The Architecture Shift Behind Reliable Enterprise AI - with Ravi Marwaha of Arango

Summary

---

The Architecture Shift Behind Reliable Enterprise AI

Podcast: The AI in Business Podcast
Host: Daniel Faggella
Guest: Ravi Marwaha (COO & Chief Technology Product Officer, Arango)
Aired: May 14, 2026

---

Episode Overview

This episode explores why AI pilots often succeed in controlled environments but falter when scaled across the entire enterprise. Ravi Marwaha argues that the breakdown is not due to AI models themselves, but to fragmented and inconsistent context within enterprise data. He shares key strategies for business leaders: focus on relevant, timely, and accurate information, rethink legacy data architectures, and prepare for the operational shift as AI agents begin creating and acting on their own records.

---

Key Discussion Points & Insights

1. AI Pilot Success vs. Enterprise Scale

• Pilots succeed in labs: Controlled, curated data enables fast AI success.
• Enterprise challenges: Real-world production environments present fragmented and inconsistent data across many systems.
  • Quote: “Real world is messier than that, right? Especially for large enterprise, your data is never in one place...your data is not curated...it’s not static.” — Ravi [07:11]

2. The True Bottleneck: Not Models, But Context

• AI model maturity: Models are no longer the main issue—they are “very, very good.”
• The ‘context’ problem: The challenge is getting the right information to the AI at the right time.
  • Quote: “Model is not the problem. Context is...Do you have the most relevant, most timely, most accurate information at the time of decision?” — Ravi [08:10]
• Fragmentation pain: Companies wrongly focus on bigger data pools, instead of clarifying what data matters for each use case.
  • Quote: “It’s not about, do I have all the information in one place? That’s not the point.” — Ravi [08:27]

3. Rethink Data Practices for Agentic AI

• Data architecture legacy: Data warehouses, lakes, and BI tools were built for human access and reporting—not for autonomous, repeatable AI decisions.
• Agentic operations require new context layers: AI agents need decision context, not just retrievable data. These agents also generate their own records that must be governed and auditable.
  • Quote: “Your data needs, which were previously designed for human consumption across multiple screens...is that still the case? It’s not.” — Ravi [32:23]

The Paradigm Shift

• From centralized data to context-driven access: Not every decision needs every piece of data.
• Governance and auditability: As agents make decisions, companies must track why decisions were made—storing agent-generated actions, context, and policies for review.
  • Quote: “Now those agents, like us humans, create information...Those decisions now need to be also stored because now they’re correct and governed.” — Ravi [12:56]

4. Governance and Organizational Impact

• Emergence of an 'Agent Resource Management' (ARM) function:
  • Agents are like employees—requiring policies, supervision, HR-like oversight, and context-specific tools.
  • Memorable moment: Host jokes, “Do we need an HR department?” Ravi responds, perhaps we need an “agent resource management” or “ARM” department. [21:51]
• Who owns the process?
  • Risk appetite and use case dictate the level of supervision—some industries will require tighter controls, human-in-the-loop oversight, and explicit policy frameworks.

5. Practical Steps for Business Leaders

• Start with business value, not technology: Don’t get distracted by AI hype or rush to master every new tool; stay laser-focused on solving real business problems with clear ROI.
• Define information needs for each use case: Clarify which context (timeliness, relationship, source, auditability) is required for each decision—don’t hoard historic data unnecessarily.
  • Quote: “It definitely depends upon your use case...if your use case is, we launched a new product...you just need the data for the product that you launch.” — Ravi [10:18]
• Map out pilot conditions and generalize them: When pilots work, study exactly how data was curated and accessed; replicate that context at scale, not just the data itself.

---

Notable Quotes & Memorable Moments

• On AI Pilot Hype:
  “A lot of companies... just want to jump in. But as most things in life, it’s better to take one step back, look at the company... You’re just integrating the AI where it can accelerate what you’re already good at.” — Daniel [05:35]
• On context, not quantity:
  “It’s not about, do I have all the information... it’s about, do I have the most relevant, most timely, most accurate information for the decision that I need to make?” — Ravi [08:27]
• On AI as a new architecture challenge:
  “You have to shift your gear from an application world to tools, context, information and decisions and specificity of what the action and the policies are that need to be to govern their actions.” — Ravi [17:26]
• On building an ARM department:
  “Maybe it’s an AR department, so not HR, but agent resource management or ARM something. But yes, that is how we will have to look at them, because they are doing the task—some of the tasks—that we as humans were doing before.” — Ravi [21:51]
• On the optimism of ‘clean slate’ intelligence:
  “You’ve got that blank slate to really build clean context...Humans enter a lot of systems with a lot of bias already built in through context...with mapping it out correctly, you can build clean context.” — Daniel [34:30]

