What does an AI agent need to understand about your business before you allow it to make decisions and take action without waiting for human approval?
In this episode of Tech Talks Daily, I speak with Kash Mehdi, Field CTO at Reltio, about the move from analytical AI that supports decisions to agentic AI that can execute them.
Kash argues that leaders should begin treating AI agents as a workforce rather than another collection of software tools. A digital workforce needs training, boundaries, oversight, trusted information, and clear permissions before it can act safely.
He uses the analogy of raising a puppy. When the puppy misbehaves, the problem may be inadequate training or poorly defined boundaries. AI agents present a similar leadership challenge. Organizations must ask what the agent has learned about the business and what authority it has been given.
We discuss why model selection may be receiving too much executive attention. Kash describes four components of an agentic system: the model, tools, data, and context. Models are improving rapidly and tools are increasingly available, but business context remains incomplete across many enterprises.
Data tells an agent a fact. Context helps it understand what the fact means within a particular customer relationship, geography, policy, or business process.
Kash illustrates the difference with a pizza order. The data may confirm that someone is logged in, the model can interpret the request, and a tool can place the order. Context tells the system that it is Friday night, the customer is watching television, and they usually order pineapple and cheese pizza.
The same principle becomes far more serious when an agent is dealing with medical equipment, supply chains, financial customers, or regulated information. It must understand which entities exist, how they relate, what information it may access, and which actions it has authority to complete.
Kash identifies three requirements for safer autonomy: a governed source of truth, a live feedback loop, and enforceable permission boundaries. Trust must be built into the data and operating rules before the agent acts because the familiar human review step may no longer exist.
We also discuss how governance changes when AI can execute decisions at machine speed. A poor decision made by one employee can usually be reviewed and corrected. A poor decision repeated automatically across thousands or millions of transactions can become a business incident before anyone intervenes.
Kash shares examples involving restaurant menu launches, medical equipment deliveries, and call center offers. Each depends on current information and the relationships connecting customers, products, suppliers, locations, and previous interactions.
For CIOs preparing today, Kash recommends building context around reusable entities rather than constructing an isolated data project for every AI use case. He points to Schneider Electric as an example where one unified foundation supported sales, shipping, operations, and marketing use cases.
The conversation ends with a warning about slow data. Autonomous agents need current context because information that arrives after a decision has been made may no longer carry much business value. Kash predicts that the half-life of enterprise data will become a board-level measure.
If a smarter agent can make a poor decision faster and with greater confidence, is your organization investing enough in the context, governance, and feedback needed to keep it on course? Listen to the conversation and share your thoughts with me.
Useful Links
[00:00:00] Are you still dealing with agentic AI semantic confusion? Well, Denodo keeps your business harmonized. And by positioning your data platform into a universal semantic layer, everybody can be on the same page. So get back to scaling your business and trusting your AI agents to create proper business decisions. And you can do that by visiting denodo.com to see how.
[00:00:26] What changes when AI stops recommending an action and starts taking it without waiting for someone to click approve? Well, today I'm joined by the field CTO at Reltio. And together we're going to discuss why autonomous agents need far deeper knowledge of a business than just another clever
[00:00:50] model. And he will explain the difference between data and context, why systems built for people cannot automatically support digital workers, and how companies can engineer trust before an agent acts. And his argument is wonderfully simple. Treat AI agents like a workforce that needs training,
[00:01:13] boundaries, supervision, and permission. Or as he puts it, think of it as an enthusiastic puppy. Yes, it can also misbehave. So checking what you taught it might be wiser than just blaming the puppy. But we'll discuss all this and also govern data like feedback loops, reusable entity models, real-time information, and whether pineapple belongs on pizza. Yeah, we've got so much we're going to
[00:01:40] explore today. And many of those questions could prove harder to resolve than enterprise AI governance, but we will have a bit of fun along the way. So enough from me. Let me introduce you to my guest now. So thank you for joining me on the podcast today. Can you tell everyone listening a little about who you are and what you do? Yeah, first of all, it's a pleasure to come on here with you. I'm Cash Medi.
