What It Takes for AI Agents to Earn the Right to Act

Enterprise AI has become very good at advising but still struggles to act. In Futurum research across more than 800 enterprises and IT decision makers, one in four cited the inability to write back to systems of record as a key barrier to putting AI agents into production. Agents can read and summarize, but most still stop short of closing the loop on real business processes.

Brad Shimmin speaks with Pavan Pothukuchi, Head of AI Platform at Snowflake, about what it takes to move from cut-and-paste AI to the autonomous enterprise. Pothukuchi explains why the agent control plane belongs inside a governed data platform. He also covers how semantic layers address accuracy and what business users need before they will trust agents to act.

His central argument is that agents earn the authority to act when they inherit identity, permissions, and audit trails from the data platform they run on. For enterprise leaders, that turns agent governance into a data architecture decision. One that should be made before any agent is allowed to write to a production system.

Key Takeaways:

🔷 The agent control plane belongs in the governed data platform. When agents execute inside a governed platform, they inherit its identity, permissions, and audit trails. Pothukuchi says Snowflake’s acquisition of Natoma, an enterprise MCP platform, and its 100-plus connectors give developers a governed path to write back to systems such as Salesforce and Microsoft 365, with no shadow MCP servers required.

🔷 Semantic layers turn conflicting definitions into trustworthy answers. Terms like revenue mean different things to sales, finance, and operations, and agents tend to hallucinate when they hit those conflicts. With 45% of IT decision makers planning to invest more in semantic layers over the next 24 months, Snowflake combines human-curated definitions in Horizon Context with semantic layers generated automatically by Cortex Sense.

🔷 Model routing only works with choice, cost control, and zero-egress inference. Pothukuchi says the data platform has to offer model choice and route each task by quality, performance, and cost. It also has to run inference next to the data so governance stays intact. He adds that routing too aggressively can erase cache-hit discounts, so efficiency depends on balance.

🔷 Trust in agent action rests on identity, scoped access, and audit. Agents need identities distinct from the people they act for, with access scoped to the task rather than inherited from a user’s accumulated privileges. Governance has to travel with the agent. Pothukuchi points to a bank in the research that took seven months to clear compliance for a single agent, a timeline he says only lineage and audit controls can compress.

🔷 Business users adopt agents that are accurate, simple, and visual. Six in 10 organizations still rely on keyword or hybrid search for their AI applications. Pothukuchi says Snowflake CoWork lets developers build agents that reason across structured and unstructured data and deliver role-scoped charts. He cites Fanatics as a customer that put a natural-language interface over more than two billion daily signals for its business teams.

Enterprises that run agents where their data, semantics, and governance already live will move from AI that advises to AI that acts, without rebuilding controls for every new workflow.

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

Learn more in the Futurum x Snowflake report on Operationalizing Autonomous AI.

‍

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

Pavan Pothukuchi:

Between the human curated semantic layers and the automatically built semantic layers, we are seeing agents able to act with a lot more accuracy and creating a lot more trustworthy responses for our users.

‍

Brad Shimmin: 

Hello, and welcome to Six Five. I'm Brad Schimann, and today we're exploring why so many enterprises and their AI initiatives are producing a lot of insights, but not taking a lot of action. And we're gonna talk a little bit about what has to change for us to be able to move from the cut and paste to the fully autonomous enterprise. And to do that, it's my pleasure to welcome Pavan Padukuchi, who's the head of AI platform at Snowflake. So Pavan, thanks for joining. Welcome to Six Five.

‍

Pavan Pothukuchi: 

Thank you, Brad. Nice to do this and yeah, would love to get started.

‍

Brad Shimmin: 

Yeah, great. Yep. I have many questions for you. So first up, let's start where I just started and talk a little bit about the current state of things. So we've had AI, generative AI now for six years, roughly. And we have a lot of intelligence. We have some very super intelligent models, you might say. But so far, we're not seeing a ton of autonomous action in the enterprise. So we're summarizing a lot of text, but not actually executing business processes. We're still in the cut and paste world, it seems sometimes. So tell me, what does it actually take to move from an AI sort of as an advisor to something that the enterprise can trust enough anyway to take action on their behalf?

