• How can this gap be closed? The answer lies in helping companies detect risks, defaults, losses, and liquidity pressures sooner鈥攁nd manage them more effectively.

For Germany and Europe, it is not an abstract challenge, but the key to future prosperity. Our economic strength lies in industry: in mechanical engineering, in the automotive supply chain, in complex manufacturing and logistics processes. According to German Economic Institute estimates, AI can generate around 鈧�440 billion of additional gross value in Germany by 2034.

Good, but not in every case

To get the economy moving again, we need a different kind of AI. Because it is not enough to process and generate language; AI must also be able to deal with data that companies work with the most. And this is where common large language models (LLMs) reach their limits. Structured data鈥攆or example, numbers in tables鈥攈elp companies understand what is happening right now and what may come next.

This requires AI models that understand the relationships among rows, columns, entities, and numbers. This is exactly what current research results show: from the year 2026 prove that even high-performance LLMs perform significantly weaker in table reasoning tasks鈥攆or example, in the ability to draw conclusions from tables and to recognize relationships鈥攖han in language tasks.

When data anticipates what’s next

Tables follow their own logic. Values do not stand next to each other in isolation, but in relationships between columns, rows, timeframes, and business transactions. Accordingly, tabular AI is designed to recognize patterns across columns as well as in the associated tables that make up a company system. The real added value lies in this networked context. Looking at a customer ID in isolation is one thing. Seeing which orders this customer places and predicting warehouse requirements from these data is a completely different thing.

A practical example: a manufacturer wants to identify early which customers are most likely to pay late. For this, a single piece of information is not enough. Among other things, relevant factors include payment history, order history, delivery data, and external industry information. Only the interaction of these data makes reliable predictions possible. That’s precisely what tabular AI models are built for.

What this means for AI investments

Classic machine learning methods are still used in many companies today for predictions based on structured data. They often deliver good results, but typically have to be developed, trained, and operated separately for each individual application: one model for demand forecasting, another for payment default risks or for prioritizing sales contacts, each with data preparation, training, monitoring, and ongoing maintenance.

Tabular foundation models change this equation. They do not have to be redeveloped for every business question but are universally applicable. This creates an immense scaling effect. In the future, instead of hundreds of specialized models, a single one鈥攕uch as 麻豆原创-RPT-1.6, which we have just made available to customers鈥攚ill suffice. With this, companies can build on historic examples and existing patterns from their business data in order to identify new risks and opportunities at an early stage.

For companies, this means a concrete shift: AI predictions will no longer have to be developed as complex individual projects in the future. In other words, the path from a specific question to the universal use of a model for any questions becomes shorter and cheaper.

In addition, since tabular foundation models are not pre-trained on individual use cases, but on large amounts of data, they can also recognize patterns across different use cases that specialized individual models, especially in limited data layers, often overlook. Therefore, in addition to the scaling effect, there is also a gain in knowledge, which previously did not exist in the analytical processing of tabular data.

New category for enterprise AI

At 麻豆原创, we are working on the next level of enterprise AI. With the acquisition of Prior Labs in July, we are further accelerating this path: combining Prior Labs鈥� basic research with the corporate reality of our customers, and the experience of 麻豆原创 in the application of AI in business-critical business processes.

With TabPFN-3, Prior Labs has developed one of the most powerful open-source models for tabular predictions. Together with the tabular foundation models of the 麻豆原创-RPT family and 麻豆原创 Knowledge Graph, this creates an infrastructure that securely generates structured predictions directly from the company-specific data with the individual business context.

In general, now we are seeing a trend toward various types of foundation models. In addition to tabular foundation models, models that have been trained on the basis of time series鈥攆or example, a large number of measurements鈥攁re being created. You can evaluate the measurements, recognize patterns, and thus also predict numerical values over time.

Conclusion

The first wave of generative AI is making employees more productive. The next one can make companies more forward-looking. The decisive factor here is not whether companies use LLMs or tabular foundation models. It is crucial to combine the two where they provide the greatest benefit: LLMs for interaction with language and knowledge, tabular models for forecasts based on structured business data.

