By AI agents deployed in enterprise environments behave non-deterministically. Without explicit evaluation criteria and continuous guardrailing, they routinely produce outcomes that diverge from intent, whether through hallucination, misrouted actions, or compromised behavior that is indistinguishable from a security incident. Using a frontier large language model to judge agent behavior compounds the problem: it is prohibitively expensive at scale, introduces significant latency, and forces organizations to sample rather than evaluate every agent action. The architectural decisions made at the start of an agentic deployment define whether governance is even possible later.
In this interview on TFiR, Kamal Hathi, SVP and GM of the Splunk Business Unit at Cisco, breaks down how Cisco and Splunk are addressing enterprise AI governance through focused small language models, 100 percent agent action evaluation, network-enforced remediation, and a tokenomics framework that measures AI investment as return on value rather than raw token consumption.
Guest: Kamal Hathi, SVP and GM of the Splunk Business Unit at Cisco
Show: TFiR
Here is what every platform engineer, security architect, and AI operations team needs to know.
Technical Deep Dive
Q: What are the two fundamentally different categories of enterprise data and why does the distinction matter for AI?
Kamal Hathi, SVP and GM of the Splunk Business Unit at Cisco, draws a clear line between document-type data, including legal agreements, presentations, audio, video, and health records, and telemetry data, the continuous logs generated by applications, network infrastructure, and operational systems. Document data carries privacy risk and is governed by regulation. Telemetry data is operationally critical for security detection, performance monitoring, and infrastructure optimization, yet it is largely untouched by mainstream AI adoption. Both categories represent proprietary enterprise IP, and exposing either to external frontier models creates IP leakage risk and degrades model accuracy because frontier models have no training context for organization-specific telemetry.
“Most of these models, they don’t know how to deal with it. They have no idea.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: Why are enterprises moving away from sending proprietary data to external frontier models?
Hathi identifies IP exposure as the primary driver. When enterprise data, particularly meeting recordings or sensitive documents, is processed by a third-party model, that data can potentially be absorbed into the model’s future training runs, making competitive intelligence accessible to anyone querying that model. Beyond IP risk, frontier models lack the domain-specific context to produce accurate outputs on telemetry, making them both expensive and inaccurate for operational use cases. The enterprise-viable alternative is running fine-tuned smaller language models entirely within the organization’s own boundary, where data never leaves and the model can be tuned to the organization’s specific signals.
“The proprietary data that a company owns is very important. It’s their IP, their ownership. But if you want AI to be effective, you need to be able to train on that data.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: What is agent observability and why do AI agents require it before going to production?
Agent observability is the practice of evaluating and guardrailing every action an AI agent takes against a pre-defined set of criteria called evals, which are configured before the agent enters production. Hathi frames this as non-negotiable: AI agents are non-deterministic and cannot be assumed to behave as instructed. Without continuous evaluation against intended outcomes, agents can deviate in ways that are operationally damaging and, in a security context, indistinguishable from a compromised system. The evaluation harness must operate on every agent action, not a sample, and must be fast enough not to introduce prohibitive latency.
“You cannot assume that agents are going to do what you ask them to do. This is just a fallacy. They have to be evaluated, they have to be controlled, they have to be guardrailed.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: What is the Luna small language model and what role does it play in agent observability?
Luna is a small language model developed by Cisco and Splunk specifically for agent evaluation within their agent observability solution. Its sole function is to assess whether each instrumented agent action meets the evaluation criteria defined by the organization. Because Luna is purpose-built for this narrow task, it delivers 100 percent evaluation coverage at 97 percent lower cost compared to using a large language model as a judge. It runs on-premises inside the customer’s environment, processes the customer’s own telemetry data for fine-tuning, and operates at a speed that does not introduce workflow-breaking latency. Luna is the model component; agent observability is the broader harness that uses Luna to perform the evaluations, enforce guardrails, and trigger or block agent actions.
“It’s 97% cheaper than using a large language model, and it’s much, much faster. Because it’s bespoke, because it’s built to do this simple set of activities.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: Why does using a large language model as an agent judge fail at enterprise scale?
Hathi identifies three compounding failure modes. First, routing every agent action through an external large language model transfers sensitive operational data outside the enterprise boundary. Second, the cost is prohibitive: organizations with 20 agents could face evaluation costs of $20 million or more at scale. Third, the latency introduced by a frontier model makes real-time guardrailing impractical, forcing teams to sample agent actions rather than evaluate them comprehensively. Sampling creates blind spots where non-compliant or compromised behavior goes undetected. All three problems are structural, not configuration issues, and they do not improve as the number of agents increases.
