Building the data foundation for the agentic enterprise

Ahmed Hafez, Director, Solution Engineering for the Middle East, Turkey and Africa at Snowflake, discusses the company’s vision for an enterprise AI control plane, the importance of trusted data and context, model flexibility, and the evolving governance requirements of AI agents.

Snowflake wants to become the control plane for the agentic enterprise. What does that mean in practical terms for customers who are already building AI agents?

There are three key things to consider, and it’s important to understand why a control plane is needed in the first place. The first is trusted data and context. To give you an example, a finance person talking about revenue and a salesperson talking about revenue can mean very different things. The way revenue is calculated in both instances could also be very different. So, getting to a point where you can actually trust the data is really important. The second is model choice. With so many AI capabilities available today, how do you choose the right model for the right use case, based on cost, quality and performance? You want to choose the right model, but you also need to manage your costs. The third is governance. You need to control what different people and AI agents are doing across the organization, limit what agents can do and deploy the right guardrails. You also need to determine whether an automation is high-risk and requires a human-in-the-loop, or whether it can run independently.

Snowflake is looking years ahead and asking what enterprises will need when AI is widely adopted and organizations are well on their way to becoming truly agentic enterprises. How do we govern these environments and protect enterprise data more holistically? That’s where the control plane comes in. We’re already supporting customers by introducing capabilities ranging from model selection and Cortex Search to ontology. We’ve always had governance at the data layer, but we’re now extending governance and control into the AI layer, with the ability to manage it centrally across the enterprise.

What are the biggest challenges you’re seeing around data context and governance as organizations begin deploying AI agents?

One of the biggest challenges we see is when organizations try to separate their AI strategy from their data strategy. At Snowflake, we have a saying: “There’s no AI strategy without a data strategy.” Access to data is one of the fundamental challenges. For AI to be relevant, organizations need access to data that may reside both inside and outside the enterprise.

It’s also about sourcing data at the right latency. Do I need it in real time, near real time or in batches? Am I making AI-driven decisions immediately, or analyzing something after the fact? At the core of it, I need access to structured, semi-structured and unstructured data in a governed environment. Think about predictive maintenance. If I’m trying to understand how weather or corrosion affects machinery, I may need external weather data.

We’ve introduced tools such as Openflow to help organizations source data. We’ve also created zero-copy integrations with platforms such as SAP, Workday and Salesforce. Whether it’s external datasets through the marketplace or internal enterprise data, Snowflake helps unlock the access that AI needs.

The second challenge is accuracy. When we talk about accuracy, the first thing people usually say is that context is key. It’s not enough for an AI agent to access data. When someone asks it a question, the agent needs to understand the context and provide an accurate answer. Otherwise, the AI agent remains an experiment. It won’t move beyond that stage into meaningful business use. This is why Snowflake is investing in semantic layers and ontology capabilities to establish the context needed for accurate results.

The third challenge is governance. Whether I’m using analytics, AI or applications, my data should be centrally protected. What I don’t want is for an AI deployment to create a situation where someone sees information they shouldn’t have access to.

Then there’s the ability to do all of this at scale. This is where the control plane, AI Gateway and Cortex Search come together to help organizations centrally govern and manage AI solutions, whether they’re built inside Snowflake or externally.

And then there are two areas that, in my view, are probably the most critical. One is knowing how to get there. This is where we work with implementation partners, our Forward Deployed Engineering teams, and professional services to help customers understand how to achieve their business outcomes through established best practices.

Finally, there has to be a clear vision from the organization’s leadership. It should always start with how AI is going to benefit the business. Not simply an experiment or a desire to say we’re using AI. Every AI use case should have a business value associated with it, and leaders need to measure whether that value is actually being realized.

Snowflake is bringing models from different AI providers into Cortex. As these models become increasingly interchangeable, does enterprise data and context become the real differentiator?

Model choice is a huge differentiator from Snowflake’s perspective. I often use the analogy of streaming services. My wife and I have this discussion every year about which services we should subscribe to. I like Prime, she likes Netflix and my kids like Disney. It’s a similar situation with AI models. Snowflake gives customers access to industry-leading models from providers such as Anthropic, OpenAI, Mistral, and DeepSeek, among others.

Today, the best model for a particular task might come from one provider. Tomorrow, another provider might introduce something better. The pace of disruption in AI is incredibly fast, and the technology is constantly evolving. So my question is, why get locked in? You should be able to choose the right model for each use case, based on its capabilities, performance and cost profile, to get the best possible outcomes from AI. Now, when you combine that model choice with a strong context layer, you can improve accuracy across models. That’s where organizations can begin to differentiate themselves and gain a competitive advantage through AI.

