How AI Agents Are Democratising Enterprise Data Strategy

Benoit Dageville had just finished conducting a seven-piece orchestra when he walked on stage at Snowflake Summit 2025 to deliver a message that would have surprised his younger self: stop building so much software.

After 12 years creating cloud data infrastructure that has transformed how companies store information, Snowflake’s co-founder now believes the secret to the platform’s future lies in what other people will build on top of it.

“My dream is that the platform becomes really extensible to run applications,” says Benoit.

“At the end of the day, most of what you’re going to do on Snowflake will not be developed by Snowflake.”

From Oracle frustrations to cloud-native architecture

The strategy stems from lessons Benoit and co-founder Thierry Cruanes learned at Oracle, where they witnessed two shifts that would reshape data management.

First, data was expanding beyond neat rows and columns to include machine logs, sensor readings and user interactions. Second, cloud computing promised unlimited compute power on demand.

“Analytics is really like ‘I want to analyse this thing,’ and you decide to analyse, and then you need a lot of compute power to go through very large volumes of data,” Benoit explains. Cloud computing offered a different model: get 100 servers and finish the job 100 times faster, while paying the same total cost.

Oracle’s systems could not handle this new reality. Hadoop – an open-source system designed to process vast amounts of unstructured data across clusters of commodity servers – was one alternative, but frustrated both founders. 

“I hated Hadoop: it was slow, it was complicated,” Benoit says. The excuse was always that complexity was inevitable given the data volumes. “We didn’t believe that.”

In August 2012, they left Oracle and rented a small apartment in San Mateo.

Benoit bought a whiteboard and for months the two engineers worked alone, trying to design a system that could separate data storage from compute power and make thousands of servers available instantly, then release them just as quickly. 

Today, Snowflake can allocate 1,000 servers in less than one second.

The data sharing breakthrough came from their own frustration while preparing Series A presentations.

They exchanged slides through email until Thierry discovered Google Docs.

“We looked at it and I said ‘Wow, this is amazing – sharing documents,’” Benoit says. They realised the same principle should apply to datasets. If the cloud could enable real-time document collaboration, why not petabyte-scale data sharing?

AI agents with built-in governance

Jeff Hollan joined Snowflake three years ago after 15 years at Microsoft. Today, as Director of Product for AI agents and Cortex AI applications, he oversees the company’s push into conversational data interfaces.

“The biggest change, especially for developers, is the amount of output they’re able to create is way higher with the use of AI,” Jeff says. “What used to take me six or seven hours to build and code now takes me 30 minutes.”

Beyond the coding benefits, business users can now query enterprise data through natural language rather than learning SQL or waiting for analyst reports. This democratisation was always Snowflake’s goal, but required AI to make it practical.

How, then, can you enable AI access to enterprise data without compromising security? Snowflake’s solution builds on existing permission systems rather than creating parallel security infrastructure. 

“Snowflake’s had really good governance for years. Now, we’re just going to put agents right on top of that, so you don’t have to do any extra work,” Jeff explains.

The approach means AI agents automatically respect existing data policies.

“I should not be able to see data for other employees at Snowflake. I don’t work in HR. I shouldn’t be able to see what the signing offer was for XYZ employees,” Jeff says. “Your agent just knows: ‘Okay, I know who Jeff is, I know what we do with our HR data. When Jeff’s talking to the agent, I’m not going to show Jeff the data that he doesn't have access to.’”

Snowflake’s marketplace strategy extends this principle to applications. Rather than building every possible data tool, the company enables specialists to create applications that run within its security framework.

“Think about hospitals,” Benoit explains. “You develop an application that can detect disease on scans, and it can be used by many hospitals. You need to bring this application to the scans, you want the application to be installed securely, within their governance.”

Skills evolution and platform maturity

Data quality remains a customer responsibility, though Snowflake provides tools for monitoring and validation. 

“Quality is a little bit tricky,” Benoit acknowledges. “Quality is something customers have to really put in their pipeline. It’s a little bit like coding: you need to write unit tests for your code.”

Data operations increasingly resemble software development. Both require testing, version control and quality assurance processes. Snowflake provides the infrastructure, but customers must implement the practices.

“Making something simple was really hard, and making something simple takes time,” he says. “Either when you say simple, it’s because the complexity has been internalised inside the platform.”

This approach requires more upfront investment but reduces ongoing customer effort.

Snowflake absorbs the complexity of managing thousands of servers, multiple AI models and security policies so customers can focus on their specific business problems.

The company now develops specialised AI agents for particular use cases. Data scientists can automate model development pipelines. Finance teams can query spending patterns across departments. Marketing analysts can compare campaign performance without writing SQL.

“I predict by next year, there’s going to be a dozen or so types of these specialised agents,” Jeff says.

“Whether you want to understand your Snowflake bill or you want to help build a pipeline, or whatever you’re doing, just have more agents helping with more and more pieces.”

Customer adoption follows predictable patterns.

Organisations start with simple queries and gradually tackle more complex analysis as confidence builds. The key is ensuring early experiences succeed, because trust lost through bad AI answers proves difficult to recover.

“Nobody wants their CEO to go talk to the data and get a bad answer, because then the CEO never is going to trust that agent again,” Jeff explains.

Jeff expects dramatic improvements in analytical speed over the next 12 months. Projects that currently require weeks might complete in hours as AI capabilities improve and workflows become more automated.

“Previously, doing this would have taken three months, now we’re at a place where it takes three weeks. A year from now, many of these things might be three days or three hours even,” he says.

The acceleration reflects both better AI models and improved platform integration. By reducing friction between asking questions and getting answers, Snowflake enables faster decision-making based on current data rather than historical reports.

Benoit’s original vision was democratising data access so every company could match Google's analytical capabilities. AI extends this democratisation to individual users within organisations who previously lacked technical skills to query data directly.

“Our goal was to democratise this data platform. We wanted everyone to be Google, but we understood it was every organisation, but not everyone within the organisation,” he says. “Now it’s everyone. AI is a very profound revolution which is very connected with Snowflake’s DNA.”