“If you do the other approach, companies end up sampling everything. The latency is very high and as a result you don’t really get what you want.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: How does fine-tuning Luna with organization-specific telemetry data reduce hallucinations?
Because Luna is already a focused model with a narrow task scope, its baseline hallucination rate is substantially lower than a general-purpose frontier model evaluating the same criteria. When organizations fine-tune Luna with their own telemetry data, the model’s frame of reference narrows further to signals it has been specifically trained to recognize. Hathi describes the result as hallucinations becoming almost non-existent, because the model is not attempting to resolve evaluation criteria against a broad, ambiguous knowledge base but against a tightly scoped, domain-specific dataset that reflects the organization’s actual operational environment.
“The hallucination is almost non-existent versus something that’s so generic, where it’s trying to understand this in the context of everything else.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: What is the difference between Luna and the Agent Launchpad, and how do they relate to Agentix SoC?
Luna is the small language model that performs the actual evaluation logic within the agent observability solution. The Agent Launchpad is a separate capability focused on creating and deploying agents, including Cisco’s own internally built agents. Agentix SoC is Cisco’s security operations center agent, which handles detection and remediation tasks, and it runs inside the agent observability harness. Hathi describes a recursive architecture: Cisco uses its own Luna-powered agent observability to monitor and govern the Agentix SoC agent. This means the same governance framework available to customers is applied to Cisco’s own production agents.
“Our own agents, we observe with our own agent observability. It’s like this recursive thing.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: How should organizations measure the return on investment of AI agents rather than just token consumption?
Hathi rejects both token maximization and token minimization as useful metrics in isolation. The correct measurement combines token usage with project-specific outcome data to produce a return-on-investment figure that reflects whether the tokens spent actually advanced the intended business goal. Cisco has deployed this tokenomics framework internally across its employee base, giving every manager visibility into how tokens are being used relative to defined project outcomes. The key insight is that doubling token spend is acceptable, or even desirable, if the return on that spend is ten times the investment. Conversely, low token consumption is not a sign of efficiency if it corresponds to minimal or incorrect output.
“You cannot do a one size fits all. The answer is what did the token give you? And the more it gave you, the more it’s worth spending on.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: How does the Cisco network layer combine with Splunk observability to enforce security remediation?
Hathi explains that the network serves two distinct functions in Cisco’s security architecture: it is both a source of deep observability data and an enforcement point for remediation. Splunk processes correlated signals to detect anomalies, lateral movement, and unusual behavior patterns. That intelligence then flows to Cisco’s network layer, where network segmentation and switch-level controls can enforce policy and isolate threats without waiting for a patch cycle. This combination is what Cisco calls Hypershield with Live Protect, which allows protection shields to be deployed directly at the endpoint or network level based on real-time signals from Splunk, often before a vulnerability is formally patched.
“Splunk can tell you what is happening, get the correlative insights, find anomalies. Cisco can control the endpoint as well as enforce policy in a very effective way.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: Why is the boundary between security incidents and AI agent malfunctions blurring, and what does that mean for enterprise teams?
Hathi uses a concrete example: a customer service agent that returns a $100 refund instead of a $10 refund looks identical whether the cause is a hallucination or a deliberate exploit. With deterministic software, the distinction between a bug and a security breach was clear because intent was encoded in logic. With agentic AI, non-deterministic behavior means that unintended and malicious outcomes produce the same observable signal. This structural ambiguity is forcing CISOs and CIOs to collaborate earlier and more continuously, since the governance of agent behavior and the security monitoring of agent behavior are now the same operational problem.
“With the nature of agentic solutions, the line between intentional and unintentional starts to bleed.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: What architectural guardrails should organizations put in place when building or rearchitecting their AI strategy?
Hathi identifies three non-negotiable requirements. First, intent must be made explicit: agents do not infer from vague instructions, and the more precisely an organization defines what an agent should do, the more predictable its behavior will be. Second, every outcome must be measured against that intent continuously, not sampled and not reviewed retrospectively. Third, guardrails must be in place to block non-compliant actions before they complete. He frames the absence of these controls as deploying without a parachute: survivable only by accident. Cisco and Splunk’s agent observability covers the full stack from GPU utilization to vector databases to agent actions to token spend, providing the instrumentation layer that makes those guardrails enforceable.