---

Timestamps for Important Segments

• [02:37] Pilots vs. Enterprise: Why scaling AI fails
• [07:11] Fragmented context: The real challenge
• [09:18] Models hallucinate due to lack of context
• [10:18] Focus on use-case-specific, timely data
• [12:10] Data fragmentation and legacy architectures
• [16:33] The paradigm shift: Architecture for AI agents
• [19:59] Governance, risk, and oversight in agentic AI
• [21:51] “Agent Resource Management”: The enterprise ARM
• [24:40] Breaking down context requirements
• [32:08] Practical advice for leaders—start architecturally, even at small scale
• [34:30] Clean context as a building block and competitive advantage

---

Actionable Takeaways for Enterprise Leaders

• Don’t chase AI trends blindly—define your business impact first.
• Assess and prioritize relevant, timely, and accurate information for each use case, rather than hoarding all available data.
• Consider the context and tools your AI agents need, including how their actions and decisions will be audited and governed.
• Challenge legacy data architectures—prepare for a world where agents become operational colleagues with their own requirements for memory, oversight, and process integration.
• Approach architectural redesign incrementally: map what worked in pilots, generalize context requirements, and apply architectural thinking before technical implementation.

---

Transcript

[A]

Welcome everyone to the Emerge AI in Business podcast. Today's guest is Ravi Malwaha, Chief Operating Officer and Chief Technology Product Officer at Arango. Arango is a multimodal data platform that integrates graph, document, key value vector and search data in one system, enabling enterprises to supply AI applications with consistent governed context that supports reliable decisions and auditability. Ravi joins us on today's episode to explain why AI pilots often succeed in controlled settings but falter when deployed across the enterprise. He underscores that the core issue is not model performance, but fragmented, inconsistent context that prevents agents from making reliable decisions at scale. He details how enterprises must reassess what information each use case truly requires, how that information is accessed and how agent generated decisions will stored and governed. This shift pushes leaders to rethink long standing data practices and prepare for new operational demands as AI systems begin performing repeatable tasks and creating their own records. Today's episode is sponsored by Arango for our Solutions partners. According to Edison Research, 79% of Americans age 12 listen to online audio monthly an estimated 228 million people. For busy executive audiences, podcasts offer a rare opportunity to capture 20 plus minutes of attention with VP leaders in America's largest enterprises. Emerge reaches 1 million listeners every year. Learn how leading AI brands convert executive attention into measurable pipeline impact. Download our media [e-mail] partner that's g o.emerj.com partner. Now the conversation with Ravi. Ravi, welcome to the program.

[B]

Thank you Maruli. Glad to be here.

[A]

So I think a lot of the leaders listening to us today has seen AI reach a certain maturity in the past six months, but they've also been hit with a real disillusionment. They see AI pilots working well and then when they try to scale that same project across the entire enterprise, they hit by certain inconsistencies or even a complete brick wall. So. So from your vantage point, what seems to be that underlying issue that's slowing them down?

[B]

Great question. So maybe I'll reframe the problem statement a little bit. So one of the notion people have that AI projects are failing or the enterprises are struggling with the adoption of AI, all of that is true, but the promise of AI is real. On the other hand, there are real challenges. First and foremost, everybody should understand, take a breath, catch a breath. This is almost no different than when the client server technology evolved or when the Internet revolution happened. So this is just new revolution that's happening and it is going to change the behaviors, it's going to change the way we work, it's going to redefine the economy. We hear all of those stories. I'm more optimistic than pessimistic that it will create a new type of economy for us. So the one is that, yes, everybody's experimenting and trying and they want to learn. Everybody's learning. Yes. Enterprise don't have as much focus on the AI because that is a technology for them to use, not so much to develop. On the other hand, we as developer or developers of this technology or people who are building applications or solutions around that technology, have a little bit more insight into the technology. So the technology is real challenges in how people are using it. The other thing is, as it's an experiment or a startup point, most of the people are experimenting and learning and doing pilots, which is to be expected. Because before you adopt the technology, you want to try it out, you want to give it a shot, and then you want to see the results and then you feel good about those results and then you want to roll it out. So there's always going to be some hiccup when you take something from inception, but then make it into production. So we are going through that journey and people are trying to figure things out. One thing that is emerging is that models is not the problem anymore. Models are very, very good. So I just want to make sure everybody understands that now one of the things that enterprise can do is instead of trying to learn and experiment with everything, because there is a lot to experiment. There is a lot of evolution happening. Every day I hear a new framework, a new technology or a new innovation that is being rolled out. The models are getting better and better. So I just want to make sure technology is not the problem. Models are not the problem is how you approach them. So one of the things that companies can do is, number one, to say, okay, if they trust in that technology, they need to say, okay. What are some of the things that I want to focus on as a company versus where I think I can leverage some help or look for people who are experts in that area. So don't try to be an expert in every single thing, because if you do, you better be a technology company than a company that builds something, manufactures something, builds cars and planes and so on and so forth. So that's number one. So don't try to be an expert. Don't try to keep up with all the trends and, and new innovations and so on and so forth. Focus on your business problem. What is it that you want to achieve? What is the roi? What is the biggest impact that you can make in your enterprise. So please start with that, not with the AI. I see a lot of companies saying, oh, we have to adopt AI. And I say, but do you have a realistic use case with, with a roi?

