What happened
On July 17, SAP completed its acquisition of Prior Labs, a German research lab founded barely 18 months ago, and pledged to invest more than €1 billion over four years to grow it.
Prior Labs does not build chatbots. It builds tabular foundation models — AI made to read the rows and columns of structured data: transactions, customer records, inventory, the tables inside every business system.
The largest business-software company on earth just made a billion-euro bet that the most valuable AI for a company isn't the one that talks — it's the one that understands its spreadsheets.
Prior Labs stays independent, and its open model, TabPFN, has already been downloaded more than three million times and published in the journal Nature — a level of scientific pedigree the chatbot race rarely carries.
The detail almost everyone will miss
The reason this deal exists is a limitation nobody markets: the large language models behind every chatbot are genuinely bad at tables.
An LLM reads a spreadsheet the way you'd read a phone book in a language you don't speak — it sees the shapes, not the meaning, and its grasp of numbers, columns, and statistics is rudimentary at best.
That is fine for drafting an email and dangerous for forecasting demand, scoring a loan, or predicting which machine fails next — the decisions that actually move a business, all of which live in tables.
Tabular foundation models are built for exactly that job. Prior Labs' latest, TabPFN-2.6, tops TabArena, the public benchmark for this category, and makes accurate predictions on structured data straight out of the box, without the months of custom model-building that used to be required.
So the quiet story is a division of labor: language models for words, tabular models for the numbers — and the numbers are where the operating decisions are.
Up close, a table is just cells — but the value hides in one of them: the number that predicts the churn, the fraud, the stockout. A chatbot skims this surface. A model built for the grid reads the cell.
Why this matters if you run a business
Most companies have spent the past two years buying the wrong thing for their highest-value problems. The chatbot was easy to demo, so it got the budget.
But the questions that pay for themselves — who will churn, which invoice won't be paid, how much to reorder — are prediction problems on structured data, not conversation problems.
You almost certainly already own the raw material. It is sitting in your accounting system, your CRM, your point-of-sale export, your maintenance logs — the tables you dismiss as boring because they are.
The strategic tell is who made this bet. SAP runs the back-office systems of a large share of the world's enterprises; it can see, better than anyone, where AI actually earns its keep inside a company — and it just priced that as tables, not talk.
The lesson for an operator isn't "buy SAP's model" — it's that your most valuable AI project may be the unglamorous one aimed at data you already have, while the chatbot everyone's excited about stays a nice-to-have.
The engine room of any business is this: aisles of structured records, humming and unglamorous. The gold light at the end is the point — the value was never in a new place, only in finally reading what's already there.
What to do about it
You don't need a billion euros to act on the same insight. Treat it as a prompt to re-sort your own AI list:
- Separate your "talk" problems from your "predict" problems. Drafting, summarizing, and answering are language jobs. Forecasting, scoring, and ranking are table jobs — and they usually carry the bigger dollar figure.
- Point the next project at a table, not a conversation. Pick one costly prediction — churn, late payment, stockout, no-show — where the history already lives in a spreadsheet, and make that the pilot.
- Tabular models now work off the shelf. The old barrier was that a custom prediction model took a data-science team and months. That barrier is falling — accurate predictions on your own tables no longer require building a model from scratch.
- Fix the tables before the model. The payoff comes from clean, consistent structured data. The least glamorous work — naming, deduping, and connecting your records — is what makes any of this pay.
The chatbot was the demo everyone could see. The money, as the biggest software company on earth just wagered a billion euros to say, was always in the tables you already own.