Giving Enterprise AI the Right Business Context With Sphinx AI

[00:00:00] The leading issue of agentic AI in businesses right now is ensuring agents act with compliance guidelines. And Denodo applies guardrails across your entire data estate. By aligning your company's data infrastructure under one system, these guardrails perform consistently across your platform. So start scaling your business and start with Denodo. Simply visit denodo.com to learn more.

[00:00:30] What happens when AI gives a polished answer using the wrong definition of a company metric? Well, in today's episode of AI at Work, I speak with the founder and CEO of Sphinx AI. And we're going to talk about the business context that is missing from many enterprise co-pilots and its AI agents.

[00:00:56] He will explain today why access to documents, dashboards and databases doesn't automatically teach a model how an organization defines revenue, customer risk, operational performance or indeed success. So today we're going to take a deep dive on context graphs, model independence, permissions, knowledge drift and the human review needed when a company truth begins to change.

[00:01:24] And he will also share a retail example of where two teams used revenue per store very differently, which left AI to choose a plausible but commercially incorrect definition. We've got a lot to talk about today, but if you're interested in how leaders can tell whether their AI agents understand their business for allowing them to act, you came to the right place. So enough from me, let me introduce you to my guests.

[00:01:54] So thank you for joining me on the show today. Can you tell everyone listening a little about who you are and what you do? Yeah, absolutely. So I'm Rohan. I'm one of the co-founders of a company called Sphinx AI. What Sphinx does is it's an AI research firm and we broadly work with our customers to help them build out context as a platform for their AI agents. Now, everyone talks about context. I think it's like relatively obvious at this point that good agents need a foundation of good context.

[00:02:24] But the hard thing about context is that it's kind of like an empty box problem, right? They say, OK, you need good context. So where does it come from? How do you build it? Who's in charge of building it? How do you govern it? How do you make sure that context doesn't just turn into the next mess that you have to manage? And that's really where we come in. We treat context as a training problem. So similar to how some people in the past and even now are treating models as a training problem, reinforcement learning, fine tuning. We treat context the same way. So you can think of what Sphinx does is we come into our customers.

[00:02:54] We look at the work that they're already doing. Very technical work, very non-technical work, everything from meeting notes and PowerPoints down to SQL code and query history. And we train on it. And from the training algorithm, what's distinct about us is instead of outputting another LLM or another kind of black box solution that you can't really interpret, we output a human interpretable context graph that acts as the foundation for all of your agents. So we're deep in the context space. And that's broadly what the company does.

[00:03:23] Before that, I ran AI teams, various institutions in financial services. Prior to that was a bona fide AI researcher at CSAIL at MIT. Awesome. I love what you're doing here. And yes, you're right. Context is a huge topic right now, especially in any conversation around AI. And before you joined me today, I was doing a little Googling on you and I quickly came across a line I absolutely loved.

[00:03:46] And you argued that enterprise AI can often sound intelligent while actually missing out, missing how the business itself defines the truth. It's such a great point. But what does that failure look like in everyday decision? You've probably seen so many different examples there. But for people listening, let's bring that to life. What does that look like inside the average organization? Yeah. So definitely, I think the root cause of this is that a lot of people use ChatGPT or something similar in their daily lives.

[00:04:14] And for most people, their daily lives are contextualized by the real world. So you'll say, what should I do today? Where do I eat dinner? And all of that information is in the public domain. It's on the internet. And it's relatively good quality answers coming out of these systems. So you get biased by that. Then you go into the office and you say, well, let me try that here too. And you say, which of my customers are going to be churning soon? Or which are at risk so I can go and focus on those deals and make sure they stay alive? Or which of my product launches is going well?

[00:04:41] And when you start asking questions like that, the trick is that this information is no longer in the public domain. It's in your company. And even if the AI has the ability to access the information, right, like it's connected with MCP or some of the similar tools to your database, it doesn't really understand how to do that work. That work is completely bespoke to your company, even though it may seem somewhat generic. Everyone thinks about their customers. Everyone thinks about churn. Everyone thinks about costs and revenue.