“Intention doesn’t equal to outcome. You have to measure the outcomes very, very carefully.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Q: What is the time series model Cisco and Splunk offer and how does it fit into the small model architecture?
Hathi references a time series model delivered to customers alongside Luna as another example of the focused small model approach. This model is specifically designed to predict anomalies within a time series of operational events. Like Luna, it is delivered to customers who can then fine-tune it with their own telemetry data, making it highly accurate for their specific environment. It operates independently of any frontier model, runs within the customer’s boundary, and is purpose-built for one function. Hathi cites it as further evidence that narrow, tunable models consistently outperform general-purpose frontier models for enterprise operational use cases.
“It’s focused, it’s small, it does not depend upon some large frontier model. And so it becomes much, much more effective and something that can be used in enterprise context.” — Kamal Hathi, SVP and GM of the Splunk Business Unit, Cisco
Resources and Documentation
- Splunk, enterprise observability and security intelligence platform from Cisco
- Cisco, network infrastructure, security enforcement, and AI observability solutions including Hypershield and Agent Launchpad
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: When it comes to enterprise AI, it seems the easy solution is, you know, just plug your model into your data and go. And that’s where things go wrong. Because moving sensitive data to external AI models is one of the worst things you can do. It creates massive security and of course, runaway cost. But when we look at Splunk and Cisco, you folks are flipping the whole game. Instead of moving data to AI, bringing AI to where the data is. And we are here at TFiR with Kamal Hathi, SVP and GM of the Splunk business unit at Cisco to break down what it means and how businesses should look at data and AI. First of all, Kamal, it’s great to have you on the show.
Kamal Hathi: My pleasure. Thank you very much for having me.
Swapnil Bhartiya: Let’s look at this area where we tend to take data, where the AI is and which actually is. It’s totally fine with consumers, but that is not how enterprises should operate. So first of all, do you see in the enterprise space the same trend that enterprises are actually taking their data to AI or no, they don’t want to, but they have to. I want to understand the lay of the land.
Kamal Hathi: There are a couple of things going on there. One, I think data is a very broad topic. Most companies think about data, they deal with documents or things that are more, I would say, human language. So you’d look at legal agreements, or there are presentations or even videos or audio, health data. These are all the kinds of things that people deal with. Now that’s whole different ballgame. You should worry about privacy. There’s all kinds of laws against those kinds of things, which is very important. Then there’s another class of data which is telemetry. So when you are doing anything in the enterprise, any application, network infrastructure, your logs are being generated all the time. This is really important because when you think about security, for example, you’re trying to understand whether or not your environment is secure against external threats, someone is exploiting your network, someone is moving around, doing things, whatever. When you are worried about performance of an application, is it running fast enough? Is it slow? Is your infrastructure being utilized correctly? Are you using too much power? Whatever it is, these things are based on telemetry. And that is the kind of data that actually is not used as much for AI, it turns out. So there’s two interesting categories, but regardless of that, the proprietary data that a company owns is very important. For one aspect is that it’s their ip, is their ownership. But the second aspect is if you want AI to be effective, you need to be able to train on that data. That becomes important for both these kinds of data and we can come back to what that really means. So when it comes to really specific solutions that are built for an enterprise, I think enterprises are trying increasingly to not send that data to an external frontier model, but maybe to have a. Sometimes they end up doing fine tuning of various kinds of models, run them themselves. They do what’s called rag, which is the augmented generation retrieval that you can do around it. There are various techniques people use in order to avoid. I think what people are really scared about is for example, having your meetings recorded on people at these third party recorders and all that. And all that data goes to understand the speech. But then it also might get used to training new kinds of models that take your IP and now a competitor understands how to go, for example, build a new drug. This is what you don’t want to happen. So I think there’s a line between what you do with the data, how do you end up exporting it, what kinds of models are using and many, many I think we’re seeing examples of organizations using smaller language models that are pre built and focused for specific solutions and they all run in your data center, in your account, in your boundary and it’s completely tuned by you, completely controlled by you and it’s not really going anywhere. And that’s an example of something we’re seeing more and more very bespoke, very focused for certain kinds of solutions.