‍

Pavan Pothukuchi: 

It's a really interesting question, Brad. As we looked at the research that Futurum did across 800 plus enterprises and IT decision makers, I think an interesting data point that stood out for me was one in four IT decision makers cited the inability to write back to their system of record as a key impediment to productionize AI agents in their enterprises. And what this tells us is agents are able to read from the system of record and, of course, enable users to chat about information, but are not necessarily acting on their behalf and closing the loop for them. What has to change in order for agents to act on the behalf of users is where the control plane for these agents actually live. When an agent executes inside a governed data platform, it has a paved path to execute responsibly. It can inherit the identity, the permissions, and the audit trails that the platform has and carries that over to the agent so that when it acts on the behalf of a user, it is doing that in a responsible manner. And that is easier said than done. You want agents to have independent identity. You want agents not to assume the privileges that a user has. And these things need to be scoped appropriately so that an enterprise is comfortable letting an agent lose on their production systems and systems of record. So at Snowflake, we have spent a lot of time on this problem with support for governed MCP servers through our recent acquisition of Latoma and Cortex tools. we are able to bring over 100 plus connectors to various systems of record so that a developer doesn't have to actually build a shadow MCP server that is not governed by their policies. So that process becomes a lot more easier and the ability to write back to systems such as Salesforce and Microsoft 365 etc. becomes easier because now there's a paved path and the governance controls and the audit workflows that enterprises have at the data layer flows seamlessly into the MCP layer and to the systems of record. And that just essentially makes the process of building with these tools that can act on a user's behalf a lot more easier. And we see that in practice. When you add MCP servers that can actually take action on behalf of a user, the engagement that users have with those agents significantly skyrockets. And I see it with my own use at Snowflake when my agents are connected to Jira, when my agents are connected to Google Drive, and the agents are able to actually write on my behalf and make complete workflows on my behalf, I see myself being a lot more productive. And my administrators are happy. They're able to look at all the access patterns and ensure that things are being done in a governed manner.

‍

Brad Shimmin: 

Yeah, it's important, isn't it, to think of agents not as people, but as artificial, autonomous systems that need their own access, need to be able to be validated, to have some sort of footprint that defines what they did and be able to work back from that in case something goes wrong. And I loved what you said about paving this for them. I think that makes a lot of sense. And MCP is one of those things that really smoothed out the road, if you will. And we at Futurum use MCP servers quite a bit for accessing all of the data that we generate, like that stat you mentioned, which I love that one, by the way. But, you know, sometimes you can seem that you have too much data. And agents and models just, you know, more specifically, can have a hard time with that. So if you have like, you know, a definition of a term, and maybe there's another definition of that same term, that's a little bit different. And if it encounters both of those conflicting definitions, sometimes these AI systems can hallucinate or contradict themselves. And especially when things like that change over time, that can be really frustrating for companies, I think, to manage. So my question to you is, how does resolving those kinds of definition conflicts really change what's possible for AI accuracy, and most importantly, trust?

‍

Pavan Pothukuchi: 

Semantic layers are essentially how this problem space is being addressed at large. And from the research, 45% of IT decision makers said that they are wanting to invest more in their semantic layer infrastructure in the next five years. 24 months. So I think what that is telling us is, I mean, I think it alludes to the problem that you're talking about, right? Agents in the same term, you know, can mean different things. Like revenue can mean different things for the sales team or the finance team or the operations team. Humans, of course, you know, encounter this problem too, right? But they talk to each other to figure out, you know, what does it actually mean.

‍

Brad Shimmin: 

Right. Argue, you mean.

