How Enterprises Should Start, Scale, and Secure Agentic AI

Travis Naurug: The big shift here is that we're moving away from this probabilistic text generation, which was kind of the core of generative AI, to more logic based reasoning. And that's a huge change for us where agents aren't just predicting the next most probable word in a sentence, they're actually breaking down a goal into various steps, they're critiquing those plans, they're using different tools to interact with other systems and they're actually achieving outcomes autonomously.

Host: Morning, good afternoon and good evening from wherever you're joining in. Welcome to our podcast. I'm joined over here with, uh, Travis Naurug and he's a partner solutions architect for generative AI at Google. He spent several years modernizing the infrastructure for clients and customers and also at Google to accelerate innovation through cloud and AI technologies. He's been at the forefront of helping customers and translate the promise of AI and real world business outcomes. Welcome to the show, Travis. Very happy to have you over here.

Travis Naurug: Thanks so much for having me. Really excited for the conversation today.

Host: So Travis, before we really dive into the tech sides of things and you know, the generative AI, the buzzword as to what is what, everybody is talking across the industry and also with customers, I wanted to take a, uh, bit of a dipstick or a check trying to understand what led you into technology. What's your story? Why do we see you working so closely with clients and customers? How did you walk your path in your life? What's your story all behind becoming a technical agenc expert at Google?

Travis Naurug: It's a great question. It's been a while, but you know, I think what originally drew me into cloud and you know, I've been doing this for about 10 years or so, was just kind of the fundamental shift in how enterprises were managing their IT infrastructure. And you know, I do kind of consider myself an infrastructure guy. That's kind of where I started, you know, and just looking at this fundamental change of how organizations deploy infrastructure to me was just fascinating and moving from these years long projects of getting a new data center up and running and you know, actually racking all of the servers and running cable and all of that stuff to clicking a button was just really fascinating to me. I think there was one moment that it's going to sound kind of silly, but it was a demo that I used to do a long time ago, like early days of cloud where I would just spin up a vm, um, you know, and I sit in Austin, Texas. I'd be on a call with a customer, I'D spin up a vm, um, in Australia. And in real time it's up and running in a minute. You can go in and ping that server immediately. And I can't tell you how many customers that I interacted with who would see that. And again, this was early days, so it was a little bit more novel back then, but their minds would be blown. They're like, hey, if, if we were to get a server up and running in Australia, like that is a huge undertaking for us today. And you're telling me I can do it with a click of a button in the browser like the future is here. And so that, that was uh, one of the moments that, that has always stuck with me of, you know, how something, you know, relatively simple can have such a, a big impact. And it's really just a different way of kind of thinking about how you're managing infrastructure. And that, that really stuck with me. Now when I look at the AI side of things, uh, I've been in the generative AI space now for about three years. I think the thing that really stuck with me, especially coming from more of an infrastructure DevOps kind of background, was looking at Gen AI as something that actually impacted me personally. It was one of those things that when Gen AI started becoming kind of a normal term that everybody knows, everybody is talking about in mainstream news, really thinking about how I can start using it as a tool in my day to day was the thing that really hooked me. Um, and ever since then it's like, I want to do more, I want to learn more. And so it's been a wild ride over the last three years. You know, things are changing so quickly, but that makes it really exciting right now.

Host: I mean, that's very interesting. You know what, and I think I share those sentiments that you just mentioned about spinning up a VM in Australia while, you know, sitting in Texas. I had a similar experience way back in the early or probably mid of the 2000s, around 2005, 2007, when I could actually SSH into a virtual machine on cloud. And virtualization at that point of time was a technology which was well tackled even at a processor level and an operating system level, but then taking it to the next level and making it accessible to people across the board at the click of a button and all of these aspects of generating keygens using SSH and using those keygens and putty and so on and so forth to create ovm, those are the initial days. And um, I'm um, completely aware of the fact that we have moved far across from VMS to now being serverless ecosystems completely. And I think on the generative AI side, I think I echo your sentiments. That we as people have started consuming the technology now is tremendous and I don't think it has only happened in the past two years. Yes it has leap forward in the adoption part of the past two to three years, which is true, but I think the technology has been in the oven for quite an amount of time. It has taken it a while to reach the space. And what's your thought process on that? How these, the generative AI aspect is moving into autonomous decision layers? Because what I hear in the enterprise is that there's a lot of focus on these agents to now become completely autonomous and not be more, let's say just chat oriented. That means we are expecting these agents to work in a timely manner at a particular time, monitoring different systems or data and so on and so forth. And then if not really take the far end decisions, maintaining human in the loop but then giving pre populated elaborated insights to the users so that they can take better decisions, more informed decisions. What's your thought process on that?