[A]

I guess that is the struggle with the hype you almost already feel behind and then you just want to jump in. But as most things in life, it's better to take one step back, look at the company. And as you said, I like the fact that you're saying, don'. Try to be the technology company. You know what you're good at, you know what your business is doing. You're just integrating the AI where it can accelerate what you're already good at

[B]

or get better at it. In fact, I mean, I, I always take this example. I, I say if you compare Uber with the old world of, you know, hailing a taxi with your raising your hand or thumb or whatever you did before in your Uber, just use the technology to do the same thing. Instead of raising my hand, I press a button on my phone, right? But the motion hasn't changed the company. People's desire to go from a place B is not the change. Cars haven't changed drastically or anything like that. And that motion still exists. That need still exists. If you solve that need, I think what your motion or what your focus should be, how do I use technology to make it better or better experience for the clients or customers?

[A]

I guess a lot of companies have done that. They, maybe they've actually pinpointed that use case and then in the lab it works perfectly because you're feeding it certain information in the lab. And I don't want to say you're making it easy for that model. That's not what I'm saying. But it is a controlled environment. Whereas if you go into the enterprise, realistically, the enterprise has been using data for a long time, so there are already systems in place. Is that where they may be starting to find? Because it's not that controlled environment anymore. You don't have the 100% control of what is, but you're not curating the information. And as exactly within the enterprise, whereas

[B]

in the lab, you do, you nailed it. In fact, you used a very, very good word called curated. So one of the thing is, of course, if you are an innovation leader, you're an AI leader, what is your incentive or what is it that you're striving to. You have a use case defined. So thanks, start with that, but let's say you define it. And now if you're an AI leader or if you're a cio, you have to show very quick results and say, hey, what is the promise of technology? What is the promise of AI? What are you going to do? You're going to say, I don't want to worry about how the data is structured. I'll just take the data somewhere, somewhere. And it's a very small set of data that I'll take. I will put it through an AI and a model. It will fit into the context window. I will run the semantic search or some of these things and use the model to understand the data. Give me the answer and then I can show very quickly how AI can be used. Real world is messier than that, right? Especially for large enterprise, your data is never in one place. Your data is not curated. Even if you have all the MDMs and you have all the data lakes and you have all the data warehouses, you still have structured, unstructured, multimodal data. Zoom data. We have now this podcast recording. Somebody might be listening to it. They might want to leverage this to inform their view about what they want to do next. So all that data and is constantly changing. So it's not curated, it's not static, it's not something that exists in one place. So that's why, number one fundamental problem for you is that you still have fragmentation of information and data. And what I normally say is model is not the problem. Context is, what is context? Do you have the most relevant, most timely, most accurate information at the time of decision? It's not, do you have the most information? It's not about, do you have all the information about history of the world? It's not about, do I have all the information in one place? That's not the point. The point is, do I have the most relevant, most timely, most accurate information for the decision that I need to make? Because the context windows are limited. Even if you have million tokens context windows, the garbage in, garbage out. Me, I was just, you know, working with my development team and I don't know if you have heard that saying that. I'm sorry, I'm writing you a very long letter because I don't.

[A]

Right, that makes sense.

[B]

And it goes with AI too. If you give them a lot of information, they hallucinate. Models don't hallucinate by themselves. Models hallucinate because you ask them a question. They don't have a context, they have too much information to, to process and then they make up stuff.

