Building Enterprise-Ready Agentic AI with Search - Six Five On The Road

How are organizations harnessing agentic AI and search to build more autonomous, reliable enterprise workflows?

From AWS re:Invent 2025, host Nick Patience sits down with Elastic’s Steve Kearns, GM of Search Solutions, to unpack how search is becoming a foundational layer for enterprise-ready agentic AI. They discuss how organizations are moving beyond traditional RAG into fully agentic systems—and what it takes to make those systems trustworthy, scalable, and operationally mature.

Key Takeaways Include:

‍

🔹Evolution from RAG to Agentic Systems: Elastic supports a seamless transition from retrieval-augmented generation (RAG) to agentic workflows by providing flexible data management and lifecycle tools for agents.

‍

🔹Agentic AI Challenges and Solutions: How Elastic addresses common enterprise pain points in adopting agentic AI, such as configuration, execution, customization, and observability, with built-in primitives and operational capabilities.

‍

🔹Real-world Applications and the Future of Agentic: Examples of how Elasticsearch powers intelligent agents today, and insight into Elastic’s roadmap for advancing agentic architectures in enterprise environments.

‍

Learn more at Elastic.

‍

Watch the full video at sixfivemedia.com, and be sure to subscribe to our YouTube channel, so you never miss an episode.

‍

Or listen to the audio here:

‍

Disclaimer: Six Five On The Road is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

‍

Nick Patience:

Hi, I'm Nick Patience, the AI Platforms Practice Leader at Futurum, and we're here at AWS reInvent in Las Vegas for Six Five On The Road. And joining me today is Steve Kearns, the GM of Search Solutions at Elastic. Welcome, Steve. Happy to be here. Thanks for having me. So let's talk about the topic of 2025 and its 2026 agents and Agentic AI. I mean, where are you seeing customers at in terms of understanding what it means and how do you define it, maybe? Yeah, it's a great question.

‍

Steve Kearns:

I think a lot of people are starting in very different places when they think about agentic AI. I like to start very simple in terms of trying to define it and then sort of attach more concepts as we go. And really at a very basic level, you can almost think about agentic AI as LLMs using tools in a loop. The Simon Willison definition is really a very simple one. And then you start to unpack it a little bit. And well, an LLM is a really smart bastion of general world knowledge trained on all the data in the internet. But on its own, it can only answer questions that are in the public data. And so as enterprises begin to deploy these kinds of LLMs to answer questions and take action within an organization, they need to provide, with tools, the right information to actually answer questions and make decisions using real business data from that organization. And I think that's really the key for me when I think about what makes this agentic AI work, what's the difference between successful and unsuccessful agentic AI. It's the teams, it's the people, it's the organizations that recognize very early that step one in building a successful agent is not having a great LLM, it's having great context that you bring to that LLM to allow it to answer the questions. And how you do that is really the magic of sort of what makes an agent successful and sort of what makes it complete. So let's dig into that, how you do that.

‍

Nick Patience:

So how do you add context to that?

‍

Steve Kearns:

Yeah, so we talk a lot, and you might hear a lot this year, the year is maybe agentic AI, but context engineering might become that term that grows in popularity throughout the year. When I think about context engineering, I think of there's almost like two wings to sort of providing context and managing context within an agent. The first, and in my opinion, the most important and the sort of most foundational is getting the right information into the LLM. And again, the context engineering, I think of this as a retrieval problem. If you think about it, I have all this data across my organizations and all of these different systems. How do I get that data together? How do I get that data accessible to an LLM? And then when it wants the data, when you need data to answer a question, an interaction, make a decision, how does the LLM get that data? And this part of context engineering of the retrieval part, bringing that data in, that I think is one of the most challenging parts. Because if you don't have the right relevant context in the context window, All of the other things that are very important that you need to do for context engineering don't matter, because you didn't get the answer right at the beginning. Once you do get the right context in, context engineering is now a very large and sort of growing part of the discipline. There's like a short-term memory within the conversation. There's long-term memory. What do I learn and retain about you that I can use as additional context to help guide the questions, help guide the research? But there's all of these other pieces. But for me, none of that matters if you can't get the right data to the agent to help answer the question in the first place.