Travis Naurug: Yeah, I mean I think that's the most exciting kind of shift in how organizations are really approaching AI. Today is really moving from, I kind of like to think about it as like historically gen AI even and just AI in general has been more of like a read only type of solution where you're getting AI to summarize data for you. You know, maybe it's generating content, writing an email, something along those lines. But I think we're at this inflection point right now in 2025 where we're really entering that agentic era and that's moving from kind of that read only pattern to more of like a read write type of pattern where agents aren't just drafting that email copy for you or doing a search across your data stores. It's taking that a step further and maybe you know, looking up customer information in a CRM, it's checking the inventory, it's updating a very, you know, a record in that CRM potentially and then it's actually sending the email on your behalf. And that's where I think we're going to see, you know, and we're already seeing it today. But just these tremendous productivity gains across the organization. And so that's the thing that makes me really excited about the future. And again like I mentioned earlier, the rate at which things are developing, uh, you know, is unprecedented. So it's always funny, I feel like spend a day reading and getting caught up to speed on new technologies. I go to bed and then wake up the next morning and, you know, 10 new papers have been published around some awesome new models or new patterns or architectures or things like that. So it's really an exciting time. And I think the big shift here is that we're moving away from kind of this probabilistic text generation, which was kind of the core of generative AI, to more logic based reasoning. And that's a huge change for us, where agents aren't just predicting the next most probable word in a sentence, they're actually breaking down a goal into various steps, they're critiquing those plans, they're using different tools to interact with other systems and they're actually achieving outcomes autonomously. And I think that's so exciting. And I think we're just starting to get into that phase. So I'm really, really excited to see what the next couple of years have in store for us.

Host: I noted down a couple of things. These are very, very remarkable as to how you think about this. The analogy that you built from being read only to a, uh, read write is so true because if you look at the history of computing, we took time to go from read only systems to read write systems. The ability to write things, which was essentially done using code and program at that point of time. But now the agents are also going through that same paradigm shift from being just read only output than to read and write. And what you also mentioned is, which is very true because we as humans started consuming this technology more from a text perspective, trying to read what the LLMs and so on and so forth are saying. But now we've come to a place where we have more structured, more programmatic output being generated by these agents, which can be articulated as goals and steps and planning and so on and so forth. I want to touch base on one point that you mentioned at the end, which was about the interaction of these agents, or, uh, generative AI as a total, with other systems. Now we know that API has become the real life ecosystem of systems interacting with one another. And we've got a wonderful security layer encompassing these APIs using OAuth. But I want to touch base upon how are we looking at agent to agent communication? How is that place evolving? Because right now we don't really have a methodical way of saying how two agents can talk. I know there are some work that is going on, but I would like to know your thought process and how do we make sure that we bake in authorization, also an aspect of that ecosystem for agent to agent communication.

Travis Naurug: It's such an interesting problem to kind of solve for and I think when we look at running agents kind of standalone, it's relatively easy. Obviously there's some challenges and stuff there, you know, running these agents. But I think what we need to plan for and really get ahead of this is when we live in a world where your organization doesn't have tens, 50, 100 agents, but thousands of agents across the enterprise, that becomes a big scale issue to solve for. And you know, you think about, you know, maybe there's a sales team over here who has built an agent using Google's 80K for example and then maybe there's an operations team who has built an agent that does something in their using LangChain as the framework. And you know, maybe one's hosted on Google cloud, another agent is hosted on a different cloud. I think that's, that's a very reasonable scenario and I think that's where a lot of organizations are going. So we really have to solve for kind of that standardized communication protocol amongst agents built with different frameworks, built in different languages, hosted on different platforms. And that's where I'm really excited about our new A AH2A protocol, you know, open source agent to agent communication protocol that is enabling this really rich communication. You can kind of think of it if you're not familiar with it, kind of that almost HTTP layer for agent communication where it doesn't matter how you built the agent or where it's hosted, but you're enabling that communication. And I think that's going to be really critical, you know, and I think it kind of comes back to a fundamental problem with agents where the agents are really only as powerful as the data and the tools that they have access to. And we all know the enterprise landscape, there's a lot of tools, there's a lot of different platforms out there and we need that ability to communicate. And so I think we're really kind of taking it from. Even in a multi agent, like a standalone multi agent, I still kind of look at that as almost a monolith move from that to modular teams, you know, with multiple agents built with again different frameworks by different teams. But having them work together I think is where the power of agents is really, really starting to shift. And it's enabling some really fantastic use cases again on that process automation side, the productivity side.

