If you're an architect or a CIO, you've probably been handed two instructions that contradict each other. The board says: "Give us AI this year." The playbook says: "First, fix the data. Migrate. Centralize. Govern. Then do AI."
The default reaction is to start moving data into a centralized platform. But this creates a problem. Every time you move data, it adds cost and risk. That should be a headache for C-level executives if they're not aware of it.
For years, moving data was the price of doing any analytics at all. To query data, you had to go through the proprietary compute of whatever system owned it. Data in Oracle required Oracle. Data in Teradata required Teradata. Cross-system analysis meant physically moving data into one place first, so centralization became the precondition for everything that followed, AI included. That sequence made sense 10 to 20 years ago, but it’s no longer a technical requirement.
"Break the Sequence" is a resource we created to explain how the architecture shifts of the past few years make it possible to start AI on the data you have today, while modernization continues on its own track.
Open formats changed the economics #
Open table formats, specifically Apache Iceberg™, broke the dependency between data and the system that stores it. Any compatible compute engine can read data in Iceberg format, decoupling storage from compute.
Iceberg has really opened up interoperability between technologies. You can publish data products in Salesforce, in SAP, in Workday and consume them from whichever technology you're using at the time, with no copying or complex pipeline building.
For a manufacturer with operational data spread across an ERP plus numerous other systems, those sources can now be queried as a unified asset. For a retailer with customer data in a CRM, transactions in a point-of-sale system and inventory in a supply chain platform, the same applies. The data stays where it is. The AI workload runs on the data where it sits. This is the leapfrog strategy in practice: AI can start on the data you have today while modernization continues on its own track. Your AI timeline and your migration timeline can move independently, without leaving existing investments behind.
Because the formats are open, the data stays yours. Iceberg tables can live as Parquet files in your own cloud storage, readable by other Iceberg-compatible engines. Reads and writes work in both directions, so data can be useful without being duplicated.
Governance that travels with the data #
The old model tied governance to location. Data had to be centralized before it could be governed consistently, making every governance requirement another reason to finish the migration first.
That dependency has broken too. Snowflake Horizon Catalog applies one set of policies across data, data products, apps, models, agents and other assets inside and outside the platform. Data accessed through Snowflake carries the same governance wherever it sits: fine-grained role-based access control, masking and row access policies, column-level lineage and access history.
Horizon Catalog now extends Snowflake row access and masking policies to external engines such as Apache Spark, Trino and PyIceberg, in public preview. Horizon applies the policies when it plans each scan, so engines receive only the data a user is allowed to see, and each operation is recorded in Snowflake's access history.
Because Horizon enforces these policies centrally, they apply automatically to every caller, whether a human analyst, a BI tool or an AI agent.
When stricter controls are needed for AI, Snowflake’s Agent identity capability lets administrators define what data an agent can access, independently from the user who invoked it. Masking and row access policies can target agent sessions directly. Actions are attributed to both the user and the agent that carried it out, so compliance teams get an audit trail of what AI touched and when.
For European enterprises, this is not yet another compliance detail. It is the architectural foundation. GDPR, data residency requirements and the EU AI Act mean that where data sits, who can access it and how AI decisions get audited are design choices and responsibilities. An architecture that is designed to govern data in place, with policy enforced consistently and lineage traceable end to end, can help organizations support those requirements without forcing consolidation into a single region or system.
The architecture #
Snowflake lets you work with your data where it already sits. One governed layer spanning compute, catalog and storage, so AI runs on your data as it is today, not after it’s been replatformed. This is the leapfrog in practice: your AI timeline and your migration timeline, decoupled.
Compute and AI: Run workloads where they make sense, on Snowflake or external engines.
Catalog: One governance layer, across open and Snowflake-managed data.
Data and storage: Query data in place, no copying required first.
Agentic Control Plane: Agents operate inside the same governed boundary as everything else.
Connecting a global estate #
NTT DATA, a global leader in AI, digital business and technology services, operates one of the world’s largest and most complex technology estates, serving clients across all major industries. With data distributed across decades of accumulated systems, the challenge wasn’t a lack of investment, it was making that investment work together without launching another multi-year consolidation program.
“For NTT DATA, the partnership with Snowflake goes beyond simply putting more data into another platform. It’s about making data easier to access, govern and consume across a complex global technology landscape. By doing so, we’re creating a high-performing, agile data foundation for AI, innovation and transformation, while maintaining the flexibility to evolve our technology landscape without introducing additional complexity.”
Nick Cheesman
CIO, NTT DATA, Inc.
Avoid the wild west #
Speed without architecture creates its own problems, much like a city built without an urban plan.
In the rush to deliver something, organizations can create a wild west of different AI tools. The result is fragmented solutions and inconsistent outputs that erode trust in AI results. The answer is to build AI workloads within an existing governance framework from day one, even if that framework is still developing. New use cases become testing grounds that provide insights for evolving the data strategy and architecture.
Modernization still matters, and AI can help accelerate it. Migrations can move faster when AI handles much of the heavy lifting. AI can be both the outcome of modernization and a tool for advancing it.
Where to start #
The estate will not be finished, so stop asking when the data will be ready and ask what another quarter of waiting costs. The practical starting point is a focused pilot. Prioritize a use case that can demonstrate value quickly and create the organizational credibility to go further.
"Break the Sequence" walks through the architecture behind the leapfrog strategy in more depth, including how AI workloads and modernization share one open, governed foundation so the work done today connects with the work completed later. Download the ebook