‍

Nick Patience:

So many companies, I mean, you said retrieval. I mean, RAG is the term that many companies will obviously understand what that means. But how does Elastic help with that specifically?

‍

Steve Kearns:

Yeah, so I actually loved RAG as a term. Unfortunately, its definition has become very narrow in terms of what it's actually doing. But what I loved about it was what it was doing was right in the name, retrieval, augmented generation. And Elastic, you can think of us as We are the company behind Elasticsearch. Many folks will be familiar with that. It's one of the most popular open source products in, I don't know, history, probably top 10. But what we do is we are almost like the data platform for context engineering. When you bring data into Elastic, you can then query that data in very rich ways and very simple ways to get the right answer, regardless of what kind of question you have. If you can bring the data into Elastic, we have a very rich and powerful query language. And so sometimes, when you're dealing especially with unstructured data, finding the one document that contains the answer, right? The easy use cases like, where's the holiday policy? Well, do I need the US or the UK holiday policy? Do I need the one from this year or last year? Those are easy versions of this question, but if you imagine the kind of unstructured data across an organization, it gets much harder very quickly. And so at Elastic, we have a number of techniques that we can bring to bear to simplify that process. In some ways, it's simplifying a very complex thing, but you need to have the power of sort of the full power of information retrieval at your fingertips when you're doing this kind of retrieval. And for us, we think about this in a couple of different ways. If you were to think about search techniques and so forth, there's traditional keyword search or BM25, this traditional way to approach search that's very powerful, very effective. And then there's a more modern approach that can give you better results called vector search or semantic search, which actually searches not just on the words themselves, but on the meaning of those words in the context of how they're used in the phrases and of the sentences. And at Elastic, we support both lexical and semantic search, and we have built-in first-party embedding and re-ranking models that make this very easy to use as a part of that platform. And so when you're thinking about how do I get the right answer, in many cases when you're dealing with unstructured data, it starts by saying, what do I need? What kind of querying capabilities do I need to actually get the best answers out? And for that, I need hybrid search. I need Lexical, VM25, and I need vector search, and I need to be able to combine those. And I also need the best retrieval models, the best embedding models, and so forth. And these are areas that we've done a ton of investment at Elastic. In fact, we just acquired Gina AI, one of the leading embedding and re-ranking model companies, to give us additional capabilities in this kind of search and relevance part of the domain.

‍

Nick Patience:

So do you help organizations use models without having to necessarily understand which model is the right model? I mean, because obviously there's a new model out every week, every day, more or less. I mean, is that part of the offering?

‍

Steve Kearns:

Yeah. I mean, one of the things that I've been very clear about as we've built out the product, especially over the last 18 months, we've been very clear that we want our products to feel very easy to use out of the box. So this like, but then when you need to customize, when you need to go further, you have the full ability to open the box, go down to the nth degree, and to configure every single piece. But if you walk up to Elasticsearch today in Elastic Cloud, we have a fully managed offering. You don't have to think about running software, things like that. You just push the data. And we will automatically generate embeddings when you're bringing the data in. At query time, you also need to generate the embeddings, which we can do that for you as well. And we'll do that with a first-party model running on a GPU-based inference service. And all of that just happens without you having to know anything at all about embeddings or quantization or any of these other techniques. We can handle that for you in this fully automated way. But when you reach that point where you want to choose your own embedding model, you want to customize the selection to say, hey, in this case, I know I'm going to be dealing with multilingual data. So let me make sure I have an embedding model that works really well for the languages that are important to me. Or I'm going to be dealing with images, and let me make sure I have a multimodal model. In those cases, you can open the box, and you can bring any embedding or re-ranking model that you choose. Same thing for LLMs and your building agents. And this idea of starting with a system that works out of the box and then incrementally allows you to customize is really important. What we've found is our customers have put these agents into production. It starts by saying, what's the proof of concept? How quick can I get a sense of, is this going to be a good idea? Can this work? But when you put this in front of users, the long tail of questions and the long tail of what they actually want to accomplish is really long and you need that power. But you don't need that power at every level, at every piece all the time. And so we try to allow people to be successful and then only have to learn more when I need to open the box to customize this piece. And that idea of progressive disclosure of complexity really simplifies the process of people building these things and getting started.

