A large language model predicts the next word. It drafts, summarizes, searches, and answers in prose. The records a company runs on sit in tables: orders, payments, stock, deliveries. On 19 October 2026, Maarten Rikken argued on SAP News Center that the next wave of generative AI has to use both, and that the gap between personal productivity and operating profit sits in those tables.
This piece covers that argument, which tabular models are available in SAP AI Core, and the question a program owner can ask this week. It leaves time-series models at the one sentence the article gives them. It leaves out API payloads.
What is available
Two tabular models are available to SAP customers now. The 19 October article names the models that already do this work.
SAP RPT-1 is SAP's tabular foundation model for structured business data. The version the article names, and the version current SAP AI Core posts call, is SAP RPT-1.6. Rikken writes that SAP has made it available to customers. A 21 September community post describes calling it on SAP AI Core for classification and regression across several columns at once. A 21 October post describes that same stock model, and names delivery delay and payment date as questions the call answers. SAP AI Core is the SAP BTP (SAP Business Technology Platform) service that runs these models.
TabPFN-3.5 Plus, from Prior Labs, has been available in SAP AI Core since 15 September 2026. Prior Labs is an SAP company; SAP completed the acquisition on 17 July 2026. The September announcement says TabPFN-3.5 Plus predicts from structured data without model training or tuning. Named uses: cash-flow forecasting, payment delays, supplier risk, upsell, and customer churn. SAP calls it the most accurate and scalable tabular foundation model available that day, citing TabArena and BeyondArena. The post does not print the scores.
The 19 October article also names TabPFN-3, Prior Labs' open-source model. The build SAP put in SAP AI Core on 15 September is TabPFN-3.5 Plus.
Background
Rikken cites a McKinsey study: 80% of people who use AI report higher personal productivity, and 6% of companies say AI significantly raised operating profit. He cites the German Economic Institute: AI can add about €440 billion of gross value in Germany by 2034, in mechanical engineering, automotive supply, manufacturing, and logistics. Those figures are his citations.
Generative AI, as deployed so far, helps a person write, research, program, or review faster. Detecting a risk, a default, a loss, or a liquidity squeeze earlier runs on the structured data companies already keep. He writes that independent benchmarks in 2026 find even strong language models weaker at drawing conclusions from tables than at language tasks.
What companies run today, in his account, is classic machine learning. One model for demand. One for who defaults. One for which sales contact to call first. Each model can be accurate. Each one needs its own data preparation, training, monitoring, and upkeep. The cost is a project per question.
A tabular foundation model is pre-trained on large amounts of structured data so the same model applies to many questions. In practice that means the forecast can stop being a separate project. It becomes a table of labeled history, plus the rows that still need an answer.
How it works
A customer number on its own is an identifier. The same number next to that customer's orders, the delivery dates on those orders, and the payments that followed is a relationship.
Option A is the language model. Rikken assigns it language and knowledge: the draft, the summary, the policy answer in words. It can write the note that follows a late payment. The number in that note is Option B.
Option B is the tabular model. It takes rows where the outcome is already known and fills the outcome on rows where it is missing.
TabPFN-3.5 does this by in-context learning: the examples travel with the question, and the model predicts from those rows. The 15 September announcement says it accepts the table as it sits in the system. Missing values, mixed types, inconsistent fields, and columns with thousands of distinct values, such as product codes or customer identifiers, stay in the request. Philipp Herzig, chief technology officer at SAP, said customers get predictions from labeled business data in minutes, with no training required. He called tabular AI the foundation of the Autonomous Enterprise. That name is SAP's destination: people set direction, and AI executes. The models a team can call are SAP RPT-1.6 and TabPFN-3.5 Plus.
SAP RPT-1.6 uses the same shape of request. The 21 October post describes it on the stock SAP AI Core deployment: some cells are marked for prediction, and the model fills them from the other rows, the examples, with the weights left unchanged. The 21 September post shows that shape on a sales-order table, several fields at once. Its example fields are sales group, shipping point, payment terms, and Incoterms, the standard codes for who pays freight and when risk passes. In practice that means one request can fill related fields that four separately trained models might answer differently.
