Philipp Herzig and SAP’s research collaborators argue that fast decisions need not come from a Jev-style text model. Their alternative is a tabular foundation model: a model pretrained to make predictions from rows of structured data, using task examples supplied at inference time.

The article reports two comparisons and a Snake demo. These are SAP’s vendor-run results; I did not reproduce them. The ticket test is the most useful evidence because the same recent examples went to both models.

Recent assignments carry information the ticket does not

SAP took 199 real tickets, supplied the five most recent tickets for each of seven possible assignees, and evaluated predictions for the remaining 164.

  • SAP reports 49.4% assignment accuracy for Jev 1.13.0.
  • Its RPT-1.6 tabular model reportedly reached 68.9%.
  • Adding CASE—text representations that incorporate surrounding table context—reportedly raised that to 72.0%.
  • The majority-label baseline reportedly reached 24.4%.

The point is not that every ticket router should use RPT. Team responsibilities can change without anything in a new ticket announcing the change. Recent assignments can carry that missing organizational context. Both models received those examples; SAP reports RPT used them more effectively.

A broader test is not an identical training setup

On the community Typed-decisions dataset, SAP reports 72.70% for zero-shot Jev, 76.75% for many-shot RPT, and 78.75% with CASE.

That compares deployment approaches with different dependence on examples—not a demonstration that RPT is a universally superior zero-shot classifier. The article does not provide enough split, uncertainty, or per-task detail to make the pooled score a general selection rule.

Do not turn the latency table into a service promise

SAP reports 322 ms for Jev, 44 ms for RPT, and 119 ms with CASE on the ticket task. Its footnote says the RPT-side timings were measured directly on an H200 GPU, without a production deployment.

  • These numbers do not establish matched network overhead, serving cost, throughput, or end-to-end latency.
  • The parallel Snake demonstration is qualitative; the article supplies no controlled win-rate or score table.
  • No calibration study establishes whether either model’s confidence tracks ticket-assignment error.

Typed output is not a correct decision. This is useful evidence for testing an example-driven specialist alongside a general decision model—not proof of safe autonomous assignment. SAP’s own next step is broader workflow testing, harder labels, and real data drift before production use.