[A]

I guess that makes sense. So in this, I think a lot of companies have felt they're behind because they don't have enough data and that you're saying that that's not even the problem. So people could be trying to compile all of their data going back 50 years, whereas maybe in this specific industry, things have changed so much over the last five years that the, the old historical data doesn't even matter anymore. Trying to ingest, let's say, paper data, trying to get that into these systems, and that really the problem. So are you saying that companies actually need to first understand, if they look across their company and all the data they have, what is actually the most important, most accurate, most timely data, and then figure out the best way to access that data in the most efficient, accurate way before they start even running the model.

[B]

That's right, because you do. So one of the things I always say is it definitely depends upon your use case. So if you are an auditor, I would not argue with you that you might want to look at 10 years worth of data and you want to do all of that processing. But if that's your use case, but if your use case is, we launched a new product, the product has some issues, it has bugs. We need to make sure the customer experience is great. We want to have an agent that is answering those questions, fixing those problems for the client, rather than we trying to do it humanly. Very different problem. You don't need 10 years of data. You don't need all the data about all the products that you had. You just need the data for the product that you launch, the knowledge base that you have that is relevant to that product, the customer issues that are related to that. You don't need all the customer issues, and so on and so forth. So I think it's, it's, again, it's just the relevance, it's the context that matters. It's not the amount of data, it's not about where the data is.

[A]

And currently, if we look at the current state. So I think a lot of companies have invested quite a lot of money in getting their data in some sort of system. And I guess it's been, could be flavor of the month, what is the best system? And sometimes it's just what fits on your current, current legacy system. You had to go with that because you didn't have the budget to go and revamp your entire legacy system. So am I understanding it correctly that for the moment, all of these, the data itself might not be fragmented in the sense that you might have all of your receipts in one place, or you may have all of your statements in one place. But when you look at the overall picture, you don't only want to look at statements. You might need other types of data. And at the moment, you might be optimized for having all of your statements in one place. But that doesn't mean that when you try to access or ask a question, you query a question, you're not getting all of the information or the picture isn't coming to you in a clear sense because it's fragmented in the sense that the different types of data correct is not sitting in one spot.

[B]

Exactly. So I think you're one. One thing I just want to make sure, because I am party and I'm guilty to some of the things that you are talking about. So I've been in data space for a very, very long time. I don't want to give away my age, my hair, give it away already. But I would just say that the way you have to think about is people have been trying to solve this problem for a very, very long time. Since the companies existed, since they started to record information, whether it was in the books. Then you had the libraries to store the books and index the books. Then you said, oh, we have now data in computers and we have it on disk. But then you said, okay, but I want that data to be information. I want wisdom out of that data. And so people have been trying that. And all of those are, I would say, need to be applauded. But when you look back as the technologies evolved, you say, oh, why did they do that to me, I don't want to look at it that way because they did it that way, because there were technology limitations and things existed and they had to deal with what existed before you. So I think your point is very valid. But now do you have to still do that? And that's my question. I think the question we should be asking is yes. When we had ERP system and supply chain systems and all these systems and humans were entering the data, talking to the data, you need screens. And then of course, all those systems were not unified. They were not talking to each other. You needed a layer. That's where the data warehouse is. The data lake came into being. But that also didn't solve the problems because then people said, I need a BI tool to be able to access that data. But then they said, I need reports. But then they had millions of reports. Then they said, millions of reports is too many. So this problem has been going on. And in fact, people said, I'm consolidating. In fact, they ended up fragmenting more. They made the copies of the data in from one place to three places to 10 places, then consolidated into another place to 11th place. Then you had BI tools or BI reports have data. So it's been going on for a very, very long. And I think now, if you really believe in the promise of AI, which, as I said, I believe, and I'm not being disrespectful to the humans and the work that we do, but I firmly don't believe that we would now be in this world where we would be sitting in front of screens or multiple screens, because you have a Salesforce system and you have to enter something. You have a JIRA system, you have to enter something. You have a Confluence system, you have to enter something, and then you have an ERP system and you have to enter something. I think if you believe in the world of agentic AI and so on, and you feel that some of those tasks, not the thinking part, not the communication part, not where the interactions are important, not that part, but the part that is repeatable, you have screens. My mind is going to change drastically because the agents will be doing the work and they won't care whether the data was in Salesforce, was it in Snowflake, was it in JIRA or Conflict. All they need is give me the data that I need to make the decision. I will make the decision. I will act on it. Whether you have a human in the loop or in. Sometimes it could be autonomous and then they will create tokens. As Jensen will say, that this is not just retrieval problem anymore. Now those agents, like us humans, create information. We create actions, we create decisions. Those decisions now need to be also stored because now they're correct and governed. Correct and governed. Which means what is the answer? Are you now going to take all of that and put it in your Salesforce system, in your ERP system, in your. In your data lakes and your data warehouses and snowflakes and data lakes? No, you will have to think differently and say, do I need all of that? Why am I writing back into Salesforce even though it's an opportunity? Because the agents now know. And where should that agent be looking at? And do all I have to do through this hop, Create it in Salesforce, then go to Snowflake, then get it from Snowflake to the next agent? I mean, I don't think so. Right?

