Kaushal Kurapati
GVP of Products,
AI Agent Platform
Oracle

TRANSCRIPT:

 

Vigoroso

Hi, everybody. Welcome to another special edition of the Enterprise Edge Podcast. This is Mark Vigoroso, founder and CEO of the Enterprise Edge. We are coming to you live from Oracle AI World 2026 in Las Vegas. Thank you all for joining on this special occasion. We are winding down. We're going to finish with a flash. We have a special guest today, Kashal Kurohati. He is Oracle's G VP of products for the AI Agent platform. Lots to talk about.

​

I think we've heard everything that's going to be announced has been announced with lots of context. to have this conversation about claw and about a genetic in general. And I've had a chance to meet with some customers and so lots to kind of bring this to a head. But yes, first off, welcome.

 

Kurapati​

Thanks for joining us. Thanks, Mark. Thanks for the opportunity.

 

Vigoroso

No worries. so let's let's dive in. I mean we've we've covered we've talked a lot this week about about claw and fusion claw and

some of your colleagues, Randy and Chris and others that have really I think done a great job of of explaining what's different about it. Why is it, what is it? Right. and I think one of the things maybe a place to start is there's a really interesting division between deterministic and probabilistic work workflows and the fact that I mean it sounds sort of scientific, but really what it is, it's it's it's

optimizing what you need when and where sort of fit for purpose for the type of work. So and I'm butchering it. But what I'd like to do is hear from you a little bit about what is the thought process behind that split between probabilistic and deterministic and what is each and how does it work in the context of fusion claw?

 

Kurapati

Yeah no thanks. I think so the way we think about it with fusion claw is if We are first of all thinking about end-to-end process that Claw is intended to solve, moving towards a business outcome, right? And so the business outcome is the is the important thing to keep in mind. The human being has perhaps specified it. They are the one who's prescribed the objective and the outcome they desire. And they're accountable and responsible for it. The system is then executing the work, right? So once you specify an outcome, this is what I want.

These are the constraints, these are the boundary conditions. That's what you would give it as a as a instruction. And then what we do is we use the power of a large language model which has reasoning capabilities, which has ability to figure out how to plan a approach to solve that problem, to get to that outcome. And we use the power of large language models and its reasoning capabilities for that. Then what we do is once that is done.

​

We know that the plan is in place. The rest of it is deterministic execution because that's something that where you don't really need to go to an LLM for every single processing. If you're processing a massive amount of data, you're processing a million rows. For every one of those rows, you don't need to go to an LLM and ask it to make some decision. There's absolutely no need. We know how to run the deterministic process, we know how to execute on that workflow, what happens next. The plan has already been given, right? So the plan is something where you leverage the power of HelloM and do it. But then you also need a bit of orchestration. If you do execute and you don't come up with the you don't reach the business outcome or whatever you're looking for, then of course there's a loop that kind of kicks in and says, okay, let me try another approach, right? Let me revise the plan.

​

That's when maybe again a reasoning po reasoning power is needed. You reformulate, recast the problem, replan, and then you go and execute. So that's the way to think about that that bifurcation.

 

Vigoroso

Yeah, to me it's it's thinking and doing.

 

Kurapati

Thinking and doing, yeah, exactly. That's right.

​

Vigoroso

I worked at a company once where the mantra was plan the work and work the plan. That's a good one. I'm putting it in terms I understand it. And I guess the question I have is when you're in the midst of execution and an exception occurs or an assumption was wrong. Or you know, I think you touched on it briefly, but like in the in the world, you know, five years ago, there there's there's sort of an inefficient process in my mind of identifying that exception, mitigating it, you know, and then getting back to work with some sort of a plan B or something. How how does that work in the context of Claw?

​

Kurapati

Yeah, in this case what would probably happen is then it would c probably reach some kind of a dead end and say, look, I did not meet the objective. It's got that much of the sort of awareness, if you will. what is the outcome that is specified? I did not reach that.

And that becomes then the loop will kind of get cut short and say, okay, let's go back and replan because this has not worked for whatever reason. So there's a bit of that memory that's also there and the learning that happens and says, Okay, we've got a replan because I couldn't solve this. Or you might go back for clarifying questions, the agent might ask the orchestrator to the user who's in the loop saying, Okay, there is this happened. I mean, how do I proceed? and if it's getting stuck repeatedly.

 

Vigoroso

So a lot of the concern if I would argue maybe even a headwind to scaled enterprise AI has always been this sort of opaqueness about how much is it going to cost me at the end of the day, right? Because of like this black box of token consumption. But now it seems like you're sort of solves that with sort of optimizing when and why you use the LLMs. And only then.