Host: Right, I completely agree. I mean have been privy to the initial documentations of A2A that came in from an A2A protocol perspective. What are the guidelines that has been built? But I think there's still a lot of work that is being done on the authorization part of it. So that, which agent, what kind of identity do agents take up based on whether a user is using that agent or it is an autonomous agent and how does that identity flow in the complete process is something that is becoming standardized with these kinds of protocols coming into picture.

Travis Naurug: Yeah, and I think that it's a great point to bring up. Security, I think is paramount in any sort of generative AI application that you're building. And one step further than security is really the governance around those agents. You know, how are we building the guardrails, how are we leveraging responsible AI and how are we controlling access to different data? And so when I look at, you know, kind of that authorization side of things and just security governance interoperability doesn't necessarily mean chaos. And I think that's a lot what, what a lot of people think. And in a standardized ecosystem that things like eight way start enabling, we can really start applying uniform governance policies across all of these various agents that are, you know, running in production within your organization. So we can do things like, you know, ensuring maybe there's a third party travel agent that your team uses for, you know, their, their travel booking, that third party travel agent can talk to your calendar agent for example. But we can start building in these policies and access control lists authorization that ensures that those agents don't have access to things like the financial reporting agents. So I think building out these guardrails and really, you know, scoping agent identity, user identity, ensuring that the right ACLs are in place is really critical here. And so, you know, any conversation that I'm having with the customer, I think security governance compliance needs to be part of the conversation from day one. It can't be an afterthought. I think that's absolutely critical today.

Host: I think I completely agree with you, but my mind is actually going in a direction moving from the technical sides of things. Travis and I know that you engage with clients and customers at such a deep level. What is the sentiment that you see in the industry on the ROI of these investments that the industry is doing in generative AI or probably if you want to deep dive. And as one of uh, the ROI AI report from Google Cloud also says 74% of the organizations achieve positive ROI within the first year of deploying Agentic AI, which is remarkable for a technology so new. 74% of ROI. It's really great. But at the same time we also have other reports coming from various different trusted sources where which are questioning the ROI for, uh, investment in generative AI and agent care. How do you create balance with your clients and customers when these kinds of different messages, some are more perception based, but some are more reality based. What's the sentiment across the industry at the moment of time?

Travis Naurug: Yeah, you know, I think that that stat that you shared with 74% of customers are seeing that ROI. It's interesting because, you know, when I look at those customers and the customers that I work with, one of the overwhelming themes that I see is the customers who are successful aren't necessarily looking for these like moonshot use cases. You know, some super complex, complicated use case. They're really just looking to reduce friction within their teams. And so those successful customers that I see, the best place to start really is where they're seeing high volume of some sort of task and there's high latency there. It takes a long time to do these tasks. There's a lot of context switching. And so, you know, I think at the highest of levels, the customers that I see having these really successful use cases are focusing on use cases like customer experience. They're focusing on use cases like employee enablement. I think one of the underlying reasons that we're seeing such adoption in these areas is that they tend to have the cleanest data and the clearest KPIs. So if you're thinking about like customer experience, like what's the time to resolution for a support case that's coming in, what are processes that you can automate on behalf of, uh, an employee who is, you know, maybe having to switch context into three or four different systems to get an answer to something. Those are the types of use cases that I think are providing just tremendous roi. And like I said, they don't necessarily need to be the most complex use. Really identifying where is that high volume friction that you can kind of solve for your employees or for that particular use case. You know, one of the things that I've seen too is kind of double clicking on the customer experience. Use cases is kind of evolving from like FAQ chatbots. This is something we've had for 10, 15 years now. It's, you know, pretty standardized, but the evolution of FAQ chatbots to more of a transactional agent. So, you know, the way that I look at that is really kind of the differentiation between action versus information. So those traditional chatbots are just reading policies, they're giving that answer back to the customer. When we look at agentic AI, we can actually execute on the policy, processing a refund or updating an address for a customer profile, or rescheduling a flight or anything like that. So there's a lot of, uh, immediate ROI there for customers. And even if you think about that use case and if an agent can resolve 30% of your support tickets, you know, without a human being involved in that process, that cost per ticket drops dramatically and CSAT scores are going to be rising because they're seeing such a quick return or resolution to, uh, you know, to their support ticket. So it's use cases like that where I'm just seeing, you know, so much roi. And again, they're not the most complicated moonshot types of use cases. It's pretty standardized stuff, but there's a huge ROI there.

