Español
Consider a regional bank whose core reporting still runs on SQL Server 2016, a version Microsoft stopped supporting in July 2026. Its leadership has two options: pay for extended security updates and buy time, or treat the deadline as the moment to modernize. The technology team has already made the case for Snowflake. What the executive committee has not yet discussed is which reports the bank should stop producing, which country’s data can live in which region, and who will sign off when the new system’s numbers do not match the old ones.
Those questions decide whether the migration succeeds, and they are increasingly the only hard ones left. Snowflake’s migration agent can now convert legacy code, deploy objects in waves and validate migrated data from SQL Server, Oracle, Teradata, Redshift and PostgreSQL. As the mechanical work shrinks, the risk concentrates in decisions that belong to the CTO, the Chief Data Officer and, increasingly, the Chief Data & AI Officer.
The bottleneck has moved from engineering to leadership
Large technology migrations have a poor track record, and the causes are rarely technical. When McKinsey studied enterprise cloud migrations, it estimated roughly $100 billion in wasted migration spend over three years, with costs running about 14% above budget and 38% of organizations facing delays of more than a quarter. The companies that outperformed, about 15% of those surveyed, shared two traits: executive sponsorship at the top and a complete roadmap before execution began.
Data platform migrations follow the same pattern. Automation reduces the cost of translating what exists, but it cannot judge whether what exists deserves to survive. That judgment falls into five areas, and each one needs an owner at the leadership level.
Know why you are migrating before you decide what to migrate
Enterprises seldom move to Snowflake after an abstract platform evaluation. Something forces the issue: a database reaching end of support, a Teradata or Oracle Exadata renewal, a Hadoop data lake that grows more expensive and harder to staff every year, an acquisition, or an AI program stalled because data is fragmented across subsidiaries.
The trigger shapes everything that follows, and leaders often lose sight of it. A migration driven by an end-of-support date has a fixed deadline and rewards speed. A migration driven by AI requires reworking definitions and governance, which takes longer. When a company runs the second kind of project on the first kind of timeline, it meets the deadline and leaves the original problem intact.
The business case deserves the same discipline. Licenses and on-premise hardware are fixed costs; Snowflake bills by consumption. A credible comparison counts the months when both systems run in parallel, the cost of retraining teams and the savings from retiring what no one uses, alongside the platform price.
Decide deliberately whether to consolidate or coexist
Many banks and large enterprises already run part of their analytics on Databricks, an in-house lakehouse or another cloud. Open table formats such as Apache Iceberg let Snowflake work with data that other engines also use, so replacing everything is no longer the only path.
Consolidation simplifies governance, support and vendor negotiations. Coexistence protects recent investments that work. Either can be right. The costly outcome is coexistence by default, where two platforms calculate the same metrics because no one decided which system is authoritative for each domain. If you choose to coexist, write down which platform owns customer data, which owns risk and which owns finance, and revisit that map every year.
Most of the savings are in what you retire
A warehouse built over fifteen or twenty years accumulates tables with no consumers, reports no one opens and overnight jobs designed around performance limits that no longer exist. Migration tools convert all of it as faithfully as they convert what matters.
The most valuable early step is therefore an inventory checked against actual usage, followed by a blunt classification. Stable, well-modeled workloads move as they are. Components that are flawed by design, such as business logic buried in stored procedures or metrics calculated differently by different departments, get redesigned. Everything else is retired. Every retired object is one less thing to convert, test, secure and pay for, and this step often generates more resistance than any technical task, because each report has a sponsor who believes it is essential.
Data residency is a risk committee decision
For a financial institution, where data physically lives can outweigh every other architectural choice. Snowflake currently operates two regions in Latin America: AWS in São Paulo and Azure in Mexico Central. A bank operating in Argentina, Colombia or Chile has to confirm that its regulators and internal policies allow data to be hosted abroad, decide what may be replicated between regions and document the reasoning for auditors.