[00:02:07] I joined Reltio as the vice president and field CTO. What that means is my role is at the intersection of product and go-to-market teams. And I spent a lot of time with leaders, helping them successfully undertake any of these initiatives around data, AI. If I were to summarize it in one line, it's really helping organizations move from the analytics era to the agentic era. And I'm sure I can unpack that
[00:02:34] a little bit more. Yeah, there's so much I want to talk with you about this because I think for years now, technology has helped people analyze information, make better decisions. And this is analytic AI. But now we're going beyond that now. It's almost like back to the future where we're going, we don't need roads because AI agents are now beginning to make the decisions and take action
[00:02:57] themselves. So are we entering fundamentally a new era, an agentic AI era? And what does that shift mean for business leaders that keep hearing all the noise that surrounds this and are unsure of where they fit in it? How do you see all this playing out? Yeah. In short, yes, the shift is categorical. It's not incremental, right? Yeah. So it's also important for leaders to understand what do we mean
[00:03:22] by analytics as well as agentic? So analytics AI helped humans decide, whereas agentic AI decides and acts. And that collapses the loop between insight and action. And it removes the human checkpoint that every enterprise system was built around. And here's what it means in the enterprise context. So every system of record, every workflow, every approval chain in the modern enterprise assumes a human
[00:03:51] is reading a screen and clicking a button. In the agentic AI, that removes the assumption because organizations that treat this as a quote unquote better analytics will miss it entirely. And the organizations that treat it as a new operating model will win. And that's really what I see is a massive shift happening in our industry. So if agentic AI is about systems that can reason, decide, and indeed
[00:04:19] take action, what does that mean for enterprise leaders that are listening now? How should they be thinking differently about the role AI will play in their organization? Maybe over the last few years, they've been playing with some of the tools and asking questions, almost treating it like a search engine of sorts. How should they be thinking differently now? Well, first of all, I would say stop thinking about AI as a tool people use. Start thinking about it as literally your workforce you manage.
[00:04:48] A tool is something that can answer questions when you ask something, right? This is what we've been doing with all of our LLMs, chat GPD, cloud code, all the good stuff. A workforce, whereas it needs boundaries, training, oversight, and trust, and built in before it acts, not after. And that's really the big thing that leaders need to realize. I use this puppy analogy that one of our chief product officer
[00:05:17] uses on our side. So the analogy, of course, lands really well because everybody loves puppies. So if your puppy misbehaves, the fix isn't the puppy, it's you because you didn't set boundaries or you didn't teach it enough. So agents today are exactly the same, capable and mischievous. And the leadership
[00:05:39] shift is from what can this model do to what have I taught the system about my business? And what have I authorized it to use without me being there? And that's really how I would suggest leaders to think about this big shift. Such a great analogy there. And I think many organizations are focused on choosing AI models or how to build
[00:06:03] their agents. But I'm curious, from what you're seeing, are many asking the wrong questions? And what should they be asking? And what should they be preparing instead? So one thing I would tell a lot of leaders today is to think about context intelligence. So maybe there are a few studies that I could quote here. Almost entirely, you know, short answer to your question
[00:06:25] is yes. So MIT study that everybody quotes is that says 95% of agentic pilots are going to fail. And that's a great headline. It's also misleading because if you dig into it a little bit further, the failure clusters in highly customized bespoke use cases, the exact places where enterprises were experimenting the hardest. Off the shelf agents do a narrow well-scoped task. And that works because
[00:06:53] almost every organization has already ruled one out. The real question isn't which model or which framework. I break it into four spaces. So one is model, context, tools, and data. So if you think about four Legos out there, models are improving on a doubling curve because inference cost has fallen 900x in the last three years. It's getting cheaper and cheaper. Tools are abundant. You know,
[00:07:23] MCP, I'm sure you've heard about MCP alone has solved a decade of integration pain, and there is no bottleneck in either. The bottleneck is entirely in the context. Do you know enough about your business in a form that agents can use to make its decisions correct? And that's where all of the money and all of the failures are right now. Organizations chasing model selections are optimizing the part