Swapnil Bhartiya: First of all, thank you for also differentiating with all data is same data. Of course when it comes to general purpose data that is used by, you know, to train because I mean if you look at LLM, we talked about the term AI but it’s basically LLMs, you know, that has vast amount of data. But when it comes to telemetry data that could be very, very specific to a specific industry specific company specific set of tools that you’re running. So first of all it’s not going to help that pool of knowledge base. And second is that data is also leaving the organizations.
Kamal Hathi: Third thing is that most of these models they don’t know how to deal with, they have no idea.
Swapnil Bhartiya: Exactly. So since they don’t have any data set, they cannot even give you right response to that as well. So talk a bit about what kind of challenge it creates for organizations because they do want to use AI, they do want to use agents to act upon it because that telemetry data can be too for some outage happen. I mean there are ton of things because that’s a. So talk a bit about what approach you know, either Cisco or Splunk is taking. You mentioned we get offset with the large language models. But the fact is in most cases you need a small. That’s right, very focused. Yeah. So let’s talk about how Splunk and Cisco are approaching this.
Kamal Hathi: Yeah, we’ll give you specific examples. So we talked a lot about this notion of agent observability. This is this idea that agents basically, you cannot assume that they’re going to do what you ask them to do. This is just a fallacy. They have to be evaluated, they have to be controlled, they have to be guardrailed. So for that you need to be able to observe and evaluate every agent action. Literally you go in and you set up what I call these evals up front before the agent goes into production. So you do this. Now the problem with that is that a what you described before, which is do you want every agent action going to some kind of large language. This is just. You are giving away everything. But the second and third important parts are that a using this LLM large language model as a judge on your agent is extremely expensive. If every single action of your agent gets now you’re paying like crazy. The second thing is, third thing is that it’s latency is a problem. These things are not fast to evaluate. So the approach we take is we have specific small language model, it’s called the lunar language model that we use for these agentic evals. And what we do is we do 100% evaluation. Every single agent action that is instrumented gets evaluated and it’s compared against what you had intended to do. And then based upon that, then you are either allowing the agent to take the action or you block it. Now this is very important because if you do the other approach, which is you take the large language model, companies end up sampling it, they sample everything. The latency is very high and as a result you don’t really get what you want. So our approach is to give you 100% sampling, 100% evaluation. It’s 97% cheaper than using a large language model. And, and it’s much, much faster. So because it’s bespoke, because it’s built to do this simple set of activities, and then we allow you to take your telemetry data and fine tune this model so that when we do, the evaluation is based upon your data and it runs all on your premises, so it’s all under your control and it’s cheap, it’s fast and it’s highly governed. So this is really important and that’s the approach we take. And I think trying to use an LLM for that, for example, has lots of reasons why it won’t work.
Swapnil Bhartiya: So when we look at Luna, so is it more or less like a harness to get started? Is this kind of. And you also agent Launchpad, also you folks launched it. So talk a bit about what exactly it is and at what point its role starts and when it ends or it is all across because you continue to fine tune model and it’s best spoke for your specific.
Kamal Hathi: So Luna is a small language model that’s part of our agent observability solution. And what that solution is, it’s all about setting up evaluations and guardrails for your agentic solution that you’re building internally or third parties that you have got. And so this is available now and its role is basically every time you should use it. When you start to build these autonomous or AI based solutions in your organization, that’s its role. It is not something that’s going to do everything, but it’s very focused on this one thing. Is it different than the agent Launchpad which you mentioned, where that’s more matter of setting up creating agents? For example, our own internal agents, we have this thing for security we call the Agentix SoC, which is a security operations center and that’s an agent that does look at security tasks for detection or remediation, et cetera. We also run those in the context of our agent observability. So our own agents, we observe with our own agent. It’s like this recursive thing. So the role of this, what I just described, the Luna model that it’s used in the context, agent observability, which is the harness that uses this model to do the actual evaluations, to do the guardrailing, to do the actual actions, et cetera. But it’s a model, the model that runs in the context of the harness, which is the agent observability solution. And that in turn is what is being used by customers and by ourselves to make sure that all the agents that they’ve got in the enterprise are being effectively evaluated and controlled.
Swapnil Bhartiya: And those agents can be totally using large Laguna models anything they want. So once you have this foundation in place with Luna, how do you help team actually scale without once again going into the whole sinkhole of cost and hallucination and performance?