‍

Pavan Pothukuchi: 

Yeah. And of course, agents don't, and they make up stuff when they encounter something that looks somewhat similar but has different context in three different places. They do not know how to deal with that. And so having a semantic layer, which is essentially metadata about the data, right? It can include metric definitions and how certain metrics are computed. It can include glossary and business terms and, you know, data dictionaries and verified queries and a bunch of different artifacts that essentially make, you know, that distinction between the various different things that humans understand but agents don't and codify them in a streamlined manner. And having something like this really helps agents get past the accuracy, at least one class of accuracy and hallucination challenges. And from the research, we have also seen that 25% of respondents cited accuracy as the top reservation to adopt generative AI more in their organizations. So I think there's all points to how do you build a semantic layer that both analysts or humans and agents can act on top of. And how do you make it easier to productionize? I think that those are essentially the challenges we see customers run into. And that is why we built Horizon Context. We built a Cortex Sense that kind of come together. So with Horizon Context, We provide a human curated and governed way to define semantic layers, which includes the things that I've talked about, business glossary, ontology, definitions, and metric computation logic, et cetera. And how do you start there? Also, typically, it tends to be challenging. And so we've given a lot of tools, including SemanticView Autopilot that enables customers to create these curated assets. And on top of that, Cortex Sense is An interesting innovation that we're seeing customers really resonating with customers a lot. And what it does is it looks at all the various business systems that exist in an enterprise and automatically generates a semantic layer by building a knowledge graph behind the scenes. And covers, so this essentially covers the larger data estates that customers might have, that might have complex, you know, legacy schemas, et cetera. So between the human curated semantic layers and the automatically built semantic layers, we are seeing agents are able to act with a lot more accuracy and creating a lot more trustworthy responses for our users.

‍

Brad Shimmin: 

Yeah, I mean, my colleagues have for years been saying it's the year of the graph, and I think we may actually finally be there. We're sure. So let's talk a minute about operationalizing these systems and bringing them into production. And, you know, the semantic layer is absolutely critical. And like you said, you know, it is accuracy is the name of the game. It used to be in the master data management world. It was all about the quality of your data. Now it's the accuracy of what the models say, given whatever data. bring to it. It is fascinating. But anyway, so no enterprise can really afford to run every single task through these super smart models. Like you take your, given any kind of Sunday, you know, Opus and Clod from Anthropic, etc. And the problem, obviously, is that they're too expensive and not always quick enough to do that, what you need in production. So we're seeing a lot of companies try to use smart routing to basically take the task that's at hand, whether it's something that's very prescriptive or something that is goal-oriented that comes up during a process and say, let's go to this model because it fits the job. So I love that, but I don't see it really taking off. I haven't seen a lot of companies doing that in full swing. So what does it really take? What's the underlying data platform got to look like for that to really work in practice, that kind of flexibility?

‍

Pavan Pothukuchi: 

Research too, right? So the data points that kind of align pretty closely. So over 48% of the respondents are looking at agentic analytics as a top trend through 2029. over LLM and generative AI model developments. So that kind of points us to the fact that customers are moving towards agentic systems versus just asking models a bunch of questions. And getting responses. And as models continue to evolve, as you pointed out, there are reasoning models, there are medium-sized models, there are smaller models, there are open-weight models, and the landscape just constantly is evolving each week. I mean, this week we had like four great model launches from multiple different providers. So I think what needs to happen from a data platform perspective is three things in my mind. First is model choice is non-negotiable. Customers are kind of inundated with lots of different models every single day of the week. And as a data platform, you got to disambiguate and kind of provide, you know, class of models for different types of tasks. And then the second thing that needs to exist is, as you're pointing out, routing, right? And it's a complex topic because you have to consider the requirements from customers across multiple different dimensions, across quality, across performance. and across cost and kind of make the right determination given a task or a request based on these parameters, which model best addresses the need of the customer. And so having a mechanism that drives the right routing logic at a data platform layer, I think is important. And the third one is zero egress inference, right? So if you have your inference platform separate from your data platform, you're essentially taking data out and running it elsewhere. And when you do that, the things that we've talked about earlier, the governance controls and who has access to what, start kind of getting very messed up. So being able to drive inference closer to where your data is, I think is an important part to solve all of these challenges in a secure and governed way. I think the last thing related to all of this is cost. And from the research, we have also seen about 17% of the respondents continue to cite cost as their number one barrier to more broadly adopting AI. And so techniques like routing at a model layer helps reduce costs. And things like adaptive compute, which we have done at a query and warehouse layer, which is where the data comes from for executing these inference tasks and kind of ensuring that that layer also is optimized. So both of these are important to ensure that you know, data platform provides controllable costs, it provides better routing, it provides the better choice for running all kinds of tasks for customers.