[A]

But. But this seems like a paradigm shift because I think everybody has been looking at. They wanted to optimize the way they store data because you get Some data that you want access quickly, some things you just want to store for a long time and you just optimized accordingly. And also looking at your spend obviously where you're going to store, but you're saying this is actually a complete paradigm shift. It's thinking about your architecture in a new way. So if you could describe this new way, I might be trying to oversimplify it, but are there maybe a set of rules that you need to look at in a completely new way before you decide on your architecture? Because I think a lot of companies sitting with a problem that they cannot scale these projects do understand that they need to take this leap. They need, they need to change something. It's not working. So what is that practical way? They need to think about this.

[B]

Good. So great question. So again, first and foremost I want to make sure all of their investments in snowflake and databricks are valid, reasonable. They need it because you can't go back. Agents will exist from now on. They didn't exist before. So but for agents to exist and to work and be effective, if that's what your promise is or that's what you believe, then you do have to re architect things. You have to rethink. So it's not just about having a model and having building some graph rag and doing that kind of stuff. Yes, it's good for PoCs, but as soon as you build those PoCs, as you to your previous point, as you go into production now you have this question to answer. How they are being governed. Where is that data going now? The data is being created, the actions are being done. I want to be able to say why that agent made that decision, like us humans. But humans you can ask right for agent, they will have to store it

[A]

somewhere because they don't have a memory unless you create a space for that memory.

[B]

Exactly. So there is a new concept of saying, okay, it's not about the data, it's about the context. So that's where I will start saying first define what type of information your agents need, what context they need to make the decision for the problem that you are asking them to solve and what are the tools they need. So you are now shifting your gear from an application world to tools, context, information and decisions and specificity of what the action and the policies are that need to be to govern their actions. So you need to make sure all of that is provided exactly at the same time, at the decision time. Because it's not good enough to ask a AI or a model or agent to Solve a problem, but then give them a framework of policies after the fact. You need to know what information is going to be now needed to be provided, what tools they need to be able to do. So if they have to go and fetch data from something, you have to think about whether I put them in one place, whether I put them in multiple places, can I get them from multiple places, but do I have the tools to be able to do those kind of things? So you have to really rethink your architecture. And then as I said, you do have to rethink and challenge yourself in saying, yes, my system of record was X or Y or Z for my sales pipeline, it was Salesforce. But now since the agent is acting, do I go and create and is that my system of record or I have to think differently? Because now the agents and they are not communicating with Salesforce, they are communicating with Salesforce through tools and lot of other hops that they have to go through. So how do we do that? So I think that is how you have to rethink this whole construct and the key focus and not because we do it. The reason we do it is that agents need context. The context needs a new architecture. If you are focused on AI, you are solving business problems through agents. You do need to make sure that you are focused on context, not the models. Models are really good, trust me, we deal with them every day. But then besides the models, you need the context. And not only the context, you need a new architecture.

[A]

And looking at companies today, this really is in a real sense a remapping. I would almost say it's remapping exactly how you are doing your business to a certain extent. A lot of companies before have created a new department and you have to map out who do you need in that department, how many people do you need, how many levels of seniority do you need in that company? And where are they allowed to go? Are they allowed to get into the financial records or not, the governance, etc. Who in the company will this be leveled down to? Each department looking at it, but it seems like this should sit higher up in the company. Who would be responsible for this remapping within an enterprise?