Access raises parallel questions inside the group. Subsidiaries in different countries should not all see the same data, and development environments should never hold real customer information. The platform provides the controls, including role-based access and dynamic data masking. The policies those controls enforce have to come from security, compliance and domain owners, and they need to exist before the first table moves. Policies written after go-live usually mean someone has already seen data they should not have.
Validation is a business sign-off
Automated validation compares row counts, totals and schemas, and it does that well. It cannot tell you whether the regulatory report, the delinquency ratio or a customer’s consolidated exposure produces the same answer as before, or whether a difference is an error or a long-standing inconsistency that the migration has finally exposed.
That second question is harder than it looks, even for the world’s largest banks. More than a decade after the Basel Committee issued its principles for risk data aggregation, its 2023 progress review found that only two of 31 global systemically important banks fully complied. Authoritative, reconciled risk data remains an unfinished job across the industry, and a migration either advances it or copies the gaps.
Business validation therefore needs named owners, acceptance criteria agreed before testing begins and a parallel run long enough to cover at least one full reporting cycle. In regulated domains, go-live should happen domain by domain, with a rollback plan for each. Starting with a contained, high-value area such as risk or finance lets the team refine its method before scaling it.
Consumption pricing needs owners
Pay-as-you-go pricing gives teams flexibility and gives the organization a new responsibility. An oversized compute cluster, a workload that never shuts down or a team running inefficient queries starts costing money on the first day.
Cost control has to be designed with the architecture: separate compute for separate workloads, automatic suspension, spending caps and tagging that assigns costs to departments. What your CFO needs is predictability, and predictability comes less from the unit price than from a governance model in which specific people are accountable for specific budgets.
AI readiness is decided during the migration
In financial services, migrations are rarely justified by reporting alone. Behind them sit credit scoring and risk models, fraud detection, anti-money-laundering analytics, a single view of the customer across products and countries, and secure data sharing with fintech partners.
Each of those depends on the decisions above. Snowflake’s AI agents answer questions by reading semantic definitions of business metrics, so an ambiguous definition of an active customer or a past-due loan becomes a confident wrong answer in plain English. Access policies determine what an agent may retrieve on a user’s behalf. And without consumption governance, the cost of compute-intensive AI workloads is impossible to forecast. An organization that makes these choices well while migrating starts its AI work with an advantage that is hard to recover later.
What this looks like in practice
When Samsonite set out to unify data across its Latin American operation, the challenge was less about converting a legacy warehouse than about agreeing on a common model. The project covered seven markets, eight brands, more than 350 physical stores and 20 online stores, with data spread across Shopify, Google Analytics 4, Oracle OMS, LINX, Retail Pro Prism, JasperX PIM and Gorgias, each structured differently.
The decisive choices were organizational. The team defined layers for raw, standardized and business-ready data, so every market would read the same definitions. Governance was built in from the start, with masking in development, role-based access policies in production and an inventory of sensitive columns. The technical build, orchestrated with Apache Airflow on AWS and transformed through eight dbt projects feeding more than 15 automated pipelines, took roughly four months from design to implementation. The result was a unified view of all seven markets and daily transactional data for marketing.
Five questions to ask before you approve the budget
Leaders who sponsor a Snowflake migration can test its readiness with five questions:
- What specific problem does this migration solve, and does the timeline match that problem?
- For each data domain, which platform will be authoritative, and who decided?
- What percentage of current objects will be retired, and who approved the list?
- Where will each country’s data be hosted, and has compliance signed off on that choice?
- Who certifies that the new numbers are correct, and who owns the consumption budget?
If the answers are vague, the project is not yet ready to start, however advanced the tooling. The automation available today makes the mechanical side of migration faster and safer. It also removes the last excuse for leaving the important decisions to the engineers.
Exomindset is a Snowflake Select Partner, Matillion Silver Partner and AWS Partner, certified under ISO 9001 and ISO/IEC 42001, the international standard for AI management systems. Its data and AI teams help enterprises across Latin America plan and execute Snowflake migrations alongside their internal teams.