‍

Brad Shimmin: 

Yeah, cost is really important. And as a practitioner, I run into this all the time. And especially when you're using multiple models to do different tasks, like doing red and blue teaming, for instance. And boy, one of the biggest challenges is how do you manage the cash that you have across those models? And as we've all learned, that's how you either lose or save money. People are cash hits are what they measure.

‍

Pavan Pothukuchi: 

Yeah, I think that's a great point, right? So because for cache hits, you get, you know, 90% discount when you run these models. And if you are too aggressively switching across models, you're essentially losing that benefit, right? And you have to, again, do your prefills and your costs could actually be higher when you do routing versus the other way. So I think making, like creating that balance is an important part of the routing process.

‍

Brad Shimmin: 

Yeah, let's go back just for a moment to talk about granting these models authority or these agents authority to take action. And we've seen over the last while that executives are getting a lot more comfortable with using agentic AI to actually query data and analyze data and make decisions on that. But it seems that to get their buy-in, to get them to really say, yes, let's give the reins over to AI, they seem somewhat hesitant. Can we talk really briefly about what it really takes for an organization to really hand over that kind of authority to agents?

‍

Pavan Pothukuchi: 

So I think there are three things that are critical for agents, for enterprises to trust agents to perform actions. So those are identity, governance, and audits. I think this is generally the who, what, and when framework, right? That generally applies in the governance world and applies to agents as well. And to the earlier point that you were making about human identity versus agent identity, I think that's an important, you know, part of doing this because you do want to distinguish between whether a human performed an action or an agent performed an action. So that, you know, like if it's a wrong action, you can go back and figure out, you know, what really happened. So that's like point number one. And then point number two is scoped access for tasks, right? which is governance and ensuring that an agent doesn't get all the privileges that a user has, which might have been accumulated over years of working at an enterprise. So you don't want to give all of that to an agent that is tasked to do one specific thing. because, as you know, agents try to do what they are asked in so many different ways. And so you want to ensure that you're giving very scoped access and not letting them go unfettered. And then lastly, you need to be able to look at what they've done and when they have done it and how they have done it. So the audit trail end-to-end is an important constituent for responsible actions. And then governance has to travel with the agent and not stay behind at the source, right? So wherever you have defined your access policies, your guardrails, those all need to apply across the journey. And when you do that in a different agent execution framework, you kind of lose that fidelity. And you have to either reapply them or rebuild them, which creates a lot of friction for customers. Also from the research report, over 22% of the users cited governance and security as their primary impediment. for productionizing agents. And so executing agents close to the data platform and ensuring that the same governance controls are applied uniformly is an important way to address this reservation. And then lastly, all these governance policies need to be ensured that they're applied in the right path, which is not like R-I-G-H-T, not R-I-G-H-T, but W-R-I-T, that path. We had, I think, One example that stood out for me from the research report is this banking customer that said it took them like seven months, I think, to go past the compliance aspects from the time they whiteboarded to actually productionize an agent that executed workflows. on their behalf. And I think that kind of time frame can only be compressed if you have the right audit trails, the right lineage, and the right controls that you can stand behind to your AI or governance board to fast track execution.