[00:05:09] The way in which you think about it is almost certainly distinct, even from a very similar company across the street. Right. And so that knowledge is what's missing. And without that, you'll get answers that are extremely generic. Right. They'll kind of regress to, you know, what you can achieve by Googling. If I'm like some of our customers are in the retail space. So if you say like if I just go on the Internet and say, like, what does cost of customer acquisition mean in retail? It'll give you something and you just copy that and say, yeah, here's my definition. It'll look fine to an outsider.

[00:05:37] But then if you're actually going to go make a decision based on it, it's amateurish. It looks like what you get from hiring an intern straight out of the school, which is effectively what the AI is. It has a lot of knowledge, but it doesn't actually understand your business. What makes your business special? What makes it tick? And so that's really where we see the gap. And then increasingly, as people start interacting with these outputs in their enterprises, they realize that, you know, they need to do something to make it better. And frankly, they can make it better.

[00:06:01] There are ways like in an ad hoc setting to say, hey, you know, like let me write up a markdown file of context for my agent. Let me give my agent custom tools. And so on a case-by-case basis, you can kind of manually chip your way towards having those agents perform.

[00:06:17] But then when you think about it at the enterprise scale, the multiplayer scale, right, where you have teams working with each other and an agent built by one team being used by another, it then becomes prohibitively expensive to say, I'm going to take every single failure mode one by one and patch it up. Right. You need something a little bit more comprehensive. Yeah, 100%. And we will have many people listening from around the world, inside companies thinking, well, we already have databases, documents, dashboards, data warehouses.

[00:06:45] So from that side of things, if we look under the hood here, why do all these systems or many of these systems fail to provide this context that an AIA assistant or agent actually needs? Yeah. You know, I think it's because they were designed for a different era, right? And it's not like two years ago, three years ago. It's not like that long ago, of course, but things have changed quite a bit. The way I would, the biggest change I would kind of describe is that the way in which machines interact with these systems of record used to be very deterministic.

[00:07:13] And now it has become very probabilistic. And what do I mean by that? Well, if you were using a database four years ago, there was absolutely zero room for interpretation, right? Your code that's accessing the database would say, I want this column from this table aggregated in this way joined to this other metric.

[00:07:31] And so there would have to be some kind of human intelligence between the end user, someone making a business decision, and the data, like a data analyst, data scientist, who is going to be the one to say, okay, like, here's how I do this translation. And the machines themselves are doing very deterministic things. Now, when you make the machines do more probabilistic things, you introduce a whole new set of problems because the machines need to understand the data.

[00:07:55] They need to understand the database well enough to choose the right things to query, choose the right things to look at. And then once they have that information, choose the right way to interpret it in order to give it back to the human and say, like, here is what you actually need to help you make a decision. That's what's missing, right? We've offloaded a lot more responsibility to machines in the age of AI. And I think sometimes we forget that our databases were meant to be used, frankly, by like an expert, right?

[00:08:21] And not just a generic expert, not just like someone who knows about databases, because that's a very generic skill, but someone who knows what's the data model of our company and how does the data ultimately relate to how we do business. I think, you know, Alex Karp was on CNBC saying something to the effect of like, oh, like, use the word alpha, right? And I think that's a great way to think about it, right? Obviously, like coming from the hedge fund world, like I use that word all the time, but like, it basically means like, what do you do, right?

[00:08:49] That is differentiated from the rest of the market, right? We call it beta if everyone's doing it and alpha if you're the only one doing it, right? And alpha basically amounts to something that GPT or Claude can't know because it's proprietary to you. It's like, what does your business do? Like, how do you keep your data so that you can find insights that allow you to serve your customers better, right? And only you know that or the people on your team know that. And so unless you can capture that knowledge, it's not possible for agents to really deliver at the same level as an expert human.

[00:09:18] And I think you've captured the problem around context beautifully when working with AI. And I've got to ask, what is it that belongs in a company-specific knowledge layer? And how do relationships, definitions, rules, and institutional experience, all these things that we almost take for granted that just seamlessly happen every day? How does these things change the quality of an AI answer? Yeah. Yeah. So I can be very quantitative about this, right?

[00:09:45] Like, for the median person working with us, like, what they normally start with is, like, a rag-based approach. A rag-based approach is, like you're saying, and this is not just us, right? There's, like, even companies as big as Anthropic have studies on how they've thought about this problem. And so the rag-based approach is kind of what you said. You have all this work that you've done already, and you say, okay, like, let the AI look at it and kind of figure out what to do next. And typically, this doesn't work very well. The reason for that is this information is raw. It's unsynthesized. There can be conflicting pieces of information.

