🦄 Making great presentations more accessible. This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.

Overview

📖 AWS re:Invent 2025 - Agentic Data Fabric: Powering Autonomous AI with Trusted Data (AIM257)

In this video, Reltio’s Chief Product Officer explains how their data unification platform creates trusted 360-degree views of customers, products, and suppliers by consolidating multi-source data into relationship graphs. Two demos showcase AI agents: one automates data steward tasks by identifying and resolving duplicates in minutes instead of 30-40 minutes, and another delivers personalized product recommendations by analyzing 70+ data points from 11 sources across customer households. Examples from CarMax and Warner Brothers illustrate real-world applications. The solution is built on AWS using Agent Core and Strands SDK, enabling scalable agentic AI that operates on pre-unified data rather than assembling information in real-time.

; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.

Main Part

Reltio’s Data Unification Platform: From Master Data Management to Trusted Profiles

Thank you, thank you. I noticed that I am now a Chief Product Officer. The R is missing. Good to have everybody here. We’re going to cover a lot of ground in the next 20 minutes. I’m going to talk about Reltio, the company. We’re going to talk about our product. We’re going to talk about all the amazing movement that we’re seeing by building agents on top of data that is within Reltio. So what is Reltio? Let’s start with that.

Reltio is a data unification platform. The idea behind data unification is that for any company of reasonable size, they will have plenty of data about their customers, their suppliers, different products that they sell, and this data ends up being in different parts of their organization. Whether it’s in different SaaS products or different databases, you have product data in a warehousing system, you have product data in an e-commerce system, and so on. So how do you bring all of this together, unify this data to produce a single authoritative view of any of these nouns that matter for your system?

What Reltio does is bring together this multi-source data and give you a trusted single view on which you can build your business operations at the highest level. Our approach has been to start with a space that is called master data management. If you know what master data management is, raise your hands. The idea of master data management was to bring together multiple different records and really manufacture a golden record out of that. This definition of golden record was fairly narrow. You were talking about the definition of an individual, where they live, their contact details, and so on.

As the space has evolved, this idea has become less about a golden record and much more about a trusted profile. These are very rich descriptions of customers or suppliers and so on, which have detailed amounts of information drawn from many different sources. This graph representation of these entities is maintained up to the second and it’s available for consumption via an API or through real-time systems all the time. It’s an evergreen data foundation that a lot of other processes can be built on.

The way we approached it at Reltio is to think of all of this as a relationship graph with different entity types, their relationships, and including a 360 view of all the behaviors that might be associated with these entities. In simple terms, if you’re talking about a customer, a customer has an address, they have a purchase history with you, they have a click stream from browsing your site, and so on. This complete 360 view in a way that is connected, where every entity is connected to everything else, really is the heart of what Reltio presents.

The story you can already see coming is that a graph is a very natural data structure for large language models to interact with. On the left-hand side, you see many different data sources that can be used to bring data into Reltio. But the interesting bit is on the right-hand side where you see the ability to consume this graph either through applications or by sending it into different data warehouses. Most interestingly, our customers are building agents on top of this data, and that’s really been an amazing experience to watch. I’ll give you two examples.

CarMax, the biggest retailer of used cars in the United States, uses Reltio to organize the data into three 360s. Why? Well, at the highest level, the promise they have to their customers is that somebody can walk into a CarMax, either buy or sell or buy and sell, and leave within 15 minutes. So they have a transaction business guarantee of 15 minutes. How do they do that? Well, to achieve that, they have to organize around the fact that they see a typical vehicle in the United States three times in the vehicle’s lifetime. It’s traded multiple times. So it makes sense for them to have a 360 view of every vehicle they possibly can so that they know as much about that vehicle as possible over time.

They also have a 360 view of prospects or customers that they’ve dealt with in the past, and they have a 360 degree view of their employees. For them, a transaction is the combination of a vehicle, a prospect or customer, and a salesperson who’s made the sale. Because their goal is not just that the customer can walk out within 15 minutes, but also that they settle with the salesperson within that 15 minute window.

The way to achieve that business outcome is to organize the data in a way that is friendly for that outcome. That is what they use Reltio for. Now, let me share a very different example from a completely different industry. Warner Brothers’ product is their intellectual property—all of their characters. If you think about Looney Tunes and pick a character from that universe, that particular character has appeared in a certain number of movies or cartoons. They have appeared as merchandise through different suppliers, and they have licensing agreements. That is what constitutes a 360 for Daffy Duck or Roadrunner. Having that data ready to be explored via large language models is what Warner Brothers uses us for.