‍

Nick Patience:

So where are, I mean, we're very, it attracts me as an industry analyst. We're very early in this agentic journey. I mean, you talked about some of the things that organizations want to do. Can we talk about what they actually want to do? What are the use cases you're seeing at this early stage?

‍

Steve Kearns:

Yeah, I mean, it's sort of fun. We're seeing so many different use cases. I think it reminds me actually of the early days of search where people started to bring unstructured data into search engines and suddenly the whole world opens up. You mean I can bring in my own personal data from Strava and I can put a dashboard to show my workouts and where I've been running and riding? That sort of thing is something that's always been fun to do on top of this data. But in terms of business systems, we're actually seeing adoption starting in a few different places. We see a lot of organizations building this for internal teams first. And a lot of the places where we're able to find and sort of demonstrate the easiest success are these functions in these areas that are sort of measured and repeatable and happen a lot. And so we see a lot around customer support. This is a common end of the use case. And it's not always just like customers, customer support. We see this internally. We had a great example at Adobe where their internal engineering productivity team. You can file tickets, they've got thousands of engineers inside their organization, and if they can make those engineers more productive by answering their ticket faster with an agent, that saves them money, that allows them to move faster as an organization, and it's measurable. And that's the beautiful thing about some of these use cases where everybody wins in these kinds of setups, because in the Adobe example, engineers often get the answer to their question right away. They saw between a 30 and a 40% case deflection. People who would have opened a case didn't. So that means all of those people didn't have to wait for a human to read a ticket, answer a ticket, and get back to them. They got their answer and can move on, stay in the flow and stay moving. And that team, the team that's answering the tickets, gets to focus on the more interesting and the more challenging problems. And so I think in those kinds of examples, we see a lot of sorts of examples across all kinds of different businesses in that area. But we're also seeing that people build these right into their core products and their core experiences. A couple of great examples, we have a number of financial services companies. And if I simplify a much more complex use case for a moment, there's one that's a financial advisor. sort of a use case. And the advisors are advising a number of clients who are calling in. They have questions about their portfolios. They don't understand the prospectus. They don't know quite where all of their money is invested. And so how quickly can we help that financial analyst understand the customer, answer the questions for them, and give them the right kind of guidance? These use cases, it's pulling back the structured information about the investments, looking up for each of those investments, the prospectus, if it's a fund. And within that, now pass that as context to the language model to get the answer back for the analyst to then use for the person. And these ideas of making employees more efficient, that's where it starts. And then increasingly we see people turn that around to say, now how do I show this directly to my external customers? How do I get these same capabilities, the same efficiency? But the use cases are extremely wide. We use it all over Elastic as well, and a number of these kinds of use cases as well.

‍

Nick Patience:

I was going to say, it's effective for any kind of company of any kind of size. I mean, everybody's got unstructured data, don't they? Whether they know they have or not. I mean, so is it really universal in terms of its applicability?

