I Tried TabPFN-3.5 Plus in SAP AI Core for EWM Predictions — Scoring Failed

SAP announced TabPFN-3.5 Plus in AI Core as a way to get instant business predictions from tabular data without custom model training — so I tried to use it against real warehouse data, and scoring failed.

Let me walk through what I did, what broke, and what I actually think about this capability after getting my hands on it.

What TabPFN-3.5 Plus Is Supposed to Do

The pitch is genuinely interesting. TabPFN is a pre-trained transformer model that handles tabular data — the kind of structured, row-and-column data that exists everywhere in SAP. No feature engineering pipeline. No weeks of model training. You feed it data, it predicts.

In an EWM context, the obvious applications are things like:

Predicting task completion times for labour management benchmarking
Classifying whether an inbound delivery will be late based on historical yard movements
Scoring warehouse orders by complexity before wave release

SAP wrapping this inside AI Core means the model is theoretically accessible via API from BTP, without standing up your own ML infrastructure. That's the appeal.

What I Actually Tested

I pulled a cleaned extract of warehouse task data — transfer order steps with timestamps, activity areas, handling unit weights, and a binary label for whether the step exceeded its standard time. Nothing exotic. The kind of dataset any EWM system with Labour Management active would produce after a few months.

The dataset was small by ML standards: around 1,400 rows, 11 features. TabPFN was built for exactly this — small-to-medium tabular datasets where training a full model from scratch isn't justified.

I set up the scoring request through AI Core, pointed it at the dataset, configured the deployment, and ran the inference call.

Scoring failed.

What the Failure Actually Looked Like

The error wasn't dramatic. No stack trace. No obvious misconfiguration message. The deployment showed as healthy. The scoring endpoint returned a 500 with a generic internal error response. Logs at the AI Core level were thin — not much surfaced without digging into the underlying infrastructure logs, which aren't immediately accessible without going through BTP cockpit and cross-referencing the AI Core instance logs manually.

A few things I checked:

Payload format — the request schema for TabPFN has specific expectations around how features are passed. I'd followed the documentation, but documentation for a brand-new capability and the actual deployed behaviour don't always match on day one.
Column types — TabPFN handles mixed types, but there are limits on categorical cardinality. One of my columns (storage type codes) had more unique values than I'd noticed. That's a plausible failure point.
Deployment region — AI Core service availability and model catalogue availability don't always align across BTP regions. Worth confirming the model is actually deployed in your specific subaccount region, not just announced globally.

I haven't fully isolated the root cause yet. That's the honest answer.

What This Tells You About Early-Stage SAP AI Capabilities

This isn't a reason to dismiss TabPFN in AI Core. It's a reason to treat it like what it is: a new capability that hasn't been through the same production hardening cycle as, say, Extended Warehouse Management RF framework transactions that have been battle-tested across thousands of sites.

SAP moves fast on AI announcements right now. The gap between "now available" and "reliably production-ready" is real, and it's worth factoring into any project plan where you're promising business stakeholders a working prediction model by a fixed date.

What I'd suggest:

Test scoring first, before building anything around it. Don't design an integration and then discover the endpoint behaves unexpectedly.
Keep your dataset clean and small for initial tests. Reduce variables. Get a successful score on a minimal payload before adding complexity.
Watch the AI Core logs, not just the application response. The 500 you see in your application often hides a more specific error one level down.

What I'll Do Next

I'm going to strip the dataset back to five features and a hundred rows and retry. If that scores successfully, I'll add complexity back incrementally. That's the diagnostic approach — not exciting, but it's how you actually find where things break.

If I get it working, I'll post the request structure and the dataset shape that worked. If it keeps failing, I'll post that too. Either outcome is useful.

The SAP AI Core catalogue is expanding fast, and TabPFN-3.5 Plus is genuinely worth watching for EWM and supply chain use cases where labelled operational data exists but nobody wants to build a full ML pipeline. Just go in with realistic expectations about where new capabilities actually stand on day one.

If you've run into the same scoring error — or got it working — I'd be interested to compare notes. You can reach me through markus-mildner.com.