[00:10:14] There can be multiple ways in which people are doing the same thing within the same company. And so rag is just okay. And then once you synthesize it into a proper context layer, you see that, at least for us, the accuracy tends to go up on the order of, like, 35% to 40% on existing enterprise benchmarks. Things like, you know, here's a question that our business leaders often ask us. Can the agent answer it correctly? And so there's a huge jump in accuracy there, right? We've seen, in some cases, that jump be as large as, like, 80%. Obviously, that indicates that the system they're using before was not very good.

[00:10:45] But there is a massive quantitative increase in accuracy by curating your context. Now, the second part of this is, like, what does it kind of qualitatively mean and what belongs in the enterprise context layer? And what I find fascinating about this is it depends on the company. Different companies have different ways of modeling their data. They have different ways of modeling their process. They have different parts of the process that they think are special. So if there's a part of the process where you are just like, you know what? We just do the industry standard thing.

[00:11:14] Like, we don't do anything special here. Maybe it doesn't need to be in your context layer. It just can be, like, a little bit of just saying, you know, we do the standard thing, right? And maybe GPT already knows that. You say, we do this standard procedure off of, like, out of a textbook. And there's no need for us to have something special here. And that's okay. And another business in the same line may say, no, actually, like, you know, we have a very proprietary way to do this kind of modeling. We have our own marketing mix modeling technique that's not just pull some Python library and run the numbers.

[00:11:41] And obviously, marketing mix modeling being a very key thing that people do with their data to translate it to revenue. And so if you have that proprietary view, great. Then that belongs in your context layer, which is why we see this as a learning problem more than, like, a documentation problem or anything else. Because every single company is unique. And by studying how they actually operate, machines can tease out and reverse engineer what is important to remember and what is important to give to agents.

[00:12:05] And just to further bring to life the problem that we're talking about and all the issues that we're raising, and I don't expect you to name any names here. But are you able to share an example where maybe two departments use the same term or metric differently and an AI system produced a very plausible but commercially wrong conclusion? I suspect you've got a few stories there. Yes, I can give you one, right, which I thought was particularly insidious because it took me a while to even understand what's going on.

[00:12:34] So retail company metric called revenue per store. Seems kind of obvious. Like, I would think that I could compute this myself without knowing anything, and so did the AI. The trick is there are two definitions of revenue per store. Revenue per store, just as you would expect, dollars divided by number of stores. And then revenue per open store. So if a store was closed on a certain day or was, like, being maintained so customers couldn't come in but the employees were there on a certain day, you don't count that in the denominator. Now, both of these metrics actually make a ton of sense to me, right?

[00:13:03] If you're on the real estate team for this retailer, like, I don't care if the store's open. I'm still paying rent on the place. So, like, I need to understand, like, are we being efficient in this sense? But if I'm thinking about it from an operational efficiency standpoint, like, well, if the store's closed or, like, no one's there, of course I don't expect to sell anything. So, like, my revenue per square foot or my revenue per store, it's, like, not that important if the store is closed. Now, obviously, like, people aren't having complicated discussions about this because there are different teams thinking about it. So when you bring in an agent and the agent has, like, a rag-based approach, it'll just kind of find something.

[00:13:33] It'll basically at random pick one of the definitions and just use it. And the trick is, like, it's very, very hard, especially if you're not, like, a data scientist or a data engineer, to even understand what's going on. It'll say I'm using a metric called, like, revenue per store and there's a bunch of code and, like, something comes on. You're like, okay, cool. And you find, well, this doesn't align with what I saw when I had my people do it manually last week. Or, like, it doesn't really make sense to my intuition, like, what's going on. And then you send it back to the data guys. They do a deep dive. They're like, oh, it used the wrong metric.

[00:14:02] And, like, so it just is a huge mess. And so this is, like, kind of the quintessential example where I think context could have solved this. Because you could, and what we ended up doing, it's very obvious. You just have, like, literally three sentences of context saying there are two ways to define this metric. One is per open store. One is for all stores. Here's when you use one. Here's when you use the other. And you give that to AI. AI is plenty smart. You give it that description. It's able to route each person's question to the right place. You can say, well, you're on the real estate. You're on the ops team. So I'm going to route your query to the ops definition, right?