Host: Then I had the opportunity of being at uh, the Google Agent AI roadshow and conference that was held in Atlanta. And Google did that across the board in different states. And we could actually see those use cases where they had this hypothetical company that was bringing out agricultural stuff for people to do gardening and all. And it was deeply integrated into Salesforce. Somebody asked for a discount and there was a ticket to be associated. The agent got into the system and then there was a human agent also interacting at the same time. And the speed at which information was flowing for these customer support agent and also for the customer. And the way we were able to resolve those tickets, as you rightly said, I could definitely see a customer wanting to go back to that company time and again. And I completely agree with you on that particular part on customer experience management. There's a huge ROI that we definitely have. Let me take a little bit of turn and go more into the Google ecosystem as to what we have. Google has a great agent marketplace and for, you know, getting the right agents and so on and so forth. How does this Marketplace helping organizations to overcome technical and commercial barriers? What is that I have from Google Market Agent Marketplace that I can quickly pick up for my organization and start working on top of it.

Travis Naurug: I'm glad you brought up Marketplace. I think it definitely simplifies a lot of the adoption of agents. And I think, you know, when we look at Marketplace, a lot of this comes down to kind of de risking the total adoption curve of starting to implement agents within the organization. And you know, a lot of times, like the hardest part of enterprise software isn't usually installing it, it's really buying it and trusting it and going through that contractual process. And I do see the marketplace solving a lot of these issues by turning agent adoption from more of a development project into a ah, procurement decision. You know, and when I look at, kind of going back to your previous question around some of the use cases we're seeing, one of the areas that I'm seeing across industry, across customers is a lot of organizations, very similar problems. So you think about, you know, a sales team for example, who needs to do reporting on their forecast and updating Salesforce for example and some of these kind of standard tasks that all salespeople have to deal with. You know, they could be in manufacturing, they could be in consumer packaged goods, they could be in financial services. These are all pretty standard problems. Where I see marketplace playing a huge role here is for a lot of these very consistent challenges that we're seeing across industry, across organizations, it makes it really easy to go and find you know, a sales forecasting agent or a financial reporting agent essentially these, you know, more basic use cases that have already been solved for. It can be a little nerve wracking for a customer starting to, you know, just start dipping their toes into AI because they're like, hey, we need to put a development team together, we need to allocate engineering. It's going to take three months for us to build something out and it's a lot of cycles, a lot of money that organizations are spending. But you know, when we look at the marketplace and especially with partners like yourself who are building these great agents that lots of companies can take advantage of, it becomes a very easy procurement decision to build something that's battle tested, it's compliant, it's got the governance in place, it's click to deploy. So it really allows customers to start benefiting from the advantages of agentic AI without the huge upfront investment that you typically would see through. Kind of a traditional development cycle, right?

