Context
Anyone migrating from SAP BW to Datasphere rarely underestimates the data volume; they underestimate the logic. A mature BW holds transformations that grew over years: rules, routines, special cases, documented at best in the code itself. This logic is the real value of the system, and it is the reason migration projects get expensive.
Transferring it by hand means: read every transformation, understand it, rebuild it in the target model, test it. That is time-consuming, error-prone and does not scale at all: the effort grows linearly with the number of objects.
Approach
Instead of translating what a human laboriously reads, the application reads the existing logic itself and transfers it into the Datasphere model with AI support. The AI component runs on AWS Bedrock; it is deliberately one building block of the application, not its center. Before and after it come extraction, verification and traceability.
The rest is the pattern all three products share: SAPUI5 in the frontend, SAP CAP in the backend, everything on SAP BTP, operable as SaaS.
How it is built
The application connects to the existing BW and reads the transformations where they sit: field mappings, formulas, routines, lookup tables. Every object runs through the same path:
- Read the logic out and break it into its parts: fields, rules, dependencies
- Translation by the AI into the target model of Datasphere
- Review: original and result side by side, every deviation marked
- Approval by a human
- Creation in the target system
The AI never gets the whole system, always one logic together with its target structure. That keeps the translation traceable and the review small.
| Component | Technology | Decision |
|---|---|---|
| Extraction | SAP BW via SAP BTP | The logic is read, not retyped; the source system stays unchanged |
| Translation | AWS Bedrock | One object per run, with the target structure as the specification |
| Review | SAPUI5 | Deviations are marked, nothing is adopted silently |
| Target | SAP Datasphere | Objects are created only after approval |
| Platform | SAP BTP, SAP CAP | The same pattern as the other products, operable as SaaS |
Learnings
- The AI carries rules, not side effects. It translates field mappings, formulas and lookups reliably. Routines that change other data along the way belong in the hands of a human.
- The review is the product, not the translation. Without marked deviations someone would have to read both sides in full. With them he reads only the differences.
- The effort no longer grows with the number of objects. The hundredth object costs as much as the tenth: one review.