Rikken's example is a manufacturer that wants to see who will pay late. Payment history alone is a thin signal. Payment history, order history, delivery data, and external industry information together are what he says a reliable score needs.
He places SAP Knowledge Graph beside the two model families. SAP Knowledge Graph is SAP's semantic network of business objects: which customer sits on which contract, which material sits on which order. The article says the graph, the SAP RPT family, and Prior Labs' models together produce structured predictions from a company's own data, with that context.
He also notes foundation models trained on time series, a long run of measurements, used to forecast a number forward. He names no SAP product for that category, and he gives no date.
Why it matters for SAP leaders
- If the objective is a copilot that writes and searches, the 19 October article assigns that job to the language model. The 80% productivity figure is the result that kind of tool is cited for.
- If the objective is an earlier signal on cash, default, or demand, the article's claim is that the signal lives in structured data, and that a tabular foundation model is the model built to read it. The 6% operating-profit figure is the gap that claim is aimed at.
- If the company already runs SAP AI Core, both SAP RPT-1.6 and TabPFN-3.5 Plus are available to request. The 15 September announcement sets TabPFN-3.5 and the SAP RPT family side by side: predictions from labeled data, no training step.
- If each forecast is still its own project, with its own training run, the shift he describes is those projects sitting on one foundation model. The posts leave the choice of table, label, and who may act on the score with the owner.
- If a roadmap item says to point a language model at the ledger and ask it for the forecast, the 19 October piece separates the two jobs by model name and by date. Language for the interaction. A tabular model, already available, for the number.
Near term
Both models are available on SAP AI Core now. A useful trial is one forecast whose history is already labeled, such as payment delay, supplier risk, or demand, before the next training cycle is funded.
Longer term
His scaling claim looks forward: one model, such as SAP RPT-1.6, in place of hundreds of specialized models, and patterns a narrow model misses. Neither news post counts models a customer has retired, or states a saving.
The sum SAP has published is the lab. More than €1 billion over four years. Prior Labs stays independent. The July announcement says the work started with SAP RPT-1.
Open questions
The 19 October article and the 15 September announcement leave the choice between SAP RPT-1.6 and TabPFN-3.5 Plus with the reader. No decision rule is published.
The benchmark claims are named, 2026 comparisons of language models on tables, plus TabArena and BeyondArena for TabPFN-3.5 Plus, and unscored in the news posts.
These sources describe a score. They do not grant permission to release a credit block, change a limit, or move a shipment. That permission stays with the process owner.
The €440 billion figure is a national estimate for Germany through 2034. A single SAP landscape still needs its own numbers.
What to do this week
Ask the owner of one forecast three things. Which table holds the history. Whether that history is labeled. Whether the next step is a call to SAP RPT-1.6 or TabPFN-3.5 Plus on SAP AI Core, or another training project.
If short explanations like this are useful, follow along.
Sources
- The Gap That No LLM Can Close (SAP News Center, Maarten Rikken, 19 October 2026): https://news.sap.com/2026/10/gap-no-llm-can-close/
- TabPFN-3.5 Plus Now Available in SAP AI Core for Instant Business Predictions (SAP News Center, 15 September 2026): https://news.sap.com/2026/09/tabpfn-35-plus-now-available-sap-ai-core-instant-business-predictions/
- SAP Completes Acquisition of Prior Labs (SAP News Center, 17 July 2026): https://news.sap.com/2026/07/sap-completes-prior-labs-acquisition/
- I taught SAP RPT-1.6 to play Tetris (SAP Community, 21 October 2026): https://community.sap.com/t5/technology-blog-posts-by-sap/i-taught-sap-rpt-1-6-to-play-tetris/ba-p/14507821
- 3 Challenges I Encountered When Calling SAP RPT-1.6 via the Low-Level Python SDK (SAP Community, 21 September 2026): https://community.sap.com/t5/technology-blog-posts-by-sap/3-undocumented-bugs-in-the-sap-ai-core-python-sdk-for-rpt-1-6-predictions/ba-p/14487970