Kamal Hathi: That’s the point. So the important part here is that this model is very much geared towards one thing only, which is to make sure that you are able to take a set of criteria you define what’s called an evaluation eval, and making sure that your agent is meeting those criteria. And that’s what it’s geared for. And because of that, its cost is much less. I said 97% cheaper than using a large language model. And so an example, I can’t give you the company’s name, et cetera, but they were spending literally order of tens of millions of dollars on large language model. And the same thing now with our model is costing maybe $100,000. I’m just giving orders of magnitude, but it certainly goes from if you’re a company who wants to have 20 agents and each agent evaluation guardrail with large language model cost $20 million, you want to spend billions of dollars on that. Whereas with our solution it becomes practical. It’s not cheap cheap, but it’s way cheaper than that. So I think it’s really important to understand that our goal and what we are making delivering with this is practicality. And also it’s about speed. If it takes seconds of time at every interaction to do this, then it’s useless. We make it very, very fast, very, very cheap. And again, it’s very focused. Hallucinations also become interesting here because now this model is very focused, but also you can fine tune it with your own data, your own telemetry. So now it’s extremely focused, it’s extremely tuned. So the hallucination is almost non existent versus something that’s so generic where it’s trying to understand this in the context of everything else. So that’s the power of using focus models and using and having them tuned with your data. And we have other examples as well. For example, we talked today, if you noticed or not, but about a time series model. That time series model is based upon again, we deliver to our customers, they can fine tune with their own data and then it tries to start predicting anomalies in a time series of events. But again, it’s focused, it’s small, it does not depend upon some large frontier model. And so it becomes much, much more effective much faster and something that can be used in enterprise context.
Swapnil Bhartiya: And if you look at Luna, who is its target audience? Who is its buyer? Is it security teams? Of course, data teams or. I think we’ve tried to break those silos back in those days, but now we have created new silos where teams don’t even talk to each other. It’s only when some incident happened, then everybody comes together. So talk about or maybe Luna can become a catalyst to bring two teams together.
Kamal Hathi: I think it’s not necessarily a matter of whether or not our model does it, but the fact remains agents are becoming like a new app when that happens. Let’s take an example. Let’s say you have a customer service agent whose job is to answer questions, do refunds, et cetera. And that agent has a hallucination, a problem, and it starts returning $10, $100, so 10 times what it should. Now imagine the same agent that has been hacked and it returns $100 instead of $10. What’s the difference? It’s the same thing, the same problem. Security, observability. In the past you would have had a hard line. So because this was a matter of writing code, observing, instrumenting it, this was all about having exploits and all that. With the nature of solutions, with agent and AI and agentic solutions, this kind of non deterministic behavior, the line between intention and unintentional, starts to bleed. And so this is what’s. So who is it for? Is an interesting question to answer. Now what we are seeing is that the CISO team and the CIO team are increasingly collaborating on these kinds of things. The CIO is geared to building agents that give value and solution. The CISO is worried about these agents starting to do malicious act. So both of them are increasingly using this as a way to combine and converge. And we are seeing this more and more. In fact, to your point about in the past, when there was a problem or incident, only when they start to have intersect, then they would start to talk, there’d be silos otherwise. So the new world, these solutions are increasingly converged already. And troubleshooting an agent or looking for security is very, very similar.
Swapnil Bhartiya: Now here I noticed that a lot of focus is also on tokenomics when we look at AI. And there was a time people used to brag that I burned 1 million tokens, do that, but what did you do with that 1 million tokens? And even Jensen will say, if you’re not using that many tokens, you’re not working. Now, when we look at tokenomics, the focus sometimes shifts on try to contain token costs. But what that also means is that okay, you have already burned those 50%, so you are left with only 25%, which is not productive. This is not what you want. So if we flip it and if you look at what does ROI on AI actually looks like and how is Splunk and Cisco looking at it to actually turning it from a cost center to actually a performance Profitable center.