[00:14:32] And that's fine. And it works very well. You just need that explicitly materialized. And so learning that information does seem to resolve the problem. And I think for many leaders, model independence, that increasingly sounds attractive because they want choice. So how can a knowledge layer remain useful across different models without reducing everything to the lowest common denominator? I suspect you've seen a few things around that too as well.

[00:14:58] Yeah, no, look, I think it's pretty critical to be model independent because I don't know what's coming tomorrow, right? Like I'm deep in this space and I have no clue what agent will be good tomorrow, what model will be good tomorrow. There's too many factors, right? There's the cost of compute. There's availability of compute. There's the quality of open source models. There's the security of open source models. Like there's so many things floating around here that are all evolving on a day-to-day basis.

[00:15:23] So first of all, completely agree with you that enterprises should be very, very aggressively pursuing abilities to switch between any of the moving parts of the system at will without getting stuck in like kind of a end-to-end solution that sticks you into one part of the stack. Now, what does a context layer have to do with model independence? I think a context layer fundamentally gives you the ability to swap between models. There are two places you can put intelligence, right?

[00:15:49] You can either put intelligence in the context layer or you can put intelligence to the model itself. Putting in the model itself would be something more like fine-tuning the model, reinforcement learning, at some limit training your own model that understands your business really well, the way that ChatGPT understands the broad world. And those are very attractive, flashy options. They seem cool. The risk is, of course, that then you're stuck with that model. And if you want to upgrade to something new, a new architecture, you're going to have to retrain it, spend a lot of money, figure out your infrastructure again.

[00:16:18] And so the way we see it is the context layer allows you to have almost a plug-and-play model agnostic layer that can plug into GPT, can plug into Claude, can plug into Kimmy if you're using these open source models and provide it with context. The real trick here is that all large language models today can communicate in plain English, right? Or whatever language you use. And so if you're able to represent your semantic context in plain English, it will work with any model.

[00:16:47] And above and beyond working with models, it also works with humans. So people can actually go in and inspect it and say, this is good, this is bad, give feedback, right? And so we see using natural language as the medium for context. It's just a really, really win-win-win situation across a couple of axes, including interpretability, auditability, and, of course, model independence. And the tools we're talking about here are incredibly powerful and create so many opportunities.

[00:17:11] But how should businesses better govern who can define or change the truth that an AI system relies upon? And particularly when experts, many will disagree or the underlying business changes. There's so many different things that could change there. So how should they better govern this? Yeah, so I think that the fundamental tension here is that while people may want there to be a single truth that they can just write down and say, cool, we have the truth. It's done.

[00:17:40] The reality is that the enterprise truth drifts very rapidly. I'll give you a quick example of this. Like, Anthropic, right, who have built some, like, a context layer internally published a blog post about this where they basically said, we had a bunch of our experts sit around and document, like, our context. And we found that a certain agent they were optimizing for was now 95% accurate, wonderful in production, good. And then one month later, they found that the accuracy had fallen to, like, 65%. Why was that?

[00:18:07] Well, the context, while very good a month ago, things had changed, right? Like, things had drifted, business processes had changed, priorities had changed. And that all needs to be captured. And if you don't capture that, the agent is just going to be kind of like someone who has paid zero attention for the last month at your company, which, as you can imagine, is not a very effective employee at all. And so the tension here is it is going to change, right? So you can't just say, like, we're going to do it right once. You need to have a process for drift management, for change management, for understanding who's going to be approving those changes.

[00:18:37] With that in mind, I do think it's very important for people in functions like data management, data stewardship, subject matter experts embedded within different teams to kind of take on the role of context steward. Now, tools can make this easier. Like, our tool, for example, like, we try to make this as easy as possible. And, like, obviously others do as well.

[00:18:59] But there is fundamentally a human component where a human needs to understand what has changed in some slice of context today, this week, this month, that's relevant to me, right? So I'm the expert on our marketing analytics. And I see, okay, our marketing analytics context has changed. We're no longer approaching our attribution as a frequentist problem. We're now doing it as a Bayesian problem. Okay. Is that correct? Yes. Okay. Good. Now the AI can use that context. And we were maintaining that it has the right knowledge or no, it's wrong. Like, actually, like, roll that back.