 

Kurapati

Only then. Yeah. And you don't need it all the time. Yeah. You don't need a hammer all the time to kind of hit it. So and we're giving a lot of tools as well along the way for you to monitor, estimate and all of that.

 

Vigoroso

Yeah, I think it's it's and the governance piece is is is very important. And and the transparency of transparency, all of that is very important.

 

Kurapati

That is spot on. I think one thing that we've really I think re I really like is the outcome receipt that we've sh shown in the fusion claw execution and it shows you exactly not that we were not doing auditing last week last year, we were doing it. But this is in a much more human digestible form that you can really take to an audit committee of some sort internally or if you were investigating it, look, this is what has happened exactly.

​

And it's time date stamped with every single thing, including the model that it's used and sources of data that it it really referred so that you can be sure that the grounding documents that were used, the policies and procedural documents that were used were were the right ones, and what it accomplished in terms of what plan it changed or or what sorry, system of record, what it changed in the system of record if it did or what it's proposing at least to to to do that. So I think that kind of an auditability and transparency gives you confidence to put something like that in production.

 

Vigoroso

No doubt. I think there's I would argue that's another obstacle or or headwind that I've read a lot and seen a lot and worked a lot around. I would argue more in the mid-market companies that don't necessarily think automatically about audit and compliance, right? Yeah. And having to to represent that right with transparency. So I think it's great. I now the question I have - I think the number is twenty five claw powered apps. Yes. And I'm thinking about interoperability. Like you think about those I don't have a list of all twenty five, but that the workforce staffing, shipping consolidation, correct account territory growth planning. So is this a two specific function, yeah, that's right. So is there what mind is paid towards Interoperability. So is there no sort of this web or network of agent claw-enabled agents that are collaborating? Is is that too much of a leap?

 

Kurapati

Not necessarily. I think it depends on what all you need for it to solve a particular business problem. Right. And we also know that it could go cross pillar sometimes. And and by design, what we've done is we have removed even from our go-to-market, like any pillar specific sort of you know constraints so that you what you can do is look I need this particular agent to go and check something in HCM and that but you're operating in maybe sales or support right then it could easily cross over and sort of get that data assuming the role-based access has been authority has been given, approvals are in place, it can go and get that. And so you can really think of that as agents talking to other agents. So I mean obviously shared context. That's what happens in an agent to cap and collaborating to get the answer. Earlier if you are a like human being, you're doing this, you have to log into a different system completely and you may or may not have access. Maybe you did, even if you did, you are the one going from screen to screen.

​

And piecing together that answer. Now with agents, I think that becomes much easier. So you really have can think you're free to think about sort of, this is the outcome I want to achieve. What agents do I need? And what permissions do they need to have to which data to essentially accomplish that get to that business outcome?

 

Vigoroso

Right. Got it, got it. I'm intrigued a little bit. The opportunity for your customers to I guess travel this sort of path on what I would call autonomy, right? So the dialogue. The dial, yes, yeah. Like are there like ceilings by function that let's just say sales. Like there's there's maybe there's a ceiling to the level of autonomy that certain workflows will ever have. But the reality might be and you can check me on this, is that when when when when clients first start out, maybe they have that dial turned all the way towards like heavy human interactions. Heavy human interaction. But then as as they test and and validate maybe they go towards that ceiling. Yeah. And then they hit the ceiling. Yeah. And then maybe there's a higher ceiling over in supply chain. There's a lower ceiling over in accounts payable or whatever. is the is that a fair characterization or how do you think how do you envision that journey for the customer?

 

Kurapati

Exactly. I think the dial that you mentioned and and I think that's what we see in a lot of customers too. Like depending on the use case where you if you are if you're processing invoices and you're cutting a check to somebody. if it's a ten dollar check, maybe it's okay to let let that autonomy be okay to the extreme. How higher the stakes. Right. How high are the stakes? What is the blast radius if it does something wrong, right? but if you're cutting a hundred thousand dollar check, maybe somebody high up should review it. It might be a prudent thing to do. Not that it's wrong, but it's just that because there is accountability here.

​

Ultimately with the human, whoever is this in that department, whoever is kind of signing off on on on those ultimately. So I think that's exactly right. And also it's it's the level of degree of confidence people get with the agents over a period of time. It's just like if you're in a car you have self driving capabilities, you would some people I know even though it ha does a pretty good job, they don't let it go. and but stakes are pretty high. Right, stakes are pretty high.

​

I've learned to trust my car with that capability. So I I hit it and then it does a pretty decent job. Experiential, not theoretical. Not ex yeah I I experience I experienced the the goodness of it. But at the same time I still have the hands on the wheel in case I need to take over and which is the accountability I need to have for my own self.