‍

Steve Kearns:

I think so. I mean, another way to think about this is if you look at almost like any Imagine any business and think about the number of different human actuated workflows that happen all of the time where they're touching different data systems or retrieving data, they're pulling that back, they're making a decision, then they're taking an action. Almost anything like that, you can start to imagine how an agent can actuate those things for you and start to automate pieces of that and make it much more efficient to sort of operate those kinds of workflows. And so I do think that it's almost like we talk about them now as agents in sort of use cases, we might start to use language more like workflows over time. How do we automate more of these workflows? Or how do we accelerate more of these workloads? And I'm not sure quite where the language will play out, but it feels like every workflow within a company that humans are doing today, where they're pawing around reading documents and doing that, that should be faster, that should be quicker, that should be easier.

‍

Nick Patience:

Yeah, that should be automated, shouldn't it? So what's, I mean, so lastly, really, so just, you mentioned, you know, which hinted at looking to the future, but what do you see as the future for Elastic in the kind of agentic AI space?

‍

Steve Kearns:

Yeah, it's a really exciting time. So we're actually right in the early stages of launching our own additional layers to our platform. I sort of mentioned before, we're trying to, how simple can we make it to bring data in, to give you great relevance, sort of on the output side of that. We just added a new layer to our products called Agent Builder. And the idea is we looked at all of the things you have to do to build these agents. I need to understand short and long-term memory. I need to understand the context. I have to get history. I have to have user-level personalization for each of the history and all of the actions that they've taken and different information about the business and so forth. And we said, how much easier can we make this process? And we said, well, what if instead of bringing data to the agent, what if we brought the agent into the data platform? And so now we've got right inside the box an agent that works. And so if you start to just bring data into Elastic, the time to start asking questions is zero seconds. you just click the agents tab and you start typing questions. And now we've built right inside the product, our own, you know, almost like a retrieval sub-agent that goes and looks at the mappings, looks at the schema, no preparation needed. You can prepare it, you can customize it, but if you don't, we'll look at the mappings, we'll look at the data within the mappings, and it will start to understand the questions that you're asking. Do I have the data necessary to answer this question? If I do, what kind of query? Is it a retrieval query where I'm trying to get the one best document? Or is it an aggregation request where I'm trying to look at information over time? How many tickets did this particular customer open? Categorize these things for me. Look at this metric over time. Look at all of the alerts that happened in the last 24 hours and tell me are any of them related to one another? These kinds of questions you can answer and retrieve context from Elastic in a fully automated way right out of the box. And then we add the ability to customize on top of that. And so we give you that out-of-the-box agent. Neat. It's very powerful. It's very useful, but it's very generic. If you're going to put this in front of that financial analyst in that example, I'd want to give that more specific instructions. I'd like to tell you very specifically, hey, here's what the data actually means. And I want to customize the retrieval pieces to say, hey, when you're retrieving somebody's portfolio, use this query because I know it's going to be correct 100% of the time. And I know you're always going to want this extra information, so let me encode that in custom tools that I can define just as part of this platform. And the neat part, at Elastic we've always sort of been very comfortable being the data layer. We're more and more comfortable over time being the UI layer, but we don't have to be the UI layer. And so as we build out this agent builder set of capabilities, we think this is going to power the next generation of experiences and applications. And the neat part for us is we're going to have a very compelling user experience and so forth, but everything that we do is API based. And so you can actually say, I'd like to connect my existing agent, whether that's an agent core agent that's calling out over MCP to query data from Elastic to use as context. Great. We're happy to sort of operate in that way. Or over A to A, you can directly connect your UI libraries or your UI agent kit or something like that directly to Elasticsearch as the back end in much simpler ways than was ever possible. And so I think this is going to be the year of people really starting to get thoughtful and creative. I think this is going to be a fun year for technology. We start to see how many different ways and places we can apply meaningful, functional, capable agents that have good relevance on top of a wide variety of different types of business data.

‍

Nick Patience:

That's great, yeah. Should be powerful stuff. Indeed, a lot of fun. It will be. Thanks, Steve, very much. That's all we've got time for today. So thanks very much for joining us at Six Five On The Road here in Las Vegas at AWS reInvent, and we'll see you again next time.

‍