[00:19:29] Why did it learn that? Oh, it's because so-and-so told it. Well, that guy is not authorized. Like, so we're going to kick out his update and tell him not to do it again. Right. So you need that level of control. And the way I see it is it's almost akin to how we think about code. Code has had reviews forever. Like, obviously, humans used to write most of the code and then other humans would review it. But now we have machines writing a lot of the code, but humans still, at least in theory, should be reviewing it. And I think context is in the same bucket.

[00:19:56] We didn't really review it before because it just lived in someone's head, so it was hard to do so. But at this point, now that it's being made explicit, a review process needs to exist to manage that change. And the great thing about adding greater context is it will massively improve accuracy. But the more risk-averse listing will maybe highlight how it could expose sensitive institutional knowledge to more systems and more users. So how do you see organizations better managing a risk like that? Yeah, no, absolutely.

[00:20:24] So I think there are two things to say about the risk management side here. One, let me answer your question first. So like any other kind of information, context requires permissioning, right? So you should be able to say for a given human or a given agent, what context is it allowed to retrieve? Now, you do want to be careful here, right? Because you can't just do it in the blanket way of saying, like, you're not allowed to see this context, so we're going to pretend it doesn't exist.

[00:20:51] Which is what typically permissioning looks like in a traditional software system, like a database. You go to a database, you say, I want table XYZ and I'm not permissioned. It'll say, I don't even know if table XYZ exists, which is a security thing because you don't want people to be able to attack and figure out what's there. On the other hand, with context, it's often better to say, hey, yeah, this does exist, but you're not allowed to see it. So that you don't really get the agent just hallucinating and saying, okay, like, I'll just make it up. But rather the agent going back to the human and saying, sorry, like, I am not allowed to see this. Please go ask so and so if you want permission.

[00:21:19] So there are some subtle differences, but you do need that full role-based access control story on what your context layer is. Without that, it does become a little bit of a risk. The second component to this is that at least the way we philosophically think about it is that the context layer has to be a virtual layer on top of existing systems of record. What I mean by this is, let's say you have a database, right? Your context layer shouldn't copy the data out of the database. It should just teach agents how to use the database.

[00:21:47] And what's the advantage of this? Well, your database already has permissions and it already has security. So if an agent goes to that system, it will already hit the existing rules and regulations of the road that you've already put in place over the last however many years of using a database, which for many companies is like 20 years of history of actual controls on these systems. And so by leveraging the controls you already have, you're able to get a lot of security and you're also able to kind of layer security in an intelligent way. So as an example, right?

[00:22:16] I might want agents to know that a certain piece of data exists in the company. Maybe there's like personal information on customers, but I may not want those agents to be able to access it. So I would tell them in the context layer, like, yeah, there's a table with our customers phone numbers. And then when the agent goes to query it, the database says, sorry, you're not allowed to see that. And this actually leads to a really good experience where now you actually get the proper feedback loop of the agent saying, here's what I would like to do, but I don't have the permissions to do it.

[00:22:44] And therefore, like you need to go ask this human to do it because this is a sensitive operation and machines ought not to be doing it in our company. And you can actually get that level of nuance, right? Which I think is like the real holy grail where you have proper secure interactions between humans and machines with good handoffs as well. That's one aspect of risk aversion. The other aspect of risk aversion is I frankly would say that the biggest thing about context that's important, like if I were using it right in my past roles,

[00:23:12] is not that it makes accuracy higher, it's that it makes mistakes much less painful. What I think is horrible is like if someone comes up to you and says, hey, like, you know, someone on your team comes up to you and says, hey, I made a mistake. The agent did something wrong. Yeah, my explanation is that my AI made a mistake. Like I probably fired the guy or like, well, that makes no sense. Like, I mean, like, what's the root cause? Like, how are you fixing it? How do you know it's not going to happen again tomorrow?