Kamal Hathi: It’s a really interesting question because it is about roi. Very easy to go and just measure how many tokens you spend or are you spending too little, too less? The last thing you want to do is, as you said, you have a solution, massive return on investment. And you’re like, oh no, no, you’re spending too many tokens or you doubled your token spend. But how much was my return on that? If I doubled and I got 10 times the return, then maybe it’s worth it, right? So what we are doing is focusing this idea of tokenomics not just on measurement of use of tokens, but on measurement of return on investment. So we take the input of what the project is for, what the outcomes are and we combine that with the actual token use for coming up with metrics that tell us how effective was the use of the token. And as an example, we’re using this inside Cisco itself. Cisco is a very big company. You can imagine thousands of employees, tens of thousands of employees. And those employees, we all want them to have agents and people are building agents all the time. So we have built our tokenomics and we’re giving it to every employee. So starts looking at what are you doing, how are they being used? Every manager can look, et cetera. And we are dialing in the specific project outcomes into that. So we know the return on investment and that’s really important. Otherwise all you have is it’s important tool. You still need to measure the use of tokens, but you cannot do a one size fits all. You can’t just say it’s bad or it’s good. In the past it was oh, token maxing, more tokens used, it’s good. Now you can say token minimizing the more used, it’s bad. And the answer is neither. The answer is what did the token give you? And the more it gave you, the more it’s worth spending on.
Swapnil Bhartiya: I always say the Cisco sits the network. Nothing goes without that. So you sit in a position that everybody would want to be in. And now when you put splunk in, that mixes the whole observability story. So let’s look at also, how does this combination position you folks in helping some of these? Whether it’s token related problem, whether the whole harnessing problem, whether it’s also taming the agents problem. How do you look at this combine the story?
Kamal Hathi: So the network is an interesting thing. Network is a source of observability, of visibility and also is the instrument of enforcement. So when you have security exploits, this is Lateral movement, it’s across the network. Usually, you know, when something jumps from here to there, the network is where it’s happening. And so we are able to go and get data from the network at a very deep level, but also by making sure that we are taking advantage of the smarts in the network that we have to filter the data, make sure the right information comes through, and then we are able to go and use network segmentation, for example, to enforce any kind of remediation that we want to do. So this combination is really important. So the role that Cisco plays is very interesting and somewhat unique. We combine the visibility, the observability, the scale of Splunk with the ability to connect across the network to get the level of detail that is nearly impossible otherwise, along with enforcement points that are on the network itself. So I don’t know if you heard a little bit today we talked about this idea of hypershields. That’s something that Cisco has been delivering. And the idea for Live Protect, which means we’re able to go off and on the network itself using various switches, able to go off and come up with these protection shields that are available right on the endpoint, but they are informed by what you’re seeing in things like Splunk that are telling you, hey, this is the kind of odd public exploits. And we immediately, very rapidly can go and deliver those even before the network is patched. So this combination becomes very powerful. Splunk can tell you what is happening, get the correlative insights, find anomalies. Cisco can control the endpoint as well as enforce policy in a very effective way.
Swapnil Bhartiya: Excellent, thank you. Now, of course, almost every company is on the AI journey, but a lot of them have started rethinking how they’re using it. So if somebody asks you, are there the specialty car rails or architectural decision that they should make if either they are embarking on their AI journey or they are rethinking their rearchitecting their AI journey, what would that be?
Kamal Hathi: I think the one thing that everybody’s learning very fast is that intention doesn’t equal to outcome. So when people go build AI solutions, they have certain intentions and they think like, oh, I have an agent, just a human being almost. I can tell them what I want to do and they will just do it. The fact of the matter is that this is not the case. The outcomes of agentic solutions actually are extremely difficult to predict unless you’re very, very explicit about what you want. But then second thing, you have to measure everything. You have to measure the outcomes very, very carefully. And so the thing to make sure is when you go build the solutions, A, you have very clear, you know, instructions, what you want to do, and B, you have got very clear measurements and evaluations of what that outcome is actually being met or not. And then you have guardrails that enforce those. Without that, it’s like jumping off a plane with a parachute. Right. It might turn out that, you know, I don’t know, you get lucky and you want to float and land in some nice snow bank and survive. Unlikely. But with the parachute, you know, you’re confident it’s going to work out. And so you need the parachute. And we provide this parachute with our observability for AI all the way from the gpu, its usage, to vector databases, to agents, to tokens, all of it. And that’s really important. Without that, it’s, you know, hoping for the best.
Swapnil Bhartiya: Kamal, once again, thank you so much for joining me. Not only share the insights on how to bring AI to the data and not to other ways and also how to also focus on smaller models, not just go all the big models. Bigger is not always better. And thanks for also sharing insight how organizations should look at their whole AI journey.
Kamal Hathi: Absolutely.
Swapnil Bhartiya: Thank you so much for your time and I would love to have you back on the show.
Kamal Hathi: Thank you so much. Appreciate it.
Swapnil Bhartiya: Thank you.
Kamal Hathi: Bye bye.