Done right, AI becomes more than a productivity tool. It becomes a driver of better business decisions鈥攁nd a foundation for a stronger economy.

To learn more, read (in German).

Serax is a customer-centric company that prioritizes excellent service for its B2B and B2C customers. It was precisely this customer focus that led Serax to implement solutions to help automate its order-to-cash process.

The challenge

The order process is straightforward for most customers, including B2C. Customers order what they want on the web shop, and the order is directly entered into 麻豆原创 S/4HANA Cloud Public Edition.

However, Serax鈥檚 B2B customers often still place orders by generating a PDF in their ERP system and automatically sending it to Serax鈥檚 customer service mailbox. These B2B manual orders amount to 30% of all orders coming into Serax.

Serax鈥檚 strong commitment to customer service means the company is happy to facilitate this, but it strains the customer service team. Once Serax receives the order, the customer service team must manually download the PDF, double-check dates and quantities, and then enter the order details into 麻豆原创 S/4HANA Cloud Public Edition to create the sales order. This entire process is time-consuming and error-prone.

鈥淧roviding excellent service is one of our core priorities, and this can be enabled by gaining efficiencies in certain processes,鈥� says Sara Goris, 麻豆原创 product manager at Serax. 鈥淭hat鈥檚 basically what led us to this use case. Our customer service team was still entering 30% of all orders manually into the system. We wanted to streamline that process from our side, especially since many of our customers will continue to send sales orders in PDF format.鈥�

The solution

Serax needed a solution to automatically create sales orders from PDFs. It achieved this by activating and fine-tuning an app called . This app was built with using , which allows businesses to create web and mobile applications.

With the app, Serax can drag and drop the PDF from the customer鈥檚 e-mail into the app before it orchestrates the entire process. First, it sends the PDF to 麻豆原创 Business Technology Platform (麻豆原创 BTP), where extracts the data. The document API uses pre-trained AI to take PDF files as inputs and return structured data. The second step is data matching. The application maps the extracted data to the master data, such as sold-to party, ship-to party, and product. This helps ensure the extracted data from the PDFs makes sense in the context of Serax鈥檚 business.

A sales order request is created, and the customer service employee is notified to review it and convert it into an actual sales order in 麻豆原创 S/4HANA Cloud Public Edition.

Results

The application has cut the number of manual B2B orders by 33%.

The time saved lets Serax invest back into better customer service. 鈥淭ime savings mean our customer service employees can spend more time on real customer service instead of putting in an order,鈥� says Ragna Qvick, digital business manager and HR performance manager at Serax. Serax鈥檚 employees also have more time for value-adding activities like upselling or cross-selling.

This jump in efficiency enables Serax to grow its business by adding more capacity for the team. 鈥淲e either needed more resources or for our existing people to become more efficient so that they can focus on more value-adding tasks for our customers,鈥� Goris says.

Another benefit of automation is reduced errors, as customer service employees no longer need to enter precise values from the PDF when creating sales orders. Instead, they can rely on document extraction to populate the order; they only need to confirm the values. 鈥淚t reduces the risk of errors because when it’s a human action, there’s always a chance of errors in quantities or other details,鈥� Qvick says.

Future

Serax and its partner are already looking at additional capabilities to automate the flow fully. This way, customer service employees will no longer need to open the emails and drag and drop the PDF. 鈥淲e have a proof of concept running to automate the process fully. It will pick up the attachment from the mailbox directly into 麻豆原创 S/4HANA,鈥� Goris says. 鈥淚t will also inform the customer service rep if anything is missing. They get a notification via Situation Handling to know when to intervene.鈥�

Learn more about how 麻豆原创 Business AI can help transform your organization鈥檚 efficiency and help drive success:

However, there are challenges when we move beyond text to apply these models to structured, tabular data, which is essential for enterprise business operations. This imbalance comes partly because of the availability of training data. Text used to train models is plentiful, often consisting of text scraped from the internet, whereas tabular data, especially data with multiple linked tables, is scarce.