[00:23:39] Like, and so what context gives you is the ability to basically say the AI made a mistake. Here's why it made a mistake. It used this piece of context that was erroneous. Here's where that erroneous context came from. If you have the proper governance process, you can even say who approved that piece of context. You call that guy up. You say, hey, why did you approve it? Am I missing something? If you say, oh, no, my bad. It was it was wrong. You go fix it. One single source of truth. And you say, OK, this is no longer the context. We've changed the context of something that's more accurate.

[00:24:05] And now you know that in the future, and you can even benchmark this, of course, that agents are not going to make that mistake anymore. So for me, the real benefit is not just like, yeah, high accuracy on day one, but it's that you have this constantly increasing accuracy. And when something goes wrong, it's not this like fire alarm calamity of like, OK, the AI has no clue what it's doing. It's a reasonable problem. It's the same as a person making a mistake. You say, hey, man, you made a mistake. Like, here's what you did wrong. You give them feedback. They hopefully learn from it. And then they don't do it again.

[00:24:32] And you'll be able to achieve that same level of feedback loop with your AI systems. And I always try to give everyone listening a valuable takeaway. So if anyone listening today, before allowing their AI agent or agents to act on company data, what kind of tests should prove that it understands the business well enough to make a reliable decision? What kind of tests should they be running there? Yeah, so I think there's two classes of tests that we recommend our customers run and that typically they'll run before moving to production. Right.

[00:25:01] One is that you yourself as the business experts should curate a couple of tasks. Right. And they can be quite simple tasks, but they should be tasks that are really end to end. Things that go from raw information all the way to something that maybe an executive uses for decision making. We recommend like, you know, 10, 20, 30 of these tasks. And like you can really kind of just curate them based on what you worked on the last week. It's like no need to like go and like develop some like out of the world benchmark.

[00:25:28] But just having a couple of things that you know about being part of the benchmarking is critical. The second thing that we do for people and like people can do themselves if they want. It's more of like an automatic benchmark curation. And the way we do that is we look at the work you've already done and we assume that the work you've done is something that you care about, especially if you've done it multiple times. And we can automatically curate benchmarks out of that and say, OK, like here's a task you always seem to do. You always seem to go and like pull these numbers and use them to draw this conclusion. Let's make sure that the agents can do that as well.

[00:25:58] So it's a combination of like a small number of gold standard manually curated tasks that you should have at all times, just so that any agent that comes through, you can evaluate it. And then second is, I'd say, like a silver standard, much wider platform of questions that can be automatically curated from the traces of the past work you're already doing. Well, I think when we talk about AI in the workplace, there's so many big conversations at the moment.

[00:26:22] And I think your conversation today talking about why enterprise co-pilots fail when they don't understand how the business work is critical. It really is crucial. And for anyone listening that would like to carry on this conversation, we started today with you. Find out more about Sphinx AI. Where would you like me to point everyone? Yeah, feel free to shoot me an email at rohan at Sphinx.ai. We're always happy to compare notes with people, see what they're thinking about with their context. And if we can help, we'll help.

[00:26:50] Well, one of the many things I've loved about chatting with you today is how you've given us an insight into the metrics, the logic, the institutional context, governance, agentic AI models and everything. It's so exciting space at the moment. But I'll have links to everything that we've talked about there, along with anything else. If I can find some videos from you guys as well, I'll embed one of those in the show notes. But honestly, thank you for bringing all this to life today and doing it in a language that we can all understand. Thanks again. Absolutely. Thanks so much.

[00:27:18] I think my guest retail example today shows why fluent AI can still be unreliable. Yes, the system knew what revenue and stores meant. Yet it could not know whether the user wanted all stores or just the ones open for trading. Three sentences of company context were enough to clarify the difference and root the question correctly. And make reasoning visible to people.

[00:27:48] And the practical lesson here is to treat context as something that must be learned, reviewed, permissioned and updated as the business changes. So before production, my guest recommended today testing agents against a very small set of expert ridden tasks and a wider benchmark drawn from the work. The organization already performs. The organization already performs. But over to you.

[00:28:15] What company term would expose whether your AI genuinely understands how your business works? I'd love to hear more from each and every one of you about how you're using AI in the workplace. Please tech talks network dot com. You'll find links to everything we talked about today and much more, including 4000 interviews that we have now. And don't forget the event calendar where you can meet me in person on the road. But that's it for today.

[00:28:45] So thank you for listening as always. Speak to you again real soon. Bye for now.