[B]

So there are two schools of thought and both are valid. I don't have a bias towards it. There is a school of thought saying agents are too dangerous, as humans are too, by the way, like if you have kids supervise, sometimes you supervise too much and you call helicopter parent and sometimes you let them too loose and you're irresponsible. It's exactly the same world we are going to have with agents. You will have to trust your agents. But as we say, trust but verify. On this other thing, teach the agents, supervise them for some time, but then let them go. So you will see different ways and different schools of thought. So to your question, I don't think there is a simple or single answer. I think it will depend on the risk appetite. It will depend upon the company's problem statement. So let's say it's not a catastrophic failure. If the agent makes a mistake, maybe you can live with it, but if it is a banking or a financial institution, you cannot live with it. You would want to put a lot of guardrails and a lot of training and a lot of making sure that everything is tested and everything is still supervised. And there's somebody sitting even before the agent, even if it is allowed to make the decision, there is somebody who says, I approve that decision and then act on it. So maybe the decision can be made by the agent, but the action, let's say example is it's a loan application. Merrily wants a loan. Agent can decide all given all the stuff and she's fine and we will give her a loan. But then somebody has to say, call Merrily, Merrily, are you a real human or you're another agent? And how do I verify? So again, you will see that progression. But to your point, do the enterprise will have to have a department? Yes. They will now have to think about like we had for humans. You had policies, guidelines, governance controls, all of this access, management, all of those layers that we built in the enterprise. They need to be now reconsidered, rethought and saying, because agents also you will have almost think about they're your employees.

[A]

I just wanted to say, do we need an HR department?

[B]

You nailed it. I was going there. So then, if you haven't heard already, people are now talking about it's not one agent. You will need multiple agents to do even one task. So for example, because of separation of duty, you need to have an agent that does the underwriting versus an agent that does the loan, like in banks. So you will follow some of those guidelines. But then those agents require the same context because they can't be going to different places to look for different things because then they're out of sync. Then their protocol A2A or whichever protocol you take agent to agent or whatever, needs to have the right context and the right information that needs to be there for them to be able to make sure that they're doing the right thing. There is a conversation saying there will be supervisor, agent, agents that will supervise the work of agents. So to your point, you nailed it. Yes, you do. You will have. Maybe it's a AR department, so not hr, so not human, but agent resource management or ARM something. But yes, that is how we will have to look at them, because they are doing the task. Some of the tasks that we as humans were doing before.

[A]

I've said before, a paradigm shift, but I think, if I understand you correctly, it's not about intelligence anymore. The intelligence is there, the AI models are there, the data is there. It's about how you feed into that intelligence. And I think that's not. It sounds like a paradigm shift, but really it shouldn't be, because if we look at humans, it really does depend if you. If you give somebody a fragmented information. Looking at the loan again, if you see my first five years as an adult, you might say, no way this person is getting a loan. But then if you give the entire picture and you come to the last five years, oh, that person is actually stable. They're getting a stable income. That's a completely different picture. Same person that you're looking at, same agent doing the intelligence on it, but it's a completely different picture. And it's only because of what you fed it. And it's not back to the quality of data. It's not about all the numbers in the right columns. That's not the data we. Or the fragmentation we're talking about anymore. It's all got the same label. But what are you actually feeding this agent to make decisions on? And if I'm understanding it correctly at the moment, got a beautiful room with all of your statements stacked, and it's all neat and you can find it. It's no problem. They've got their metadata, but it's in a different room from Marley's purchasing history because it's different data. And at the moment, you have to walk to all these different rooms to pull the data. And sometimes you forget about one room you forgot she made some complaints. And you might want to pull from there, where if you've got all of this in and they're all speaking to each other, and if you say Marley, it pulls up everything, and then it may decide what the agent needs to see based on the level of seniority, but at least it's pulling everything for Marley and then deciding what. What it shares with the agent. Am I even close to understanding?

[B]