[00:07:51] of the stack that was never broken. So what I would say is leaders really need to think about building their digital representation of the business in real life, knowing who are all your customers, what interactions they're having, what products do they have? Because all of this information is typically in an organization that has not worked on building this foundation or context. This is spread across various architecture settings. You need to create all of this context in a unified
[00:08:21] place and supply this to your agent because previously, I mean, humans, we've gotten used to this swivel chair technology, I call it. A human would eyeball a CRM screen to learn about a customer. Then we'll go to an ERP to learn about product supplier. Then we look at support tickets. We were the ones that were stitching all of this context, but agents, they need this exact digital
[00:08:46] representation. So my first ask for leaders is really while there is time still build that context, create that digital representation, help agents understand who your customers are, what products they have, what interactions they're having, provenance, history, semantics, all of these details in a unified place for agents to act on. Fantastic advice. And it's easy to see why it can be so
[00:09:13] overwhelming for many because most enterprise systems that all enterprises are using today were designed for people, not autonomous software that were making decisions. So again, what else needs to happen before an organization can maybe trust an AI agent to make decisions or take action on its behalf? Because again, that is one of the big bottlenecks. I've been brave enough to trust it, to go out there and make those decisions, right?
[00:09:39] Absolutely. And this is a great question and I have a great answer for you. So three things have to exist before autonomy is safe. A governed source of truth about your entities. Of course, that's what I was explaining to you just now about knowing your customer's product and all the good stuff. A live feedback loop that captures what agents learn while they work and then enforce permission boundaries on
[00:10:03] top of both. So most enterprise systems were built for a human to eyeball a decision before it goes live. The review step disappears with agents. So the trust has to get engineered into the data and the guardrails ahead of time, not caught after the fact. And this is exactly, you know, it's not a tool treated like your workforce. And that means unified governed data, your agents pull from instead of
[00:10:31] guessing. An ontology that acts as a permission layer. An agent can only traverse as far as it's authorized to go. And a governance loop where agents learn, you know, gets whatever the agent is learning gets validated and folded back in rather than dumping raw data into a knowledge graph and left to be bloated out of control. And that's exactly where I see these actions need to happen.
[00:10:59] And we often hear that AI needs better data. Last year, a lot of tech conferences, they were saying no data and AI. And you tend to talk about AI needing context as well. So what's the difference here? Why is context becoming so important? Just for people listening that are unaware of the importance of this too. Yeah, this is great. And I got tons of examples for you. So data is something that tells you the fact. Context tells you what the fact means. And it's genuinely a different problem.
[00:11:29] Like imagine, you know, I like pizza. I like to order a pizza. So if you think about the four Lego framework, right, your model, your context, tools and data, right? So take this pizza example. So the data space knows I'm a logged in user into a system, right? The model space translates my sentences into like an API call and the tool space places the order, right? Because you have systems
[00:11:58] and applications out there. You know, I type something, it places the order. Those three are mechanical. And we have solved these problems years ago. The context space is really knowing that, hey, it's Friday night. I'm in front of the TV and I always order pineapple cheese pizza. Maybe this is a debatable topic about whether you need pineapple on your pizza or not. And that's the one nobody's, you know, build systems for. And this is where the differentiated
[00:12:27] value sits today. It's just really knowing your customers. So you could personalize things, you could understand and take the demand in. Or I can also give you a sharper example. If you think about this question that you could ask an agent, like how many clients are over 50? It's a simple query. You know, I could query it against the date of birth field and I could get the
[00:12:50] right answer. But if I ask the same question a bit differently, how many clients will retire in the next 50 years? This is a completely different problem because now you need to know that what does retire means in each geographic location under each HR policy? That's not a data question. That's a meaning question. And meaning is exactly what most enterprise estates were never built to carry.
