Snowflake’s AI agent strategy is built around enterprise data security

ChatGPT may have introduced millions to the power of conversational AI, but the real revolution is happening behind enterprise firewalls. AI agents are autonomous systems that can break apart complex business challenges, make data-driven decisions, and execute multi-step workflows that represent the next leap forward from simple chatbots. These digital workers promise to transform everything from customer service to scientific research.

But there’s a catch. Despite all the excitement and investment pouring into agentic AI, three fundamental problems are keeping these systems from radicalising their potential. They can’t remember what happened five minutes ago, they make critical decisions based on internet knowledge rather than proprietary business data, and they have an unfortunate tendency to act first and ask questions later.

The Agent Advantage: Beyond Simple Chatbots

The distinction between chatbots and AI agents becomes clear when we understand what drives their behavior. While chatbots excel at answering questions, agents are designed to solve problems. As Jure Leskovec, Professor, Stanford University & Co-Founder and Chief Scientist, Kumo explains, “The key capability or aspect of the agent is to be able to autonomously take actions and then observe the result of those actions in the environment, reason on top of that result and take the next steps.”

This autonomous loop to act, observe, reason, adapt fundamentally changes what’s possible. Instead of asking a chatbot for travel advice and getting a list of suggestions, you can instruct an agent to plan your entire New York trip. The agent breaks this down into smaller tasks: booking flights, finding hotels, arranging transportation, making restaurant reservations. It coordinates these activities while juggling your budget constraints, dietary preferences, and schedule requirements.

But here’s where it gets interesting for businesses. These agents don’t just follow scripts they can specialize using the unique data that only your organization possesses. “If we are all using the same base models, then differentiation comes from the data,” explains one industry expert. “So it becomes so much more important that these agents, their decisions are not this kind of, you know, I read this on the internet common sense type of decision, but it’s really a decision that’s rooted in the enterprise data and only that will lead to, I think, accurate decisions and also decisions that then bring back the value.”

This data-driven specialization creates genuine competitive advantages. An agent working for a financial services firm doesn’t just know general investment principles, it understands your specific risk models, client histories, and regulatory requirements. An agent in a manufacturing company doesn’t just know about supply chains it knows your suppliers, your quality standards, and your production bottlenecks.

The Memory Problem: Digital Amnesia

The first major roadblock is surprisingly human in nature these sophisticated AI systems suffer from what amounts to digital amnesia. Every conversation, every task, every interaction starts from scratch. Jure Leskovec, a researcher working at the intersection of academia and industry, puts it bluntly: “We as humans get enriched, we remember, we interact with the world, we learn directly with the world. Most of these models, they are frozen. It’s almost like banging you on the head with a hammer after every interaction. It doesn’t go away.”

This memory gap creates frustrating limitations that anyone who’s worked with AI systems will recognize. A customer service agent can’t remember that you called yesterday about the same issue. A financial analysis agent can’t build on insights from last week’s report. A project management agent can’t learn from the mistakes it made during the previous product launch.

The problem isn’t just inconvenient, it’s economically wasteful. Without memory, agents can’t develop expertise, can’t avoid repeating errors, and can’t build the kind of contextual understanding that makes human experts valuable. Every interaction requires rebuilding context from scratch, which slows down processes and increases the chance of miscommunication.

The Enterprise Data Challenge

The second bottleneck cuts to the heart of what makes AI agents valuable in business settings. For all their sophistication, most agents make decisions based on generic internet knowledge rather than the specific, proprietary data that drives real business value.

Amit Prakash, whose company focuses on data-driven decision making, explains the challenge with a concrete example: “If you say I’m a customer support agent, I need to estimate what’s the likelihood a customer is going to churn. Is this a high-value customer or a low-value customer? If I want to offer them a promotion, what promotion do I offer them? Will they take it or not? These are all decisions that the agent needs to make in order for it to be effective.”

The stakes get even higher in regulated industries. “If an agent is operating in a hospital setting, the agent needs to decide what’s the risk of readmission of this patient because based on that, they are going to do one thing or the other thing,” Prakash continues. “These decisions are not something you read about on the Internet. They need to be grounded in the data. For example, data stored in snowflake and so on. It’s private data.”

This isn’t just about having access to data it’s about understanding the nuances, patterns, and relationships that exist within a specific organization’s information ecosystem. An agent that can tell you general best practices for customer retention is useful. An agent that knows your specific customers, their purchasing patterns, their communication preferences, and their likelihood to respond to different incentives is transformative.

