow you to use their catalog, but that catalog is self-contained and can only be used inside that vendor's platform. It becomes expensive and it slows down innovation. With the advancement of AI — and yes, we had to use that word at some point — there are new technologies, new vendors, and new capabilities coming out all the time. You as an organization will want to experiment with them and take advantage of them. You don't want your data catalog locked away in a closed vendor ecosystem. You want it open and available so you can connect it to the entirety of your platform.

STEVEN M. ELINSON: Once you have those insights, the question is: how do you drive outcomes without adding additional complexity into your system? A key durable truth here is the idea of dual path. Primarily the conversations we've all been having today are about using Amperity to perform functions like revenue recovery — a traveler was looking for a flight from New York City to Hong Kong, they abandoned that cart, how can we get that signal quickly enough to have the opportunity to reengage that customer before we lose them to a competitor?

STEVEN M. ELINSON: We've heard about all the amazing ways Amperity allows us to do all of that. But the key point is that Amperity is also extraordinarily open and API-driven — and as of this morning, we've now learned about Model Context Protocols as well. So you take all the attributes Amperity has created for you and then put them back into that enterprise data store. The reason: today when I speak with customers, they say they have three different versions of who their traveler or guest is. Marketing has one idea, operations has a second, and customer care has a third. The future requires that we eliminate that and have a single version of truth. Although the revenue recovery pieces we're talking about today are important, it's this second piece — the unified truth — that's going to be most important in the agentic future that's coming.

STEVEN M. ELINSON: Think about the agents we're building to automate massive volumes of transactions. Your agents are going to need access to that single version of the truth — a real idea of who the profile is, where they are in their journey, customer sentiment, and everything you've ever known about them in the past — so that the agents can take those actions on your behalf.

STEVEN M. ELINSON: And I want to share that this is real, not theoretical — because we always have to prove our work. That's an Amazon thing. So I'll let you read through most of this slide. I'll focus on the bottom where I have my own experience. Lotte Mart is a South Korean hyper-retailer. By moving to this event-driven architecture, they were able to increase their conversion rate on personalization by more than five times — because they actually knew who the customer was, where they were in their journey, and how they could activate them. I was on stage two years ago at Amperity with Dane Matthews, the Chief Marketing Officer from Taco Bell. Taco Bell gets a large uptick in sales during special events like the World Cup or the Super Bowl, and also a really massive uptick when they introduce limited-time offers and special menu items. Dane tells the story about when they reintroduced the Mexican Pizza — they saw four times the sales they normally see in a single day. So you also need a system that's elastic enough to scale up to handle that volume, but also elastic enough to reduce back down so you're not inheriting additional operating costs.

STEVEN M. ELINSON: And last but not least, Expedia — one of our travel sellers. They have a slight advantage over some folks in that they're more digitally native: no physical retail locations, it's entirely website and mobile app driven. So it's all clickstream. But Rick Fox, who I've worked with quite a lot, talks about the fact that they have 100% coverage over the entirety of their data pipeline. So when he talks about being able to give information to his channel managers, revenue managers, distribution strategy leaders, marketing officers, and the rest, they have 100% coverage and can access all the insights to drive the outcomes. We know the architecture works — it works at scale, and it works for some of the most loved and well-known brands in the world.

THOMAS KOEP: Now I get to go. Thank you for walking through that. I think it's critical to understand features and functionalities alongside actual use cases. Let me dive into how this works. Most likely, what I'm about to describe is what you currently have — or what I had as a customer for years. Most folks have batch streaming services in one event-stream cycle, and a data warehouse where they're batching all their upstream systems, running identity resolution, getting segments out, and sending that to other areas. But what's interesting now is that we've gotten to the point where we merge those two. You don't have to have two distinct systems — one middleware, one event stream, one batch. You can run it all directly through Amperity, putting all your known and unknown data together.

