Process Mining works on event logs: one row per activity execution with case id, activity, start / end time, resource. BAW has that data in three places:
- BAI - the cleanest source on CP4BA: every process, activity and task event is streamed; the process events can be exported (from the BAI indexes) or read from Kafka into a CSV / the Process Mining connector.
- Process REST API - IBM's open-source BAW accelerator for Process Mining pulls instances and tasks through the REST API and produces the CSV in the expected shape; good for traditional BAW without BAI.
- PDW / tracking data - the Performance Data Warehouse tables (or your own tracking groups) with custom SQL - flexible, but you build the log yourself.
# event log columns Process Mining expects (CSV), one row per task execution
caseId,activity,startTime,endTime,resource,processName,attribute_amount,attribute_region
2072.345,Enter order,2025-03-01T09:00:00Z,2025-03-01T09:06:00Z,jane,Order handling,12500,DE
2072.345,Approve order,2025-03-01T09:06:00Z,2025-03-02T14:30:00Z,mark,Order handling,12500,DE
# accelerator (github IBM/ibm-process-mining-BAW-accelerator): configure BAW URL, credentials, process app and date range -> CSV -> upload to Process Mining
What it delivers: the actual process map with frequencies and durations (rework loops, skipped approvals), conformance against your reference BPMN (imported), the variants and their costs, bottleneck and waiting-time analysis per team, and simulation ("what if approvals under 5,000 are automated"). Task mining adds desktop recordings for the steps outside BAW.
Back into Workflow: Process Mining exports the discovered or improved model as BPMN 2.0; Workflow Center imports BPMN files (as a process app or into an existing one), giving a BPD skeleton with the activities and flow - you then add the implementation (services, coaches, teams). Keep the round trip pragmatic: the mined model documents reality and drives the change; the BPD is re-authored, not generated, for anything beyond the skeleton.
References