[00:13:18] And that's the change. And you're going to see more and more of this as more AI capacity is brought into an organization. Leaders are quickly going to realize, hey, these are meaning questions, not just data questions. And that's how I would define the difference between your data and context. David. And something else I'd love to bring in here is how does structured, governed enterprise data,
[00:13:42] how does that provide the foundation that enables these AI agents that we're talking about here to better understand customers, products, suppliers, and even business relationships, and do all that well enough to make those reliable decisions? Anything you can share around that too? Yeah. So it gives them really a map before they start guessing. So think about the ontology and semantic layer as a lightweight map an AI agent can consult first. So it really gives them the
[00:14:11] ability to understand what entities exist and how do they relate with one another. Entities meaning you have a customer entity, you have a product, you have organization. How do they relate to one another? Or if somebody is living in a household, what entities exist and how do they relate with one another? And then what am I allowed to touch? Because this is where you need to start thinking about governance boundaries. So only after that does the agent decide how far down it needs to go to answer
[00:14:41] the question straight from the unified data or out to the data lake or in rare cases all the way down to the source application. So we call that the progressive targeted retrieval and it matters because it keeps the agent inefficient. It keeps the agent efficient and it keeps it accurate. A customer profile isn't just today, you know, demographic anymore. It's risk score. It's a summary of last 10 support calls that the customer made.
[00:15:10] It's every product they've bought and who else bought that product, right? So knowing all of this information from different dimensions. That richness is what lets an agent tell the difference between a consumer buying for themselves or an individual purchasing on behalf of a company in one lookup without a person in the loop explaining the difference. And that's really how I would think about that.
[00:15:37] And as organizations move from AI that informs their decision making to AI that can actually execute them, how do governance, trust and explainability, how do these things need to evolve? Because IT can be very rigid and stuck in this is how we do things, but presumably this needs to evolve too, right? Yeah, it's a great point because, you know, I've spent several years in the data governance space and governance has a bad perception sometimes.
[00:16:05] It's sometimes the four-letter word love or the four-letter another word that I'm not going to call out on this on air. But it has had this negative connotation. Like anytime you talk about governance, people thought like it's police action. You know, you're here to define rules, boundaries, how I should be using the data. And in most cases, I've seen governance being a compliance checkbox.
[00:16:32] Now with AI, we're seeing an inflection point in this space because it becomes the thing that makes autonomy possible at all. And this is very exciting for governance people. So when a human executes a bad decision, you have a conversation, right? You could talk to somebody, you could course correct that. But when an agent executes a bad decision at scale, you have an incident at hand and it happened faster than anyone could intervene.
[00:17:01] So I'll give you three real examples because, you know, we're working with some of the most complex, largest, global Fortune 100 type of organizations that are heavily regulated. But also they're going after both offensive and defensive use cases. So to start with one example, a quick service restaurant chain models new menu items like in a unified data before a single burger gets shipped.
[00:17:27] Checking ingredients approval and regional supply chain fit ahead of the launch because I need to understand like, you know, whether I could support this launch or not with the ingredients that I have at hand and the supply chain that could fit in this particular model. Get that wrong. Get that wrong. You're not fixing one order. You're fixing millions of units, right? Same thing in another industry, a medical device manufacturer ships CT and MRI machines, right?
[00:17:55] And if a machine lands at the right hospital, but the wrong loading dock, that's thousands of dollars to redo. Accuracy is in optional today just because AI is now the one acting on the data. And another one could be, I'll give you an example of call centers. You know, they only want to pitch a new product if the customer's last interaction was positive.
[00:18:18] If you had call in complaining about something that you haven't received or something's broken, not working, it's highly unlikely that you're going to get pursued and, you know, buy anything from this company unless and until your issues resolve. Which means unifying data across support, marketing stack that were never built to talk to each other. So explainability in each case isn't a nice to have report after the fact.
[00:18:46] It's the governed data, the ontology enforcement that made the correct action possible in the first place. And this is not just, you know, across these three industries that I've described, but you're going to see more and more applicable use cases across the board. And it's really an exciting time where leaders can think about these real-life problems and center how they're going to approach data governance. Because it's no longer a compliance checkbox.