THOMAS KOEP: The interesting thing going forward: you can now use all those unknown signals, bring them into your tenant, store them over time, and the second a guest logs in or provides a piece of PII, you unify all that historical browsing behavior to an Amperity ID. And the benefit is that the Amperity ID is already your record of truth. You can bring that together with all that historical context and start to use it in real time, creating journeys that don't have to wait for a micro-batch process or a daily refresh. You can create profiles through all of those unknown signals and treat them exactly the same way you would have with a batch process to create profiles in the platform today.

THOMAS KOEP: It gives you so much more flexibility. You're running all your event streaming, owning all of those pieces of context about your customers, off the foundational identity resolution you're already using for your business. The fun thing is you essentially stop relying on stale information and run everything through subsecond profiles — bringing that together with all your other information. And it eliminates the window of time where someone might choose another brand because you didn't market to them as fast as possible. The last point is that it's universal and open. We can now bring in all that information for different applications — it's all API-based. So you can remove a lot of different point solutions and start to build your tech stack the way you want it, without being beholden to a stream service here, an event processing service there, and batch systems feeding into them separately. And for customers, there's real efficiency gain — not just in time and service, but in MarTech spend. Audit your tech stack and ask: do I still need event processor X and event processor Z? Because you can take that money and reinvest it back into the platforms that hold all your customer information in the first place.

THOMAS KOEP: A few guidelines on key capabilities. Using the web and app as examples: event streaming lets you bring all that information in and manage large-scale datasets. Web information is massive, but we have the speed and processing ability to bring it all in at the Amperity ID level. The other is unifying all your real-time and batch processes. You don't have to wait for a stream service to send information, bring it back into your upstream system, and then batch it into Amperity. All of that now runs side by side, keeping all your data together at the Amperity ID level, so you can trigger different journeys based on real-time context variables. Whether it's web, app, email opens, clicks — all of that becomes a stream service you can use to trigger additional journeys and send real-time information out.

THOMAS KOEP: And the last is still using profiles that are known to you from all your upstream systems and all your historical context, and additionally layering in sub-second information from a web or app. Getting directly to: yes, this person is looking for something special. We know you're shopping for X, Y, and Z because you've already given us some of your information in the past. Now we can reach out to you instantaneously with a very specific offer that is unified at the Amperity ID — so you're not merging a different web ID, an Amperity ID, and a million different IDs at each one of these systems.

STEVEN M. ELINSON: Yeah, awesome. One of the other fun durable truths we talk about relates to market realities. Travel and hospitality customers like to look at adjacent industries for inspiration — retail, consumer packaged goods, sometimes even healthcare, because patient care is very similar to how we want to care for travelers and guests. But what's super unique about travel and hospitality is something we call the look-to-book ratio. Anyone familiar with look-to-book ratios? A few people from travel and hospitality. If you think about yourself dreaming of a spring break or a summer vacation with your family — when did you start planning? The average consumer starts eight months in advance.

STEVEN M. ELINSON: You start to get the inspiration. Maybe you go onto a website and just start browsing: I'm looking for a warm beach vacation. Where are the best beaches? You find Mexico, Portugal, Florida, maybe even California. A couple months later you're back, you've settled on Mexico — but where in Mexico? You start narrowing into cities. That gives you a sense for the look-to-book ratio. Historically, over the last 15 years, that ratio in travel and hospitality has been 10,000 to one. It sounds phenomenal, but when you think about it: when you're looking for an airline trip, it's actually a collection of segments. What you see offered is actually customized and assembled in a bundle only at the time you request it — one stop, direct, two stops. As we've moved into the AI world, with more meta search and online travel agencies like Booking and Expedia, we've seen the look-to-book ratio explode from 10,000 to one to 100,000 to one. Go on Expedia, put in "beach vacation," and see how many options come back — that's how you get to those numbers.