Oh, you. You're very close. So let me maybe break it down for you because I think you almost define context already. So let me maybe make it in a more structured way, because one of the things people will say is, okay, do I need to know everything about marriedly? I could question. It depends on the question, right? So I don't need to know everything about married to talk to Marley or do something for Marley. I just need to know in what context. I need to talk to Marilee. What do I need to know? So first is that. Start with that. So I know the earlier world was there was a race to get all everything about Merrily, but I'm saying you don't have to do that today. You can say, if it's about a loan, I don't necessarily need to know all about everything about Merrily and your social media profile and postings and so on and so forth. I just want to know what is your compensation? What is your money that you're making? Do you own the house? So there is a basic set of principles and rules and things. And in fact, some of the things I'm not supposed to know, I'm not supposed to ask you about your race. So in fact, in some cases it's also things that you don't need to know and you should not know or should not have. Right? So that becomes important. So again, the first thing is, what is the context that you need to look for that information? So it's not all the information the relevant. The number two is when we say Merrily, do I pull up everything about Merlee? Wherever Merrily is, Marley could be a flower, Merrily could be a girl, Merlee could be a woman, Merle could be a name of a painting. No, I need to first define what. So semantically, if you are looking at it, because Merrily can be used in multiple ways, I don't need to just match semantically. I need to say for this context what information I need. So that's number two. So that I want to make sure that it's not about just the semantic meanings, but in the context of the question or the problem that I'm trying to solve, what information that I need? So it's not just semantic matching or just everything about Merrily. Number two, and I think you said it before as well, what is the timeliness of that information? Do I need to know everything about Merlee since she was born, or I only want to know five years, or I only need to know about one year, or I only need to know what she does now? So that's that what I call timeliness? Or that basically freshness or what is that context and how appropriate or what information I need to serve that context, what is the time window, the time framing of that context. Then I think you could basically say, but I just don't want to know about Merrily. I want to know if the context is about a loan. Then I want to know does Marily own a home, does she own a car, does she have loan payments? So I need to know other things about Merrily that relate to that context, how they relate to Merrily, how Merrily relates to them. And for example, if I'm doing, if you're a South African or a different country, then I want to make sure that I'm allowed to do business with that country. If you're not a citizen of the United States. So these are very different things.

[A]

It's a lot of context.

[B]

Yeah. And you need to know those relationships which are in the, you need to establish those relationship in that context then. So we talked about freshness, we talked about, it's not the semantic meaning we talked about that you need to know the relationship, how they relate to that context. Then of course you want to make sure, do I have the controls or do I have the auditability or is this source verified, Is this trusted? Do I know this is true? And so on and so forth. So you do want to make sure, is it trusted? Is it verified? Can I, can I control it? Is that, does it have guardrails? And so on and so forth. So you don't want to just give away information or Merrily. And then it's on the dark web, I mean that will be a disaster. So then you have to have some guardrails. And then of course, and you said it then even if you did all of that stuff, the old world of putting it into and sequestering it in some kind of a SQL or a lake warehouse or lakehouse or a data warehouse doesn't help because agents don't talk like that. Agents are more conversational, they understand text. So then you have to now make sure that information can be provided without still keeping it private, not giving to the model to train the data, which is another concern people have. So this is where the rags and the graph rag. But then basically you have to do something with that information to make it consumable by the AI. So that's what we call it. You have to have some kind of an AI services layer or integration layer with AI or the models and the way to prepare the data for those models. And then I think last and like we talked about this, you do have to make a decision for efficiency, even on the retrieval side. When I want to know everything about Merrily or some things about Merrily, is that a runtime or am I going to need that information again? Is it more efficient for me to store it in one place or go to all these places again and try to find it and keep it fresh? And the other part is, if now I did act on loan application from Merle, where is the decision stored? And then where is the context stored? What information that I used, regardless of where it came from, regardless of the retrievers we bulk, where is it stored and how it is stored? Because at a certain time I might need either an agent or agent or a human to look at that information saying, did we do the right thing? Is Merily the right recipient of the loan and is it the right Merily? What are the considerations? What are the criteria that was used? What are the information? Was that information verified? Where would you store all of that? That so this auditability, provenance. So that is where we say we have a multimodal database. So not only we provide you the context, the freshness, the temporal aspect of it, not just the semantic, but also preparing the data for integrating with your AI or agents, but also being able to store that data. So that's how I normally describe it.

[A]

I guess we take it for granted that as you mentioned, if you say Marley, a human, sort of, if you're in loans, you automatically know this is a person and you're speaking about that. But you cannot just assume that from an AI agent because they do look at information differently than we do. And again, we already sitting with context. As humans, we have at least 20 to 25 years of context built up already. When you enter a job, never mind your studies, et cetera, and your training within the company. And we do take that for granted. And I guess it's good to know humans still have that edge. But we can teach and we can govern this in a certain way, that we do give that at least enough context or the right context for AI agents. So we've touched on exactly what you feel that leaders need to start doing today. And it seems like there is still a bit of work for enterprise leaders to be done. They need to map this out for the AI. You've got the pilot, so you already know what you want to achieve. So I think don't disregard that. You've done the work, you've seen the pilot work because you gave it the right data. Maybe just walk it back and see what was the exact conditions and governance and context. At the end of the day, you gave that, that pilot walk that back as a map. And then you need to go and figure out how you're going to map this across your business. And it would make most sense if you then have an architecture in place that does all of this. Because I don't think people. I definitely didn't consider the fact that you are creating so much more data just from using the agents. And that needs a place to go and it needs to be searchable as well, and it needs to be auditable as well. So I do think you've given a real clear picture for what AI leaders need to do right now before implementing the architecture and contacting those companies that can help you. You don't have to be the technology company. Map out your business and the technology company can help you with the rest. Is there anything else you'd like to add for AI leaders, something we might have missed that would really benefit them in considering all of this? Because it is a lot to consider.