[00:19:14] It's really an inflection point today that can really enable autonomy for your organization. And I always try and give people listening actionable takeaways. And if there is a CIO listening today trying to keep up to speed with things and looking for tips and advice there, can you name a few things that they should be preparing for now for their organization, for autonomous AI agents? Anything you think they should particularly be focusing on right now?
[00:19:44] Yeah, absolutely. I'll stand the obvious. So first, stop optimizing your models or tool selection. Start building your context layer. And that's the actual gap. The model and tool markets are already solved for you. So you're going to have abundance of access to models and tools. Second, I would say, unify your data around entities, not around use cases. Build the customer, the product, supplier graph, right? Well-governed.
[00:20:12] Every subsequent agentic use cases gets cheaper to turn out. And that compounding is a real ROI story. So cost per use case goes down over time. Total cost of ownership trending down is the honest signal that you've actually built a foundation, not just bought a tool. Third, I would tell you is build a feedback loop now while the stakes are low.
[00:20:36] Let agents observe proposed changes to your ontology for real interaction, structured or unstructured. And govern those proposed changes before they go live. Bottoms up, observe ontology beats a rigid top-down one because real business vocabulary is never as limited as the first version of your glossary assumes. So I would say think about these things where start building your context layer.
[00:21:06] And this is going to bring in a lot of different use cases for you because as you connect your data, as you define those relationships between your entities, your customers, product, organization, or whatever industry that you work in, whether you're in health and life sciences where you may have HCP, HCO, healthcare, professional healthcare organization. You will start to understand your data more effectively because now your data is not just simple facts.
[00:21:36] It's also giving you the context because each of these relationships carry trust attribute that gives additional meaning for agents to act on. And I'm curious, as all this gathers pace and organizations evolve and change, what characteristics do you think will define the organizations that successfully embrace autonomous? I'm going to say over the next few years. Do you see anything in particular that's going to start happening and taking shape here? Absolutely.
[00:22:05] I already see this happening. I was recently hosting one of our agentic executive roundtables in our customer marketing initiative. What I've seen is this is already happening. To answer your question, there will be the ones, organizations, right, who treat unified governed data as reusable infrastructure, not just a product. Not just a one-time project. So that's one. I'll give you an example from Schneider Electric as the pattern.
[00:22:33] So they started with one use case, upsell and cross-sell in B2B sales. Then reused the same unified foundation for shipping accuracy and back office operations. Then again for marketing segmentation. Same foundation, three different departments, three different returns, and the cost of each new use case dropped because the foundation was already there. So the losing pattern is the opposite. So a new point solution for every use case.
[00:23:02] Agents gets bolted onto the same fragmented estate. That's been fragmenting for 20 years. Expecting a smarter model to somehow compensate, it won't. So better models widen the capability reliability gap. They don't close it because a smarter agent making a decision on bad context just fails faster with more confidence. And that's something that you have to watch out for.
[00:23:32] And like everyone, you will see a lot of things online. There's so much noise around agentic AI and agents. And a quick scroll down your LinkedIn feed, you probably see a few things you agree with and many more that you don't. And from all those conversations that you have, what would you say is the biggest misconception about autonomous AI agents out there? Anything that springs to mind there? Yeah, it's a simple one for me. That's the constraint. People thinking that the constraint is the AI.
[00:24:02] It isn't. Models have gotten so capable, they're now outpacing organizations' ability to absorb them. And inference cost has collapsed 900-fold in the last three years. So the constraint has never been the intelligence. It's whether your business has made its own knowledge accessible enough for that intelligence to act on safely. That's the big question that you need to be asking.
[00:24:27] Which really goes back to, are you building that context layer so your agents can confidently act on this context, knowledge about your business? And I'd love to have a bit of fun with you here. If I was to pull out a virtual crystal ball and ask you, what is one prediction that you have about the future of enterprise AI? And something that might surprise a few people, what would you say? So one thing I would say is, we call it slow data, and I'll explain what that is.
[00:24:57] Data that arrives late is about to become data that doesn't count at all. So not just data that's useful. You know, the old warehouse pattern, refined data where hours or days turn it into a report is going to lose ROI fast because agents need immediate access. The way humans, you know, never did. They need to act in real time. So I put it as if it's late, it's dead, right? You should put this on a T-shirt, baby. If it's late, it's dead.
