Let me start by sharing my screen. I'm going to walk through a presentation as part of this session. Can everyone see my screen okay? Right. Let's get started.So my name is Aneri Shah. Just an introduction about myself. My name is Aneri Shah. I'm an engineering manager at Amazon where I lead the back end infrastructure and data platforms that power customer engagement and personalization at consumer scale. Thank you so much again for being here today. You know, this is my favorite kind of audience because you are here because you actually care about building things and not just talking about them. Today, I wanna share one insight that fundamentally changed how I think about AI. It is not about models. It is not about the fancy prompts.

It is about something that happens before any of that. And getting it wrong is the hidden reason most AI investments under deliver in production. So the talk is called flat signals to user graph, what the AI capability frontier actually requires. We have got twenty minutes together, so let's make them count. Cool. So I wanna start with something we all know about, but do not say out loud enough. AI in 2026 is under delivering, and we all know about it. We see the demos. You know, they look incredible. We deploy. The real world results are fine, sometimes useful, often disappointing. And the gap between the demo and the production reality is something almost everyone in this room has experienced. Right? So what do we do? We try harder at the AI layer. We find a better shiny model. We spend more time on prompt engineering.

We maybe try a new rack setup, a new fine tuning approach, a new agent framework. And those things help a little, but they heat a ceiling fast. So here is what I have learned from building AI systems at scale for years. The real reason AI under delivers is almost never the model. Sorry to break it to you. It is that the AI cannot see a coherent picture of the person or thing it is supposed to reason about. The data underneath is fragmented, and no models in the world, no matter how capable, can reason well over fragments. So that is the insight I want to unpack with you all today. So what does fragmented data actually looks like? And I want to be really clear here. This is not just a big company problem or an Amazon problem or like a tech giant problem.

This happens everywhere in every organization that has been building software for more than a few years. Think about the users or customers or employee that your AI is supposed to help. What does the data about them actually live? Sorry. Where does the data about them actually live? Their behavioral history is probably in one system. Their preferences and what they have liked or shaped is probably in another system. Their social connections or relationships or colleagues are somewhere else. Their purchase history or, I don't know, actions are in yet another place. So and here is the thing, none of those systems were built to talk to each other. Each one was built by a different team, at a different time, to solve a different problem. Right?

So they are all doing their job, they just have no awareness of each other. So when your AI agent tries to answer a meaningful question about a user, like, what do they actually care about most? Who influences their thinking? How have their needs changed? You know, it cannot. Not because the AI is not smart enough, because the data was never assembled into a picture it can reason over. It is seeing fragments and it is doing the best it can with fragments, but fragments are never enough. So I want you I want to give you a real number now, right, and I want to want you to sit with it for a second. When we audited the data foundation at our platform, a platform serving tens of millions of users, we found that 75 of monthly active users had zero meaningful signal available to AI personalization systems.

That's right. Zero. 75% for for them. Three out of every four users invisible to the AI. Not because they are not engaged with the platform, they had. They had, like, write things, follow people, purchase, listen, interacted for yours, but that data was scattered across disconnected systems and the AI could not piece it together into a coherent picture of that person. So what does that invisibly actually invisible actually mean in practice? It means the AI serves them generic content because it has nothing to personalize with. Every session starts from scratch. No memory, no context, no relationship. The user feels like the AI does not know them because it does not. And and here is the thing that is easy to miss. The more invisible users you have, the less your entire AI investment delivers. You might have extraordinary AI capability on paper. Right? If it only works for 25% of your users, you are leaving three quarters of your potential on the table.

Now I want to ask you something, and this is something I want you to actually do this week. What is that number for your system? It was 75% for our system. What is that number for your system? What percentage of users have enough signal for your AI to meaningfully reason over them? Most organizations have never measured this. I promise you the number will be surprising, and it will reframe how you think about where your AI investment actually needs to go. So the shift from modeling events to modeling relationships. This is the slide I want you to take a photo of. This is the shift. Look at the left side. Modeling events. This is how most data systems are built. You record what happened. User liked something on Tuesday. User swallowed someone on Wednesday. User purchased something on Friday. User listened for forty five minutes. Each action is a separate record in a separate system.

The AI can see a log, a list of things that happened, but it cannot see a person. Now, look at the right side here, modeling relationships. This is fundamentally different. Instead of recording isolated events, you model a user as a node in a graph, as a person in a graph with every signal type connected to that person as an attribute or a relationship. So likes, follows, you know, social connections, history, purchases, context, all connected, all traversable. I I know this sounds like conceptual. You know, the AI does not see a list of action, it sees a person. I want to be clear that this is not conceptual. It is a data architecture decision, and it is a decision that every team building AI has to make, whether they knew they are making it or not.

If your signals are in separate systems with no unified model, you are on the left here. If you have built a unified graph where relationships are traversable, you are on the right. Modeling events gives you a log. Modeling relationships gives you understanding. So that is the shift. Everything else I'm going to tell you flows from this. So I want to spend a moment on the proof because everything I'm telling you is grounded in real production experience, not just a whiteboard exercise. So we consolidated 69,000,000,000 customer interaction records from more than 10 disconnected legacy systems into a single unified graph. The platform now processes one forty k request per second. The percentage of users with meaningful signal available to AI increased by 35 points. Customer experience issues, the end customer experience issues dropped by 30%.

And these are not projected numbers. These are measured outcomes from a production system serving tens of millions of people every day. And I want to be honest about what made this hard because it was genuinely hard. You know, we worked on this for a couple of years, three years. It was a three year long project and it was genuinely hard. The reasons being the platform could not go offline. We had tens of millions of active users throughout the entire migration. There was zero tolerance for downtime. You know, customers are paying for our services, so there is zero tolerance for for any downtime. Three separate systems had to stay perfectly synchronized in real time during this transition to achieve this unification. And any inconsistency was immediately visible to the users.

