Claude reads extracted SAP BW metadata and drafts the Fabric lakehouse, pipelines and semantic model that replace it. Instead of six weeks of screenshots and a Word document, the engagement starts with an object-level inventory and reviewable code. Everything it was unsure about is flagged for a human, not silently converted.
A BW landscape that has run for a decade is documented nowhere except inside itself. The logic sits in transformation routines, start and end routines, DTP filters and process chain dependencies, written by people who have mostly left. So the project opens with someone paging through transactions to rebuild a picture the system already holds.
InfoObjects, DSOs, ADSOs, InfoCubes, CompositeProviders, transformations, DTPs, process chains and BEx queries, with no reliable view of what feeds what, and no way to tell which ones anyone still uses.
Start routines, end routines and field routines hold rules nobody can describe without reading them line by line. There is no automated path from that ABAP to PySpark or T-SQL.
Without an object-level inventory, "how many days is this migration" can only be answered as a range. The proposal gets padded, and the client reads the padding as risk.
The accelerator ingests a BW metadata export: object definitions, transformation rules, routine source, DTP and process chain definitions, query definitions. It builds a dependency graph, classifies every object, then drafts the Fabric target on top of that graph. Nothing lands in a client tenant unreviewed, and every generated file names the BW object it came from.
Parses BW object metadata into a lineage graph across InfoObjects, ADSOs, DSOs, InfoCubes, CompositeProviders, transformations and DTPs, so any report field traces back to its source table.
Separates objects that are actually loaded and queried from the dormant, technical and duplicated ones, and produces an object count an estimate can be built on.
Converts BW transformation rules and ABAP routines into PySpark or Spark SQL notebooks, with the source object and rule referenced in the code so a reviewer can check the translation against the original.
Emits a medallion Lakehouse structure, bronze and silver table definitions, and Data Factory pipelines that mirror the process chain execution order and dependencies.
Turns BEx query definitions into a starting Power BI semantic model, mapping measures, hierarchies and variables to DAX where the mapping is unambiguous.
Every artifact carries a confidence level. Non-cumulative key figures, currency and unit conversion, hierarchy logic and heavy ABAP are escalated to a named human review step rather than auto-accepted.
The source is read through metadata extraction rather than live system access, so nothing runs against the production BW box. Claude drives the parsing, classification and code generation as an agent with file access to the extract. Generated artifacts are version-controlled and promoted through Fabric deployment pipelines like any other code.
The fastest first step is a metadata extract from your BW system. We run it through the accelerator and hand back a scoped object inventory plus a sample of generated Fabric code you can judge on its merits.