‍

Brad Shimmin: 

Yeah, because when you open up your system and like you rightfully said, it has to have optionality. You need to be able to bring in the model you want. You need to bring in the framework you want. You may need to bring in another platform. You may have agents running elsewhere that you're connecting in. And that's one of the hardest things because the minute you sort of allow that agent to take action based on something else or just to reach out, that boundary you had for security is compromised unless you do something smart, like set up a gateway to handle those like you guys have. But I love that product, by the way. Anyway, I want to finish up, last question for you. And it can be quick fire if you'd like, but it's one that I think is on everyone's minds. Let's assume that we figure all of this out, all of these technical issues out, and we understand how to secure AI and how to govern it and how to give it the ability to write back to our systems of record. What do business users care about that? How do we get them on board? Because there is some concern and worry amongst business users. And worse than that, there's the lack of concern and worry amongst business users who are very gung-ho about using this. And they might bring in their own agentic systems to work on company data. What do we do to help them, but not hinder them?

‍

Pavan Pothukuchi: 

Yeah, I think that's a great point. I think all the things that we have talked about applies to a developer that is building agents. And as a business user, you care about accuracy, you care about ease of use, right? If we expose all of these constructs to you as a business user, you're not going to use the system. So it needs to be really simple to use. and it needs to be able to give you results that you trust, right? And once you get a response that you cite somewhere and you're corrected, you kind of lose trust, right? Like you don't want to go back to the system. So there's a tight balance between ensuring that the systems are accessible and usable and they provide right responses. And speaking of which, one of the data points that was interesting to me from the research report was, I think six in 10 organizations cited they're still doing keyword-based search or hybrid search. for their AI applications, which just seems like it's just decades old technology at this point. And so there's a leap that goes from keyword-based search to semantic search and the semantic layer aspects that we have talked about. While a business user might not care about all of these things, it is important to ensure that you're providing, as a developer, you're providing the right state-of-the-art technology in your agent workflow, so that the business user can actually trust the underlying result that you're providing. So that is why we have built Cowork, which is essentially a product meant for business users. So as a developer, you can build an agent that can go across your structured data and unstructured data, and retrieves the right results and distills that and even create charts because, you know, like you don't want a wall of text if you're a business user, you're busy, you're going about your day and we as humans are visual people and so we need to see something visual to process information. And so that's why business users gravitate more towards, you know, charts and dashboards and reports and things like that, right? But they're super static. And with the new way of being able to do things with agents, Core gets this data, distills it for you, and creates beautiful charts that make you process the information a lot more faster. And the data that you're seeing is scoped to your role and all the role access policies that an enterprise might have set all apply so that you're not seeing data you're not supposed to see. So that's extremely important in the enterprise context, as you very well know. And there are, we have customers like Fanatics that are using the system to put, you know, there's 175 million users, they have over 2 billion daily signals that they get across these users, and they put all of this together in a natural language interface in front of many of their internal teams so that they're able to quickly derive insights and identify monetization opportunities and cross-sell opportunities. So that has been really interesting to see how fast they went from creating agents to actually putting those agents in front of their business teams.

‍

Brad Shimmin: 

That is incredible scale. My goodness, that's a lot of people. And I like very much, because what I heard from what you were saying just now is that you need to meet people where they are in business. And you can't always go, well, you'll just have to export this to SPSS and then come back with some sort of chart or wait for somebody to do it. That era's gone. Yeah, so this was fantastic. Thank you very much for joining me today, Pavan. I really appreciated the great conversation.

‍

Pavan Pothukuchi: 

Likewise, I've enjoyed it. Thank you so much. And I've learned a lot from the report. So appreciate the partnership.

‍

Brad Shimmin: 

Great, me too. And thank you everyone for tuning in to our Six Five broadcast today. I really appreciate that. And don't forget to hit subscribe. We love that. And follow us on our socials, which you can see probably right now. And do check out all of our coverage on sixfivemedia.com. All right. Till I see you next time, have a great day everyone.

‍