Thousands of tables and data objects required engineers to trace dependencies and rebuild established patterns. CLOUDSUFI introduced task-specific agents for lineage and Bronze-layer development, giving teams structured outputs to review while retaining specialist validation.
Swarovski’s data modernization programme involves migrating thousands of tables and data objects from a legacy enterprise data warehouse to a modern cloud data platform.
Each group of data moves through several stages: analysis of the existing environment, Bronze-layer development, data modelling and transformation, and Silver-layer implementation. Within a medallion architecture, the Silver layer is where source data is cleaned, validated, and converted into consistent structures using established names, data types, transformation rules, relationships, and source-to-target mappings.
As the migration progressed, the same types of analysis, mapping, development, and validation work had to be repeated across a growing number of objects. The scale created an opportunity to use AI for defined, rules-based activities while retaining specialist review for business context, exceptions, and complex transformation logic.
Swarovski was not simply transferring data. Every group of objects required teams to understand existing technical logic, trace dependencies, rebuild transformation rules, validate outputs, and prepare data for standardization. CLOUDSUFI and Swarovski focused on reducing repeated work without removing the judgement required for complex cases. The programme needed repeatable outputs, visible review points, and a practical way to compare agent-assisted effort with established manual estimates.
CLOUDSUFI did not attempt to automate the complete migration from end to end. The team selected activities that became repetitive after their rules, inputs, and expected outputs had been established.
Data engineers worked with the AI team to document how each activity was performed, what a correct output should contain, and where exceptions usually appeared. Each agent could then be designed around a specific task instead of a broad, open-ended objective — establish the process, identify repetition, define boundaries, build the agent, compare effort, and retain review.
“Any automation introduced into the migration had to work within our existing environment and controls. The agents gave our teams a more consistent starting point for repeatable work while keeping review and decision-making with the people who understood the data.”
Fabrizio AntonelliVP and Global Head of Data and AI, Swarovski
The implementation began with two capabilities that could be incorporated into active migration work and evaluated against real delivery activities. LineageGPT is in production and being used to analyze selected technical documentation and generate source-to-target lineage. Interactive and tabular views give engineers a structured starting point for review, refinement, mapping, and impact analysis. The Data Transformation Validation Agent supports repeatable Dataform development in the Bronze layer using established project patterns. Generated outputs remain subject to engineering validation, with complex and unsupported cases handled by specialists.
Deployment was incremental and depended on the data, access, and security requirements of each activity.
The agent-assisted workflow changed how eligible activities were completed. Teams could begin with structured lineage or development outputs and direct more of their time toward validation, exceptions, and complex logic.
For a simple Bronze-layer Dataform development object, agent-assisted effort was approximately 20% lower than the established manual estimate, decreasing from approximately 0.75 to 0.60 person-days. The result is specific to a simple development object and should not be interpreted as a uniform improvement across every migration activity. Across a migration involving thousands of tables and data objects, reducing effort on eligible recurring activities can improve throughput without increasing manual work at the same rate.
LineageGPT enabled teams to generate source-to-target lineage from selected documentation. Interactive and tabular views made relationships easier to review, refine, and use during mapping, development, and impact analysis. By improving repeatable Bronze-layer development, the team also created a more consistent foundation for downstream Silver-layer cleansing, mapping, and standardization.
The value extended beyond the time saved on an individual object. Once the rules for an eligible activity were defined, teams no longer needed to recreate the same work for every table or object. Structured lineage and development outputs provided a consistent starting point, while engineers remained responsible for review and for the cases that required deeper technical or business judgement — greater scalability, more consistent execution, stronger traceability, better use of expertise, and a clearer path to Silver.
“At this scale, improvements in the effort required for each object compound across the programme. Reducing repeatable development work and structuring lineage gave our engineers more time for complex transformations, while supporting a more consistent path to the Silver layer.”
Fabrizio AntonelliVP and Global Head of Data and AI, Swarovski
For Swarovski, applying AI within the migration meant introducing automation that could work within the existing data environment, security requirements, and review processes. Focused agents supported defined work, outputs remained reviewable, and human accountability stayed with the engineering and business teams.
Swarovski and CLOUDSUFI created an agent-assisted migration model built around reuse, traceability, and human control.
Let's build what's next.
Talk to us about your data and AI challenges — and how CLOUDSUFI can help solve them.
Talk to us →By submitting, you consent to CLOUDSUFI processing your information in accordance with our Privacy Policy. We take your privacy seriously; opt out of email updates at any time.
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