Host: With all um, collaboration with Google and Freedoms being over there. The kind of agents and the ecosystem, the agent ecosystem that we have built and we've been able to bring it forward on Google Cloud for clients and customers to pick it up from. Marketplace has been a, ah, tremendous effort from our side and we are very happy to have our uh, accelerator ecosystem available on Google Cloud. That's a work that. But I guess both of our teams have done a huge, huge job collaborating and bringing that to life for our clients and customers. I'd like to come towards a more conclusion aspect, more business aspect of our discussions. We've talked about the technology sides of things. We talked about the ROI and how customers can take what we have as an offering on Google Cloud, and we all know about the hype curve and the reality aspect as to how industries move in a technology space. When a new technology comes in, what's your recommendation to an enterprise who's probably, if you bracket them in three groups, there's an enterprise that's taking first step towards generative AI. What would be your recommendation? There's an enterprise who has done certain experimentations but is now looking at productionization, now trying to make it more real so that it could scale it across their enterprise. That's the second bucket and then you've got a third bucket who have been walking this path for a while, they have matured their ecosystem quite well, but now they're looking in the long term as technology changes, that is a third bucket. How would you recommend the next steps and the engagement and the processes that these three kinds of group of enterprises should take, let's say in the next coming one to two years. And we know that technology is changing

Travis Naurug: very fast to your point. Technology is changing fast and there's always going to be a new model. It's evolving very, very quickly. I actually think, you know, when we look at organizations who are starting to, you know, kind of go down this agentic AI path and looking out into the future, where do we really need to spend our time and get ready for this future of agentic AI? It actually kind of comes down to a fundamental problem of data modernization. You know, and kind of going back to the comment I had around moving from this paradigm of like read only to read write, seeing that on the data side as well, just kind of the foundational enterprise data that we're using in a lot of these applications. So, you know, I look at it as, you know, we're seeing this shift in data preparation. So if we think about kind of early days of generative AI, we really focused on a lot of unstructured data, so PDFs, documents, that sort of thing. Generative AI was very, or is very, very good at handling that unstructured data, which historically has been a challenge. So, you know, I think that's all really exciting. But when we're looking at agentic AI, the priority is really shifting from unstructured data to structured data and APIs. And what I mean by that is, you know, an agent can't update a customer record for eXamp, if that record is stuck in a legacy SQL database. Without an API endpoint, you can't really do much with it. So I think looking at the existing data estate and Ensuring that's modernized to be leveraged by these new agentic tools that are out there is absolutely critical. You kind of hear the notion of garbage in, garbage out with AI and so making sure that your data estate is modern, is, is accessible. The data is, as you know, as clean as can be. That's going to set you up for success in the, you know, kind of the long term. And that's going to be something. Regardless of which new models come out or which new architectures that underlying data and access to systems through APIs, that's still going to be a critical piece. The next thing, you know, when we kind of look at like, where do you start? Your data is in a good spot. You know, again, I see the most success happening with customers who you aren't necessarily pursuing the moonshot, first time anyone has done it sort of use case. It's really finding that high friction, low risk type of use case as, you know, kind of your lighthouse project. Start somewhere, start small and don't start with the most complex problem. Find something that's really visible that can illustrate the ROI of agentic AI and start building that out. I think that is the best way to start approaching these and look for processes that require employees to switch context. You know, something where they may have to navigate to three or four different systems to complete a single task or you know, get the required information that they need to complete that task. And that context switching sort of use case is really where agents thrive. So, you know, start small. But I think the time is now, the technology is here, time to start adopting it. So that would be my recommendation for an organization who's looking to, you know, kind of start going down this agent AI path.

Host: Thanks a lot for that answer, Travis and I completely agree that's one of the best ways for people in enterprise to take their first step in the technology. But, but as organizations experiment and look at these low friction, high roi, uh, use cases to make things happen once things start becoming real and when they start engaging with their actual business ecosystems. Consider a company, an organization which has a large supply chain data ecosystem and a supply chain data model could be large. So many tables, so many columns, so many kinds of business KPIs and metrics. How does an organization that is feeling overwhelmed with its data size, whether breadth wise or depth wise, or the volume size, they're feeling overwhelmed with that data. How would generative AI go ahead and uh, assist in bringing meaning, business ROI and actionable insights on that kind of an ecosystem? What's your let's say Genai101 architectural design for our listeners over here.