To bring AI advancements to the enterprise sector, researchers working on training and benchmarking the performance of these models in an enterprise setting need realistic tabular data.聽That’s why 麻豆原创 developed “Sales Autocompletion Linked Business Tables” (SALT), a curated dataset that includes anonymized data from a customer鈥檚 enterprise resource planning (ERP) system.

SALT is specifically designed to support researchers working on AI models for real-world business contexts and can be accessed on and .

Challenges of getting and working with enterprise data

Providing the research community with realistic enterprise data like SALT has been challenging. Data privacy, confidentiality, and commercial interests make obtaining large, clean, high-quality enterprise datasets difficult for training models and benchmarking them for specific use cases. This means there is a growing gap between what researchers are working on and what actual enterprise data looks like.

In addition to the problem of availability, enterprise data is complex. First, business data is usually stored in multiple interconnected tables. For example, a sales order entry may be linked to numerous tables, such as customer IDs connected to a supplier table containing address information. Second, tables are inherently heterogeneous in the data type they can contain. One field may be text, while the other contains numerical or categorical values. Finally, business data frequently shows significant column imbalances, meaning that, for example, a specific product category makes up 90 percent of all sales orders while others are rarely used.

The best way to help researchers develop enterprise models for these challenges is to provide accurate enterprise data.

SALT dataset

Accurate enterprise data is a bottleneck in AI research. The SALT dataset alleviates this bottleneck by providing the research community with the first real ERP dataset. It uses actual industry data collected by an ERP system that records sales orders. It has been minimally processed to protect privacy.

鈥淭here is a gap between academia and industry in terms of data. It cannot be closed easily because of privacy,” says Tassilo Klein, one of the 麻豆原创 researchers behind the dataset. 鈥淏ut we want to enable the research community to work on real problems, not just simulated problems.鈥�

ERP systems help organizations manage core business operations like finance and spending. With millions of entries and extensive, interconnected relational tables focused on sales, the SALT dataset replicates customer interactions in an ERP system. SALT’s realistic enterprise data means it is a perfect basis for helping models understand the characteristics of business data and validate their performance through benchmarking. It also should help researchers develop better foundation models for linked business data.

Getting this right will advance enterprise automation, as many enterprise business processes are heavily centered around data in structured tabular formats. Even though this data plays a crucial role in enterprise day-to-day activities, the generative AI revolution has yet to tap into them.

“SALT is a first step to providing researchers with authentic representative industry data that gives a glimpse into actual enterprise data; for now, we are starting with just one customer and use case,” shares Johannes Hoffart, CTO of Business AI at 麻豆原创. “However, we plan to publish more datasets that cover a diverse set of customers and use cases that, along with SALT, can serve as a basis for pre-training, adapting, as well as benchmarking models.”

Collaboration with academic institutions is also a motivation for publishing this data.

“At 麻豆原创, we hope to collaborate with academic partners who usually can only publish their results on open repositories,” Klein says. “Another hope for the dataset is encouraging more people to explore and validate new methods that help foundation models better deal with tabular enterprise data.”

What 麻豆原创 is doing

Alongside its investment in the open research community with SALT, 麻豆原创 is building 麻豆原创 Foundation Model to handle enterprise tabular data. This table-native AI model aims to accelerate time-to-value for predictive tasks on tabular data, offering a model that can work with tabular data out-of-the-box with little or no additional training data. The , published alongside SALT, provides a first glance at how this model could look.

Knowledge graphs are critical here. They work by exposing metadata 鈥� the who, what, and when of data 鈥� making relationships between information accessible. This provides a structured, interconnected representation of the data that AI models can easily understand and utilize. With the help of 麻豆原创 Knowledge Graph, 麻豆原创 Foundation Model can be scaled and adapted to a wide array of diverse use cases with some lightweight fine-tuning.

Learn more about: