Most AI budgets this year point in the same direction. Chatbots, copilots, assistants that draft the email and summarize the meeting. Software that talks.
That spending is not wrong. Language models are genuinely useful, and the companies putting them in front of employees are learning real things. But if the talking software is the whole of your AI plan, you have quietly made a decision most boards never debated. You have decided that the highest-value AI in your company is the kind you can have a conversation with.
Your company does not run on conversation. It runs on tables.
Demand forecasts. Churn scores. Credit risk. Inventory that has to be in the right warehouse before the order arrives. The fraud flag that fires in the half second before a payment clears. The decisions that actually move your profit and loss live in rows and columns, and they get made thousands of times a day whether or not anyone is talking.
The category most executives have not heard of
Last month SAP committed more than one billion euros over four years to acquire Prior Labs and build a frontier AI lab in Europe. Prior Labs builds something called tabular foundation models, a category that barely existed two years ago. These are not models that chat. They are pre-trained to make predictions directly on structured business data, the kind that sits in your ERP, your warehouse system, and your spreadsheets.
The most telling part was not the price. It was the reason SAP gave for paying it. Its chief technology officer, Philipp Herzig, put it plainly. The greatest untapped opportunity in enterprise AI, he said, was never large language models. It was AI built for the structured data that runs the world’s businesses.
Read that as a budget statement. The company that sells the systems your operations already live inside just told the market that the most valuable AI may not be the one with a chat box.
Why this is strange, and why it matters
The model behind the deal, TabPFN, was published in Nature in early 2025 and has been downloaded more than three million times. The headline result is worth sitting with. On the datasets the paper studied, a single TabPFN prediction taking 2.8 seconds beat an ensemble of the strongest traditional models that had been tuned for four hours.
The number that matters there is not the accuracy. It is the four hours becoming 2.8 seconds.
For most of the last decade, getting a good prediction out of your business data meant a project. A data scientist collected the data, engineered the features, picked an algorithm, tuned it for days or weeks, and shipped a model that then slowly went stale. A tabular foundation model collapses that. It is not trained on your data and it is not tuned for weeks. It reads your existing table and returns a prediction in seconds, the same way a language model reads a paragraph and returns the next one.
That is the shift. Prediction on your own operational data stops being a multi-month engineering effort and starts being something closer to a lookup.
What it looks like when it works
The early results are concrete, though they come with a caveat I will get to. Prior Labs reports that Hitachi Rail improved anomaly detection by 40 percent for predictive maintenance, the kind of failure-before-it-happens signal that keeps trains running. The retailer Éxito reports that its spend forecasting became more accurate and less manual after older models lost to a tabular foundation model on forecast error. In healthcare, a research group reports 90 percent accuracy identifying cancer patients from a routine blood panel.
Notice what those have in common. None of them is a chatbot. Each is a high-stakes prediction on structured data that a business was already collecting, the exact place a tabular model is built to work.
The honest caveats
This is early, and pretending otherwise would be the same overpromising that makes executives cynical about AI. Three things are worth keeping straight.
First, those customer results are reported by the companies through Prior Labs, not independently audited. Treat them as promising signals, not settled fact.
Second, the canonical Nature result was on smaller datasets, up to ten thousand rows. Newer versions of the model now scale to a million rows, but the category is months old, not years. The tooling around it is still maturing.
Third, none of this means language models were a mistake. The point is not to defund the copilots. It is that you have probably been treating language AI as the entire menu when it is one dish. There is a second category, aimed at the data your business actually turns into money, and almost none of your budget is pointed at it.
What to do on Monday
You do not need to buy anything yet to act on this. The first move is a list. Write down the five predictions that, if they were sharper, would move your numbers the most. Demand. Churn. Credit. Pricing. Whatever fraud or quality signal costs you when it is late.
Then ask one question about each. How is that prediction made today? For most companies the honest answer is a gut feel, a static spreadsheet model from three years ago, or a data-science project that was scoped for six months and never shipped. That list is your real AI opportunity, and it has very little to do with the chat box.
The open-source versions of these models are free to experiment with, which means the cost of finding out is a single afternoon for one of your analysts, not a procurement cycle. This connects to a point I have made before, that the AI strategy that pays this quarter is usually the unglamorous one, and to the deeper argument that the model itself is no longer where the advantage lives. The advantage lives in pointing the right tool at the decisions that already run your business, the same logic that should make you suspicious of importing someone else’s forecast instead of predicting on your own data.
So before the next copilot license renews, one question is worth the room it takes up on the agenda. Are you putting your AI budget where your business actually lives, or where the demo was loudest?
Questions this article gets
What is a tabular foundation model?
A tabular foundation model is a pre-trained AI model built to make predictions directly on structured business data, the rows and columns in your databases, spreadsheets, and ERP, rather than on text. SAP described the category as AI purpose-built for structured data when it committed more than one billion euros to acquire Prior Labs, the maker of the TabPFN model. Where a language model predicts the next word, a tabular model predicts the next value: a demand figure, a churn score, a credit risk.
How is a tabular foundation model different from a chatbot or large language model?
A chatbot or large language model works on language and is the AI you have a conversation with. A tabular foundation model works on your structured data and returns a prediction without a conversation. The practical gap is speed and setup. The TabPFN model published in Nature returned a prediction in 2.8 seconds that beat an ensemble of traditional models tuned for four hours, and it does this without being trained or tuned on your specific dataset. The two are complementary. The point is that most AI budgets fund the talking kind while the highest-value predictions in a business live in tables.
Where should a company start with tabular foundation models?
Start with a list, not a purchase. Write down the five predictions that would move your numbers most if they were sharper, such as demand, churn, credit, pricing, or a fraud signal, then ask how each is made today. Where the answer is a gut feel or a stalled data-science project, you have found a candidate. The open-source versions of these models are free to experiment with, so the cost of a first test is an afternoon for one analyst rather than a procurement cycle.