But even with different models available, couldn’t individual use cases still become locked into a particular model? How does that flexibility work in practice?

They can, but the benefit with Snowflake is that customers have the flexibility to change models without being locked into a particular provider or contract. We bring industry-leading models within Snowflake’s security boundary, giving customers secure access and the flexibility to choose the model that best suits their use case.

Let’s say I’m using a vision model from a particular provider today. Tomorrow, another provider introduces a vision model that’s better suited to my use case. It might not necessarily be better in general, but for my use case, it’s better. I can change my process to use that new model and continue operating as before. We’re giving customers the flexibility to adapt as their business requirements change and as AI models evolve.

Would switching between models affect the outputs? How do you ensure consistency when an organization moves from one model to another?

That’s a good question, and this is where the context layer becomes extremely important. The context layer is model-agnostic. It’s defined at the data layer, not at the AI layer. So, irrespective of which AI agent you’re using, whether it’s inside Snowflake or outside, the agent can query the data and rely on Snowflake’s semantic layers and semantic views to provide the right business context. That’s what helps maintain accuracy, irrespective of which model or provider you’re using.

Is Snowflake ready to govern AI agents even if they aren’t built on Snowflake?

Absolutely. We’re approaching the platform holistically, with openness as a fundamental principle. And that applies to AI agents as well. We’re trying to help organizations build the agentic enterprise, and that requires an open approach. We recognize that enterprises will deploy agents both within Snowflake and across external platforms.

As organizations adopt AI across departments, use cases, and datasets, the key question becomes: how do you create a central governance layer that works across all these environments? How do you apply policies centrally, irrespective of where an agent has been built or deployed? That’s what we’re working to enable. Organizations need a consistent way to manage their AI agents, whether they’re Snowflake-based or external. This is a very real requirement, and it’s something we’re taking extremely seriously.

As AI agents gain the ability to query data and take actions across enterprise systems, how does Snowflake extend traditional data governance to govern what these agents are allowed to do?

Our governance approach has always been centered around the data. We refer to this concept as data gravity. The idea is that you bring AI to the data, you bring analytics to the data and you bring applications to the data. You don’t move the data to each of these things.

Why is that important? Because I might run an analytics query today and an AI query tomorrow, but my data should always remain protected. If I’m not supposed to see a particular column of data, I shouldn’t be able to see it through any application or AI interface either.

Now, with AI agents, we need to extend that thinking. We need to start treating agents much like human users when it comes to access and permissions. What does an agent have access to? What limits should apply? What controls and guardrails do we need to enforce?

Consider automation. I might have an agent performing a low-risk operation, and I’m comfortable allowing it to run fully autonomously. But if another agent is carrying out a high-risk operation involving important information, I may want human oversight. That’s where human-in-the-loop becomes relevant. So, our capabilities start with governing the data and extend into governing what AI agents do. We also need to audit their activities. What identity is an agent operating under? What information did it access? How did it access that information? What actions did it take? The question isn’t simply, “What can the agent do?” It’s, “What’s the agent allowed to do?”

What’s the opportunity for Snowflake in the region, and how is an event like Ai Everything Abu Dhabi helping you showcase what you offer?

The opportunity in this region is remarkable. I often say that data transformation can be driven very effectively from the top down when there’s a clear vision. And when you look at the UAE, the ambition of the country and the investments being made are incredible.

Take Abu Dhabi, for example. It has a strategy to become an AI-native government by 2027. There’s also a clear national AI strategy, backed by significant investment and a strong commitment to implementation. One example is a delegation of AI officers from government entities who travelled to the United States to visit our headquarters. They met our founder, Benoit Dageville, and our CEO, Sridhar Ramaswamy, to discuss how Snowflake could support their AI ambitions and the best practices for successful implementation.

Events such as Ai Everything Abu Dhabi are also helping drive this momentum. And when you consider how these initiatives extend across government entities and into the wider enterprise sector, the opportunity becomes even more significant.

What excites me most, beyond the level of investment and ambition, is the nature of the conversations we’re having with customers. Everyone wants to adopt AI. The question they’re increasingly asking is, “How do we get there?” And our answer is straightforward: there’s no AI strategy without a data strategy. The real conversation is about how we help organizations achieve the business outcomes they’re looking for.