So these are two very different examples, but the point is that the way the data is organized reduces the effort in ingesting and organizing the data. It makes the consumption of the data much more front and center. When we think about this space of bringing together data from multiple different sources and assembling it together, this problem has existed ever since we have had multiple SaaS systems or multiple databases. This data fragmentation problem is at the core of what we are trying to solve. We are saying that there might be two, three, or four records that all indicate or are all related to the same individual. How can we pick and choose and connect these records so that we have as much of a trusted view of this customer as possible?

You are using first-party data, and you are also bringing in third-party data to enrich the output of this process to be as high quality as possible. However, it is not all automated yet. Even in 2023 and 2024, there is still a lot of manual effort being put in. If you think about larger companies, there is even more manual effort. Data stewards, which is a particular role, are spending a lot of effort making sure that the data is as high quality as possible. Of course, data quality itself is a very interesting topic of discussion because you can spend a lot of money getting data quality to the highest level possible, and the ROI on that is really determined by what the application of that particular data is.

Data quality has to be mapped to the business outcome that quality is looking for. So how much effort you put into data quality is the second piece that has sometimes been highly manual and sometimes very expensive. The third thing is complex data models. As I described, this space is about bringing together different data models or data from different data models and combining them together so that you have an authoritative view of that particular entity. There is complexity in being able to map all of these together, and typically that has been done from a manual point of view.

These are the real problems. It is very easy to say that if you have your data organized well, agents are going to behave well, or that trusted data leads to trusted AI. But the how part is left unsaid a lot of times. The approach that I am going to share with you really goes to the heart of that how. It is a very pragmatic approach to how to get trusted agents. The goal that we set out on is really to take this manual effort in organizing this data down by an order of magnitude or two orders of magnitude and automate as much of that as possible.

Building Agents on Unified Data: Demonstrations of Data Management and Personalized Recommendations

So here is what I am going to walk you through. I am going to walk you through a couple of demos pretty quickly here. But just to orient you on our product set, everything runs on a unified Reltio Data Cloud. Any of the products that run on top of the data cloud inherit the scalability and the security aspects that are built on top of AWS. The two products that we offered a year ago are multi-domain MDM and our Intelligent 360 products. The way to think about it is that multi-domain MDM encapsulates everything that has to do with entity information. Then, if you add behaviors and wrap that around to build a true 360 in real time, that is the Intelligent 360 product.

What we are going to talk about today is really the agents and the agent flow, which is the layer on top of these two capabilities that delivers fully functional agents in two regards. The first agent class is for managing data itself. So the data that is in the graph that exists within MDM or within Intelligent 360—how can we manage that autonomously? I will show you an example of that.

Once the data is managed in as pristine a state as you can make it, let’s put it to use. How do you build an application effectively or an agent that’s partially doing what a CRM would do or what a customer success system would do, but do that as an application that’s built on top of data? This is the data-first application movement that’s very much in play right now.

Let’s look at the first scenario. In this case, we are trying to solve for data management and duplicate data. As you see, it’s a conversational interface with really no mention of data anywhere in the interface itself. We’re just asking a business question. We’re saying, "We’re seeing service delays in our top 100 segment, and we suspect that there is some data duplication going on."

At that point, our agent is looking at what that top 100 means and it’s getting that definition from the underlying data platform. It’s able to then propose an approach where it says, "I found what that top 100 means for each organization. I’m going to systematically look at why that duplication is happening." It gives you simple options you can pick, and as I’m sure everybody’s very familiar with these sorts of interfaces at this point, it’s identified that there are two organizations that actually have some duplication in the data.

Further, it proposes a couple of actions for each one of those. For this particular organization, Solar Turbines, there are two versions of information about this organization. Looking at it attribute by attribute, I can see that there are some mismatches that are probably causing suboptimal shipping, like you’re shipping to the wrong location and so on. By the way, it’s also enriched this data by looking it up on the Internet. That’s one of the requests we’ve always had from customers. There’s such a rich and ever-updated source of data on the Internet. Why can’t that be brought into the data management platform?

So it’s done all of that. It’s brought in this data and it’s created this level of automation where if somebody had to go and look this up, they would very easily take 35 to 40 minutes to resolve this data. A data steward would very easily spend 30 minutes resolving this, and this took less than one minute. Just the savings in terms of doing 100 of these a day is pretty enormous. Of course, needless to say, the whole thing can be automated. It doesn’t have to be conversational. Everything is API driven, so you can very easily wrap this up in an even next level of automation.