Travis Naurug: Hm, that is such a great use case and it's something that I see all the time as well. And I think it's almost a universal problem. It's hard to find the information that you need especially in the enterprise landscape where you've got petabytes of data across a bunch of different siloed systems. And so there's a lot of ways to kind of approach this use case. I mean there's everything from build your own solution and you know you're going to be using potentially like a vector database. Maybe there's an agentic aspect that's doing, you know, let's assume we're connecting into a uh, SQL database and we need to be able to do natural language to SQL. There's the ability to do that certainly and that's actually one of the use cases that I love is using an agent to do natural language to SQL. So as a non technical person who doesn't know SQL at all, you can still go and ask questions of your data and get those insights. I think though, you know there's build your own which for some use cases makes sense. But I think we're starting to see especially with things like Gemini Enterprise, uh, in particular which is Google kind of agentic platform for internal employee productivity. One of the big aspects of Gemini Enterprise is really connecting with pre built connectors to these existing data stores, to these existing third party enterprise systems that an organization is using and making that data searchable, leveraging all of the great technology that generative AI brings. So you know, in the Gemini Enterprise use case, Gemini Enterprise is essentially managing all of the parsing of the data, the chunking of the data, the embedding of the data, the storage of the embeddings and actually enabling essentially out of the box rag search, uh, retrieval, augmented generation for those enterprise data stores. So when I look at path of least resistance I think Gemini Enterprise is a fantastic solution because you know you've got these pre built connectors into call them the usual suspects of enterprise applications and data stores and it makes that democratization of, of search and data much more accessible to the organization. So it's a hard question to answer with one specific answer but I think platforms like Gemini Enterprise are really making this so much easier than it was even a year and a half ago where you know, you'd have to go out and build your own, you know, rag architecture. You have to manage all of the chunking, the parsing, the embeddings all the models that are involved in that process, all of the storage that's involved in that process. It wasn't a trivial project. But with things like Gemini Enterprise organizations are able to quickly connect to those enterprise data stores and get it in the hands of their employees. Um, and I think that is absolutely critical is just, you know, how quickly can we get our employees enabled and get access to these tools to improve their productivity, kind of smooth out their workflows and make sure that they have access to the data that they need access to.

Host: Wonderful. This has been really insightful and thank you very much for your all insights. I just want you to provide your finishing thoughts, your closing thoughts for all our listeners. What's your recommendations? What's your prediction for 2026? What should we be looking forward to Google that is bringing across in 2026? Uh, your closing thoughts on that and I'm sure everybody is quite excited as to what Google is going to bring in 2026. We've seen some great announcements coming at the last quarter in the last few months on the capabilities that we are getting, as he has rightly said on Gemini Enterprise price. But what are we looking at 2026 for from Google?

Travis Naurug: That's also a great question. I can't share too much on what's coming in 2026, but I'll leave it as agentic AI isn't, isn't going anywhere. I think now's the time to act. Now's the time to start exploring these solutions and identify where it fits within your organization. And for the listeners, I mean, one thing just to keep in mind again, I know I've said this a couple of times now, but don't necessarily go for that moonshot use case as your first use case. And the engineer in me hates saying that because, because those are the really cool, exciting things. But there's so much value to be had finding those relatively simple friction points that your employees are dealing with on a daily basis. And that is a great place to start with meaningful roi. The one other thing that I'll mention too is, you know, just within the Google cloud ecosystem, partners such as Treatence are absolutely critical to our mutual customer success. We can't do what we do without you guys. And so, you know, I really look at Treatence and partners in our ecosystem as being the glue that brings, you know, all of the great technology that Google is building into meaningful solutions for our customers. And you guys are doing it every single day out in the field. So talking to the customers out there, if this is something that you're interested in getting started with. Definitely. Reach out to treatence, reach out to your Google team. This is really how we operate best is is working with our partners, working with our customers and co creating these solutions. So now's the time to really act, you know And I think 2026 again I can't share too much. It's going to be a very exciting year. Things are evolving quickly on our end, which is exciting.

Host: I'm so happy and so humbled to be as part of this podcast and as an organization. I say it for everybody, for uh, the whole of the organization. We are so happy and we are so content and we ah, are taking the next steps on the collaboration with Google. It has been a tremendous journey. We look forward to 2026. We've been the chosen retail partner of the year, uh, from Google in year 2025. We look for something same as 2026 as we go more forward and, and thank you very much for your time here, Travis. It was incredible talking to you. Thank you for sharing all the insights and thank you for your time this morning.

Travis Naurug: Absolutely. Thanks for having me. Love the conversation.