STEVEN M. ELINSON: As we move to an agentic future, many are predicting that number will become 1 million to one. The agents are going to go out and scour the entirety of the world, pull together combinations we can never even imagine, and bring them back to you as the consumer to book. This explosion in the look-to-book ratio and how travel and hospitality has to respond is something we believe every other industry could learn from — you can take those skills and apply them to industries that don't have those kinds of conversion ratios.

STEVEN M. ELINSON: First and foremost — and we've heard about it a few times today — customers have multiple personas. I'll go back to myself and Taco Bell as an example: my Taco Bell order during the Monday-to-Friday lunch hour is vastly different from my Taco Bell order on the weekend when the Knicks are playing. Think about the data points Christian was capturing, and then think about what Taco Bell could do with that first-party data, second-party data, and even third-party data. If they started linking my ordering habits to external events and recognized that I place orders every time the Knicks are playing, think about the personalization and offers they might make to me with those insights. It's not just enough to know who the person is — you have to know the occasion and why they're coming to visit you.

STEVEN M. ELINSON: The second piece: there is absolutely a convenience tax on top of all this. The expectations of all travelers and guests just keep rising. As consumers, we expect every single brand to understand this, and we expect them to do it across all channels. At Amazon we use the term "unified commerce" — most of the rest of the world calls it omnichannel. The reason people call it omnichannel is actually counterintuitive: "omni" is Latin or Greek for "one," and the reason people use the term is because they actually don't have one channel, and they want to try to treat it as if it was one. But that's what customers expect. If I put a promotion in front of you and you call the contact center, that agent better know that promotion was just sent to you.

STEVEN M. ELINSON: Last but not least: travel and hospitality has the most perishable inventory of all. If I were a retailer, I could leave that item on a shelf. But in travel and hospitality, that's not the case. For a cruise line that's unsuccessful in selling a cabin today, it means that for the seven days of that cruise, they'll be unable to sell auxiliary items — no spa, no restaurant, no casino revenue. The amount of revenue actually lost through the perishability of travel and hospitality inventory is remarkable. So we've talked about it a lot today, and I just want to reinforce: that first objective — revenue recovery, revenue generation, revenue recognition — is important. But it also leads us back to the dual path we talked about earlier. You want to give your entire organization access to that governed, single source of truth: high-quality, secure information that lets you continue to accelerate outcomes and make sure you're ready for what's coming.

THOMAS KOEP: Awesome. Wrapping up: the future is unlimited. I went through marketing and digital use cases because those are the easiest to comprehend in terms of stream services and digital personalization. But you can do this for loyalty or operations as well. You could stream geotags and check-in information, combine that with everything you already know about the guest, and trigger a personalized welcome. Or you might know that at the time they check in, they qualify for a higher loyalty tier — and that's now instantaneously updated. By the time they walk from the check-in desk to their room, their updated loyalty status is reflected on the television in their room. You could also add upsell based on loyalty qualification. All of those things become much more operational. The signals are just events and traits — you can send that information at the speed your customers actually expect, hyper-personalizing those downstream systems as fast as possible.

THOMAS KOEP: The last point is setting up success for AI. Historically, AI has not been bringing in things like event streams and web information. This can now, in the future, bring in context about web personalization to train models in real time. You don't have to sit there and process in a four-hour batch or wait for a daily refresh. There are now real-time inputs that AI can go read and update back into your own tenant. And finally: closed-loop data. You can bring in your email opens, clicks, and web behavior instantaneously, without waiting for upstream or downstream systems to process it. It all goes through your Amperity tenant.

THOMAS KOEP: With that — three key takeaways. First, finally being able to unify unknown to known: you can bring that information in, tie it to an Amperity ID, and turn that on at the speed your customers expect — no waiting for upstream micro-batch. Second, you can actually cycle that through a real-time journey at scale. And last: applications outside of just the digital experience. You don't have to do only that — the idea is that we're setting this up so you can run so many different programs through this tool. With that, thank you. There's time for Q&A, but we've got a minute before the next session, so if you want, we're happy to talk about this afterward.