[00:25:27] The half-life of your data is about to become a board-level metric. And because agents are going to act in real time, they need all of this context foundation, just like we did back in the day where we were just sitting on this wheel chair, looking at all of these different screens, making the connections. Agents need to have this in real time. And that's really the big difference today.
[00:25:53] I'm going to steal that idea and make some money selling those T-shirts, I think. And if we were to take everything that we've talked about today, put it all in an LLM there, if listeners could take away just one key message from everything that we've talked about other than don't put pineapple on pizzas. Is that anything that you'd ask them to or that you would like them to take away? Yeah, absolutely. So if your agent today is misbehaving, don't blame the agent.
[00:26:21] Roll up the newspaper and check your own data foundations first. Every trust problem, every hallucinated answer, every wrong action an agent takes traces back to a context gap that you can actually close. That's what I would say. And also, order pineapple pizzas, please. Oh, there's a whole other podcast waiting to happen there. We'll have some comments on that, I'm sure.
[00:26:45] Latney, for anybody listening wanting to learn more about Reltio, how you're preparing enterprises for autonomous AI agents. We've just dipped our toes in the water today. For people that want to get a little bit deeper on everything we've mentioned, where should they go? Where would you like me to point them? So great resources, just relteo.com, where you can learn a lot about our customers, the products that we have today. And then also exciting is we've joined the SAP family.
[00:27:14] So Relteo is now an SAP company. And there's abundance of opportunities out there because we're working with some of the most largest, heavily regulated, complex organizations at scale. So you can learn about a lot of use cases from our website. We also, I would point you to, if you're interested in getting more acquainted with the Relteo ecosystem, we have a data-driven conference that's happening in 2027 in March. Guess where it is, Neil?
[00:27:44] It's in Orlando, Florida. Not in Vegas. So you can join that and learn directly from our customers, partners, and our other participants there. And the date for the conference itself is March 2nd and 4th, 2027. So I would highly encourage you to check out our website. And if you're interested, join that. We will have several sessions happening out there, especially some of the innovation that we're bringing to the market,
[00:28:12] as well as you can hear firsthand from our customers. Well, I applaud you for choosing Florida over Vegas there. Excellent stuff. I think your passion for helping and preparing enterprises for autonomous AI agents and busting through the noise really comes through today. So I'll add links to everything that you mentioned there, including the big event coming out next year as well. So please check that out. Feedback to me. Let me know what you think for everybody listening, including pineapple pizzas.
[00:28:42] I think that is a debate that will go on. But more than anything, thank you for bringing all this to life today in a language everyone can understand. Appreciate your time. Yeah, of course. Thanks for having me, Neil. I loved how Cash left leaders with three practical priorities in our conversation today. Build a context layer, organize data around reusable business entities, and then create a governed feedback loop while the stakes still remain manageable.
[00:29:09] And the wider lesson here is that increasingly capable models cannot compensate for incomplete knowledge of customers, products, suppliers, policies, and relationships, because they could end up making the wrong decision faster at machine speed and with greater confidence. And I also loved his warning about slow data. When agents act in real time, information arriving hours or even days later,
[00:29:39] they've lost much of its value. And I love that phrase, if it's late, it's dead. It may indeed need the t-shirt treatment, although the pineapple pizza campaign, I think that is something that will divide the audience. So a big thank you to my guests for bringing energy, clarity, and practical examples to the conversation. You can find all the resources he mentioned there at reltio.com. Connect with him on LinkedIn. But over to you.
[00:30:05] If an AI agent misunderstood your business tomorrow, would you know which missing piece of context caused that mistake? As always, Tech Talks Network, if you want to contact me, I will be back again tomorrow with another guest. But I'd love to hear from you. So keep your messages coming across in. Everything from these t-shirt ideas, pineapple pizzas, or a genti ki. Whatever it is, I'm nice and easy to find. But that's it for today. So thank you for listening as always.
[00:30:35] Bye for now.