The Impatience Problem

The third bottleneck reveals an almost comically human flaw in our AI systems: they’re terrible at asking follow-up questions. Current agents have a tendency to jump to conclusions and start executing tasks based on incomplete or ambiguous instructions.

Leskovec uses a restaurant analogy to illustrate the problem: “These models are very proactive. They jump to conclusions and just do the thing. It’s almost like if you go to a restaurant and order food, you go there and you say, I’m hungry. The food does not just show up in front of you. The waiter asks, do you have any food restrictions? Would you like this type of food, that type of food? Maybe through a few patients, then you get the food you like.”

Instead of this natural back-and-forth that characterizes effective human collaboration, agents often deliver results that completely miss the mark, leading to what Leskovec describes as “error and correction. Rather, how do they actively seek information? How do they seek additional feedback? So that when they go and act, that action is much more active.”

This eagerness to please creates inefficiency and frustration. Rather than spending a few moments clarifying requirements, agents rush into execution mode, often producing work that requires significant revision or complete redoing. The irony is that a few clarifying questions upfront could save hours of downstream correction.

The Infrastructure Challenge

Even if the core technical problems are solved, deploying AI agents at enterprise scale introduces a whole new set of infrastructure challenges. The vision of agents working seamlessly across different systems, companies, and platforms requires standardization that barely exists today.

“If you want agents to work at scale, it needs to work across companies, across ecosystems and platforms,” explains one industry leader. “So having standardised ways to actually interoperate with secure way is going to be super critical. And we have seen standardisation starting to emerge like the MCP and other things. But I think that is going to be one of the main critical aspects.”

The complexity multiplies when you consider governance and access control. Organizations need granular control over what information agents can access and what actions they can take. This becomes exponentially more complex when agents operate across company boundaries or within shared ecosystems. As one expert puts it: “How can you build a governance that works across agents, across companies, across ecosystems, so that they actually are truly a co-worker in terms of how they interact with agents.”

Then there’s the evaluation problem. Traditional machine learning systems are predictable, given the same input, they produce the same output. Agents are different. “Agentic systems are way more complex than traditional machine learning systems. Like the classification system or the regression system, you can run emails. They are deterministic. Agentic systems are traditionally non-deterministic,” explains one researcher. “And so how do you run emails against a system? And that’s where I think the industry needs to spend more time.”

Safety adds another layer of complexity. As the same researcher notes: “These agentic systems can be extremely powerful. They can be connected to different systems, different environments. You can do tool calling and so forth. And how do you build ingrained safety in the system? And so that is also something which we are taking very seriously.”

Snowflake’s Enterprise Approach

Understanding these challenges, Snowflake has built its AI agent strategy around a fundamental principle: never let enterprise data leave the secure boundary. “What we do is we bring all of the large language models to run inside Snowflake so that no data leaves that Snowflake security boundary,” explains a Snowflake executive. “So this means you can trust from a security perspective that your data is secure. You can trust the governance of it. So all of the access controls that our customers build in their unified data platform is respected.”

Snowflake Intelligence, the company’s flagship agentic system, tackles the decision-making bottleneck head-on by ensuring agents have access to the full spectrum of enterprise data while maintaining strict governance controls. All the access permissions and data governance policies that apply to human users automatically extend to AI agents, creating what the company calls a “unified data platform” approach.

But Snowflake’s team recognized that having access to data isn’t enough, users need to trust the agent’s reasoning. Their solution involves multiple layers of verification and transparency. “When we automatically generate code, can we make sure that if there is a match to a verified query, we’ll tell the user so that they can have high confidence in the answer,” explains one team member. “Whenever we give an answer from any type of documentation, we’ll always show citations with a specific snippet from that documentation. Again, to increase confidence.”

The practical impact shows up in day-to-day usage. As one Snowflake executive shares: “I’ve been using Snowflake in terms of this actually every day. One example is I meet a lot of customers this week. Before meeting customers, I actually look up and ask Snowflake, how is a customer actually using Snowflake? And that is actually pretty insightful and make it much more productive in terms of actually me giving help and offering help to the customers.”

