Our summary
Routing an incoming invoice is a different job from finding patterns in past invoices. This SAP Community guide explains that distinction: a decision model such as Jev can route one case, while a model trained to read tables estimates outcomes from historical records.
In the invoice example, Jev categorizes the incoming document. TabPFN, a table-reading model, uses past records to estimate what may happen next. The two can work together when each receives the evidence its job needs; neither is interchangeable with the other.
Unfamiliar cases and serious consequences still call for rules or people, whichever approach you use. The article cites a small fraud test from an interested party. It is a starting point for further testing, not evidence of a reliable ranking across business tasks.
Key takeaways
- Ask whether the job is routing one case or predicting an outcome from past records.
- A table-reading model needs historical data, not merely a description of the new case.
- Use rules or human review when cases are unfamiliar or the consequences are serious.