[B]

So, yeah, I would say if I have to leave. Final thought. I am not saying that you have to start with a big bang approach or just rethink the architecture and so on. I would say first, yes, start with a small scope, but apply an architecture approach. Even if you don't implement that architecture, you do need to have an architectural approach. And the idea should be it's not just about AI. You have to fundamentally think that the process requires a redesign. Because now the way you will do things will be much different. Your data needs, which were already or previously designed for human consumption across multiple screens and so on and so forth. Is that still the case? And it's not. If it's not, then is that. Would you follow the same process that you did before, Take the data across multiple systems, put them into. Put a BI layer, put an intelligence layer, do a search? No, because you have to now think about saying, who's the recipient of that information? And what is that? Is that just data? Because humans are very good at context, right? Because we are on that way, we are built that way. We have lived through this. Agents are almost like newborn babies. They have no contacts, they are babies. They have no context, they have no understanding, they have no memory. So that is one of the things. They have a context window which, which is finite, right? So humans have billions of neurons, right? We're talking about millions of tokens with the biggest context window right now. So there's a lot of ways to Go right. So they can't store the context, which means your focus has to be not the data, but the context. Your focus has to be not thinking about how you serve the humans, but thinking about how you serve the agent. So what kind of services, what kind of integration, what kind of architecture you need to have, they don't have memory. So you have to figure out what does that agent memory looks like, where should I store that? Because you do need to store that. So those are the kind of things that you should think about. And then the idea is model is just one moving part. I think the whole point will be the enterprise, which think about the context, think about the governance. And to your point, not this very controlled lab environment, but saying messy world. And that mess is now getting inherited by the agentic world. Agents will still live in this world and this world is messy. So your job has to be, we do have to have policies, governance, controls, operational discipline, exactly the same way we had before, but now in a very different architecture.

[A]

When you think about context, and to a certain extent your optimism has rubbed off on me because even though there is no context, you are left with a blank slate where you can build a lot cleaner context than you might have with humans. Humans enter a lot of systems with a lot of bias already built in through context, where you can with mapping it out correctly, having the governance in place, setting the rules as well as you can. You've got that blank slide to really build clean context, if that's a term, I'd say you can.

[B]

If I'm hearing you correctly, are you saying it's a good thing that the agents don't have memory?

[A]

I think we do because it's often deterministic. I think it is better to start with the clean slates because the world is messy. But I think people's context can be extremely messy. So it's a clean slate, I think is probably, especially in regulated industries, I think that's a very good starting point, is a clean slate. But you can't leave it clean. You've got to build something in there. And I'm hearing from you, it needs to be accurate, it needs to be relevant and it needs to be timely. And those do shift depending on the decisions and the intelligence you need. So you really need to map this out clearly and have that in mind. Ravi, this has been a great conversation, looking forward to more conversations. And I'm sure the leaders that are stuck in pilot phase now can look at the problem optimistically but differently. So thank you for this.

[B]

Thank you, Marlee. Pleasure was all mine and thanks for giving me the opportunity.

[A]

Wrapping up today's episode I think there are three key takeaways from our conversation with Ravi. First, scaling AI breaks down not because of model limitations, but because enterprises rely on fragmented, inconsistent information that prevents systems from making reliable decisions. Second, leaders need to define the specific timely data each use case truly requires to rather than attempting to centralize or perfect all historical data. And finally, preparing for agent driven operations means rethinking long standing data practices, including how decisions are stored, governed and synchronized across the organization. According to Edison Research, 79% of Americans aged 12 listen to online audio monthly. An estimated 228 million people. For busy executive audiences, podcast offer a rare opportunity to capture 20 plus minutes of attention with VP leaders in America's largest enterprises. Emerge reaches 1 million listeners every year. Learn how leading AI brands convert executive attention into measurable pipeline impact. Download our media [e-mail] partner that's go.emerj.com p a r t dash e r for further executive level analysis and to join our network of leaders delivering workflow impact with AI, visit emerge.com on behalf of the team at Emerge. We'll see you on the next episode.