Research Frontiers

The academic and industry research communities aren’t sitting idle. At Stanford, researchers are pushing agents into uncharted territory automated scientific discovery. “In my research at Stanford, we are really trying to push this forward to automate scientific discovery process where the agent decides what experiment to run next, observe the result of the experiment and reason about the next steps,” explains one researcher.

This work has implications far beyond academic laboratories. The observe-reason-act cycle being refined in scientific contexts could improve how agents handle complex business processes, from supply chain optimization to financial modeling.

On the technical infrastructure side, breakthrough research is addressing fundamental performance bottlenecks. Snowflake’s research team recently published work on “shift parallelism”—a technique that automatically adjusts processing approaches based on demand. As one researcher explains: “One of the key elements of choosing between latency and throughput is for latency, you need tensor parallelism. For throughput, you need sequence parallelism. So what the research team actually came up with is automatically based on the traffic, and a move between these different parallelism paradigms without changing the memory thing, and that actually almost doubles in terms of throughput and latency for us.”

Meanwhile, Meta’s open-source contributions are helping standardize the tools needed to build and evaluate agentic systems. “We have made our contribution by building LamaStack, which has an agentic emails framework, which we want to standardise all open source,” explains one Meta researcher. “We launched LamaGuardian. We basically came up with the Defenders Programme, the Safety Defenders Programme.”

The Transformational Decade Ahead

Despite current limitations, industry leaders paint an optimistic picture of the next ten years. The vision extends far beyond workplace efficiency, it’s about fundamentally democratizing capabilities that were previously available only to large organizations with significant resources.

“AI is going to democratise actions on top of the data,” predicts one industry executive at a Snowflake event. “So I do see there is going to be places where everyone would be empowered to kind of do things without actually having to think about where things are, where the data silos are. Because all of that will be kind of solved by agents and AI.”

The democratization theme resonates across multiple sectors. In education, one researcher envisions profound change: “Just imagine a student in a rural village somewhere in Africa or India is able to get an AI tutor to learn exactly the same thing what somebody in an urban city for an expensive private school is getting that same knowledge, right? It is going to democratise the education, healthcare, finance, everything across the world.”

But perhaps the most compelling vision involves AI agents as collaborative partners rather than replacements. “AI is going to, more than taking away your jobs, it’s going to amplify the human potential,” explains one executive. “If I have an AI coworker sitting there helping me debug my code, analyse my code, or if I’m a financial analyst, do all the financing tasks, and I can control it, I can actually elevate what I’m doing to a completely different level with great precision, great creativity, and really bring in tremendous productivity to my job.”

The shift toward continuous optimization represents another frontier. One researcher describes “this kind of notion of continuous decision making where let’s say supply chains and things like that, they get continuously optimised by these agents in real time, getting information, making decisions, and streamlining these processes.”

The Data Advantage Era

The conversation around AI agents ultimately returns to a fundamental truth about competitive advantage in the AI era. As one industry expert puts it: “In every revolution, we always see something getting commoditised and something becomes kind of a bottleneck or something that becomes hard. The way I see now going with this one is that kind of skilled work is getting commoditised. It’s like kind of the white collar jobs are getting commoditised. But I think where the bottleneck will become is in terms of making the right decisions and getting the value from the data that makes any enterprise unique to make those decisions rooted in that. So I think the value of data will become even more important than it already is.”

This insight cuts to the heart of why solving the three critical bottlenecks memory, decision-making, and over-eagerness matters so much. Organizations that can successfully deploy AI agents with persistent memory, access to high-quality enterprise data, and sophisticated reasoning capabilities will have significant competitive advantages.

The technology foundations are being laid now through research initiatives at companies like Snowflake, Meta, and academic institutions like Stanford. The infrastructure for secure, governed, cross-platform agent deployment is emerging through standardization efforts. The evaluation and safety frameworks needed to deploy non-deterministic systems are being developed and open-sourced.

But the real transformation won’t come from the technology alone it will come from organizations that learn how to integrate these capabilities with their unique data assets and business processes. The next frontier isn’t just about building more capable AI models. It’s about creating the infrastructure, governance frameworks, and trust systems that allow those models to operate autonomously while remaining grounded in the specific data and context that makes each enterprise unique.

The companies that solve these challenges first won’t just have better AI tools, they’ll have digital teammates that remember, reason, and act with the full context of their organization’s knowledge and experience. That’s the promise worth pursuing, and the bottlenecks worth breaking through.