 

Vigoroso

Can you share your thought process around the commercials for Fusion Claw.  I don't think you guys have actually published or announced any kind of pricing art. So I I think that's true. But what's the thought process that if you maybe could give some insight to how are you going to commercialize this to existing customers, new customers? Is there anything you can share on that?

 

Kurapati

Yeah we can. I mean I think we've we've already made some changes to our pricing list. So it's a continuation of our consumption based approach. So it's consistent with that.

​

And we're trying to keep it as simple as possible. Since we use premier kind of frontier models in this case, it's essentially the actions and the outcomes. The outcomes matter. So it's it's essentially if you're accomplishing a particular outcome, what did we we call them AI units and and how much AI units has it consumed? And there's a particular kind of rate card multiplier for for for for the fusion claw related tasks. So we have added that.

​

And essentially it's consistent with the earlier kind of model, so that you're familiar with it. And because there are hard costs, obviously, because there's we're using frontier models which tend to be definitely more expensive. And we're spinning up a lot of compute resources when you launch a claw. Each each one j each one of those jobs spins up a container on our side, which is a secure container.

 

So to make sure that the value is there. it's it's a continuation of the AI units approach. And there's a business value multiplier for a claw-specific business outcome. Again, you don't need claw for every single task out there. Yeah. you would use it when you're solving the more complex high-value problems that exist in an enterprise workflow.

 

Vigoroso

Yeah, I mean that's the other thing we haven't mentioned, it's probably the most foundational long running multi step multi complex exactly. You know, we're not talking about task completion.

 

Kurapati

If I'm asking a simple question saying, what is my leave balance or for the year? You don't need a claw task, it's a simple thing. Just a simple conversational thing. It's microagent can execute, get me the answer.

 

Vigoroso

Yep. And that's what I think  people are starting to realize is that AI has really become sort of this fabric or this portfolio of use cases with I mean fit for purpose answers to some of these use cases. Some might be Gen AI, conversational AI and Gen AI was in without claws. I mean it's yep, you know, and it proliferates and I think with good reason. With good reason. Yeah. Each is a is each has got its own purpose and place.

 

Kurapati

Right. So when embedded AI was I mean like Gen AI tasks were introduced, that was that had a particular you know reason for it. And as things have evolved and capabilities have improved, model capabilities are very dramatically changed and which is enabling us and our harness capabilities have changed significantly and our ability to sort of execute deterministically and things like policy models that we've launched, which allow you to follow strict rules and execute. I think those have made a significant difference in introducing these new agentic patterns.

​

And so you it's choosing the right course for the the you know particular course. Yeah essentially. No, it's it's in its in rapidly evolving. Rapidly evolving, right? Yeah, yes.

 

Vigoroso

Well let's do this. Like I mean we're sort of at dinner time here, but I'd like to give you a opportunity. Like you've been here we've been here two three days. tell us a little bit about what are some of your top takeaways, your observations, maybe something you didn't expect or surprises from the last couple days.

 

Kurapati

First of all, the excitement is palpable. Right. I think I've come across many customers saying this, many partners and even media folks saying this, saying this is it feels so much it's real, not just AI is really doing the work in the enterprise. And they're actually seeing people putting this in action. Yeah. So and I've seen it, we've seen it, but it's when you hear that feedback. That's very rewarding, satisfying because people are actually getting some some things done with AI. It's not a theoretical construct any longer. So that's first of all, I think the most satisfying piece.

 

And the second is the feedback on the ideas resonating. They're immediately looking at something like Fusion Cloud or a companion, or even Fisker that we launched. Like, I can make use of that. That that absolutely makes sense. You don't need to have elaborate gymnastics to explain the concept. Once you're like a few seconds into it, if people are nodding and they're agreeing with you and they're getting it, then you know that you've done a the you're on the right track. Yeah. And and our pace of innovation is something that has really not only surprised but delighted me and also everybody. Yes, it is frenetic pace to keep up, but at the same time, I think.

Are appreciative because we are really solving the problems that matter to them and they know that they can get ROI from AI you know when you kind of adopt all of the things that we're building.

 

Vigoroso

I mean I've talked to a few customers and I will say that most of them said unprompted that they can't keep up. They can't keep up, right?  The long pole in the tent is no longer the software vendor. It's it's now the customer's ability to absorb it. It's on them in a in a way. So I thought was kind of a simple but yeah important takeaway.

 

So well Kaushal, thank you so much. Thank you everybody for listening and for streaming. we are towards the end of our AI World 2026. We will catch you on the next episode of Enterprise Edge. Take care everybody.