And the pipeline architecture we needed to make this work did not exist in any existing playbook. We had to build something new. And I'm not sharing this to, like, impress you with scale. I'm sharing it because I want you to know that this principle, unifying signals into a traversable graph, works at consumer scale under real production constraints. It is not easy, but it is worth it. And the principle applies regardless of whether you are working with 69,000,000,000 records or 69,000. So once a graph exists, a new class of AI becomes possible. Now here it here is where where it all gets exciting. Look at this comparison because this is the thing I want you to really feel, not just understand.

The left column shows the questions AI can answer when it is reasoning over events. You know, what has this user done in the past? What content is most popular overall? What did this user click on before? These are backward looking questions, historical. They can tell you what happened. They do not tell you anything about what matters right now or what is meaningful to this specific person in the context of their relationships. Now look at the right column. The question AI can answer when it is reasoning over a graph. What does this user's world look like right now? Not what they did, what their world looks like. What is more popular among the specific people that this user trusts, not what is popular in general. What is popular in their world? What would the people who share this user's taste do next? This is not the same AI. I want to say that clearly.

That that is not a marginal improvement on the same capability. It is a qualitatively different kind of intelligence. And critically, the model did not change. The prompts did not change. The data foundation changed. That is what made this possible. You know, this is why the foundation matters so much. Not because it is cool infrastructure or, you know, something cool to just talk about. Because it is the thing that determines which column of capabilities you are working with. So now let me show you what this actually enables. And, you know, just want to be specific here because this is the part that gets people excited. The first capability is conversational AI that actually understands context. I am talking about a chatbot or a voice assistant that does not start from zero with every single session. It knows who you are. It knows who you trust. It knows what you care about because the graph already has that picture before the conversation begins.

When someone says, show me something my friends loved this week. The AI can answer that, not because it searched a database, because it traversed a relationship graph. Natural language becomes a query over real relationships. The second is AI agents that generate deeply personal experiences. Think about giving an agent a goal. Create something personalized for this user based on what their social graph has been engaging with. The agent can execute that goal because the relationships are traversable. Without the graph, building that feature requires week of data preparation, you know, before you can start experimenting. With the graph, the data is already there. It becomes an afternoon worth of work. And the third, this is the one I think is most underappreciated, is fast experimentation. The biggest hidden cost of fragmented data is not the features you cannot build. It is how slowly you can experiment. Every new AI idea starts with the same question.

How do we get the data? With a unified graph, that question is already answered. Teams move from idea to production experiment in days instead of month. And over time, that compounds. Every new capability you build is faster than the last because the foundation is already there. So, yeah, you know, three major benefits that you're or, like, things that you can build on top of of of the traversable graph that you create. Conversational AI that understands context. This is what we did. AI agents that generate deeply personal experiences and fast experimentation. Again, just want to reiterate, fast experimentation is the key here. And you need to make sure, like, your underlying infrastructure should be able to support fast experimentation that the AI agents can really just pick on. Cool.

So is your foundation ready? Three questions to find out. Before I close, I want to give you something you can use tomorrow morning. These are the three questions I would ask about any data foundation before investing more in, you know, models or, like, prompts or fancy agent frameworks. They are a diagnostic. They will tell you where the real work is. Question one, can your AI see the same user across every surface they interact with? If someone expresses a preference on mobile, is that signal instantly available to your recommendation engine, your chatbot chatbot, your email system, your voice interface? The answer should be yes, instantly, everywhere. If the answer is it depends or mostly or I think so, that is your fragmentation problem right there. Question two, can you traverse relationships between signal types without joining multiple systems?

Meaning, can your AI answer the question, what did the people this user trust engaged with recently in a single query without having to stitch together data from three different systems? If that question requires data engineering work to answer, your AI is working with fragments, not a graph. And question three, the one I keep coming back to. What percentage of your users have enough signal for AI to meaningfully reason over them? Most teams have never run this query. I'm asking you to run it this week. Whatever the numbers whatever the number is, it will tell you more about your AI's true ceiling than any benchmark or evaluation suite you have ever looked at. If you cannot answer yes to the first two and if you do not know the answer to the third, that is where the work is, not the model. Cool. So I want to close, you know, with one thing, just one thing.

The conversation about AI right now are almost entirely focused on the model layer, you know, which model is the best, how to prompt it better, how to build better agents on top of it. And those conversations, you know, matter. I'm not dismissing them. But they are sitting on top of a foundation that most organizations have never seriously examined, and that foundation is the actual ceiling on what your AI can deliver. The model is not the ceiling. The data is not the ceiling. If your AI is reasoning over fragments, isolated events, and disconnected systems with no way to traverse the relationship between them, it will hit that ceiling quickly no matter how good the model is. So if your AI is reasoning over a unified graph where every signal is connected, where relationships are traversable, where the AI can see a whole person and not just a list of things they did, the capabilities that become possible are qualitatively different.

And, you know, this is something that we have experienced post, Nami. And everything you invest in in AI on top of that foundation compounds over time. So build a foundation that lets AI see your whole user. Raise the data ceiling first, then everything else follows. Thank you so much. And, you know, I would love to connect with all all of you. Find me on LinkedIn. It is the best place to continue the conversation. Thank you so much. Cool. I think we are at time, so I'm going to cool. I'm I see a lot of thank yous coming in. You know, thank you. You were an amazing audience. Thank you for taking the time to to be present for my session. I think we are at time, so I'm going to log off and end this session with that.

Again, you know, please find me on LinkedIn and happy to connect over there. Thank you.