The next demo I’m going to show you is around that next level of building, being able to build applications given that the data is already organized in a proper manner. In this demo, what we’re showing is there’s customer data and there’s product data and a call is coming in from a customer. Our agent intends to show the next best product recommendation or at least propose that to the agent so that they can talk to the customer. Of course, there’s no assumption about whether it’s a human agent on this side of the phone or an AI. It doesn’t really matter. The point is, can we produce a recommendation down to the personalization of one? Given as much data as we have about the customer and their past behavior, can we actually build something that is very, very customized for them?

Very similar to the previous demo, we’ll drop into the conversational interface. In this case, the question we’re asking is for a certain customer, we provided a name, Sarah Anne, and please provide three compelling product recommendations based on her profile and the different relationships that we see in the system and different interactions that they’ve had with us. The system’s looking for and searching for that individual and it finds a perfect match. It’s going through our MCP layer and talking to our API. It says, "We found a unique individual. Here’s some information about them. They’re a 46-year-old married female from San Francisco. We’re seeing from them that they have purchased from us recently. There’s a pattern there. They’re actually an enthusiast. They interact with our blog or some of our content."

Here is a set of recommendations that she would most likely be interested in. The thing on the left is not interesting to me. The thing on the right is more interesting because that is the why.

Why is the agent recommending certain things? This is more interesting because that is the why—because at the end of the day, we have to be able to defend our decisions and present something to a customer, right? So we go a step further and say this is great, but if you look at the household, perhaps she’s interested in buying something for her household, somebody else, a gift for somebody else, not just something for herself. So what would you recommend in that case? We look at the next surrounding circle or level in the tree, and we find that the family consists of Michael, Olympia, and Sarah. Olympia, who’s the daughter, has expressed consent for new product launches, so we could potentially position something for her in this case, right? The system goes through and says for Olympia and Michael in this case there is a specific kit that we’re recommending, and then the rationale on the right-hand side is why we’re making that recommendation. It also scrolls here in a second, and what’s interesting to me is also the suggestion to the agent who’s having the conversation. There’s a communication strategy, right? It’s about how to present, not just what to present, but how to present this, and that’s based on a certain amount of data that we’ve already seen in the system behavior in the past. We’ll actually ask the agent, "How much did you consider? How many data points did you consider while making these recommendations?" It can come up with an answer, and in this case, I think it’s used 70+ data points from multiple different sources—11 different sources—and multiple interactions that it’s had access to in the past. What this shows you is that if you have data organized a certain way, then using agentic AI on top of that almost seems simple and magical. Conversely, think about what you would have to do if you’re trying to create an agent that is delivering you this outcome, but your data is distributed in five different databases and multiple applications, right? You’d be stitching together all this information when you need it, and that becomes a very expensive operation. The outcomes become harder to predict because on the other end there’s a customer you’re trying to solve that problem for or present a proposal to, and your system is busy assembling all this information together.

Powered by AWS Technology: Leveraging Agent Core and Strands SDK for Scalable Solutions

That to me is the real power of a data unification system where you have a graph ready to be consumed at all times and you’re not really spending any engineering resources. Once you’ve set up the graph, it’s ready to go at all times. So a little bit about how we built this—of course we are powered by AWS technology. We were one of the earliest users of Agent Core and the Strands SDK. The core of the agent is built out using Strands, and for us, Agent Core as it has developed has become really this sort of super easy way to scale up and not have to worry about a lot of the non-functional aspects. We don’t have to worry about whether security is taken care of or access to memory, so we don’t have to assemble memory and web access tooling. All of that comes as a package, and we can focus on really the value that we’re creating on top of that.

In a sense, the model used doesn’t really matter because all these models are so capable. I think as an ISV for us to create this, it really is about getting those non-functional aspects to be testable because today we’re using Model X, tomorrow it’ll be Model X plus one, and we need to make sure that we’re able to actually give you the same performance at scale. Remember, our customers are some of the largest companies in the world, so the ability to operate at scale and yet go from one version of a model to the next version of the model and guarantee no regressions—that’s the sort of important thing that a solution like Agent Core provides us, right? So we’re very excited about some of the policy-related announcements that were made yesterday, and I’m really looking forward to incorporating some of that into the mix.

With that, I’ve shown you a bunch of videos. If you want to see the real thing in action, we’re at booth 1227. I’ll be there for a little bit and happy to talk, but the Reltio team is there, and we’d love to take questions there. Really appreciate everybody’s time. Look forward to seeing you.

---

; This article is entirely auto-generated using Amazon Bedrock.