This model marks the next leap in tabular AI, giving organizations access to unmatched tabular AI predictions on structured business data without model training or tuning. For 麻豆原创 customers, TabPFN-3.5 Plus is now available in 麻豆原创 AI Core.

Critical enterprise business decisions,聽such as cash flow forecasting, payment delays or聽supplier risk scoring, run on structured, tabular data. While LLMs聽excel at language and knowledge,鈥痶abular foundation models聽(TFMs)聽are purpose-built for聽structured data and聽can accurately predict business outcomes based on tabular data such as payment delays, supplier risks, upsell opportunities, customer churn risk and more.聽

With TabPFN-3.5, 麻豆原创 customers can work with data as it exists in their systems. Missing values, mixed data types and inconsistent fields are handled by the model, returning聽accurate聽predictions without preprocessing.聽This聽newest model from Prior Labs鈥� family of tabular AI models聽uses in-context learning to make predictions from raw tabular data, handling columns with thousands of distinct values, such as product codes or customer identifiers, and mixed data types natively, without the trial-and-error configuration that traditional models require.聽TabPFN-3.5 Plus聽is the most聽accurate聽and scalable tabular foundation model available today, based聽on聽TabArena聽and聽BeyondArena, two聽external聽benchmarks designed to evaluate predictions over real-world data.聽

鈥淩eal-world data is rarely perfect.聽Datasets often聽contain聽complex relationships and varying conditions that make traditional machine learning difficult. TabPFN-3.5 is聽specifically聽built for these challenges, providing industry-leading accuracy and scalability for tabular data with less manual effort鈥攎aking it the聽most effective聽tabular AI model in the industry聽to date,鈥澛爏aid Philipp Herzig, Chief Technology Officer, 麻豆原创聽SE.聽鈥淲ith TabPFN-3.5聽and the 麻豆原创-RPT model family, customers get聽accurate聽predictions from labeled business data in minutes, with no training聽required.聽For us,聽tabular AI is not a supporting feature of the Autonomous Enterprise, it is the foundation.鈥澛�

For more detailed information and to learn how to get started on 麻豆原创 AI Core, see the , and view the .

麻豆原创 completed its acquisition of Prior Labs in July 2026, bringing one of the world鈥檚 leading TFM research teams into the 麻豆原创 family. Prior Labs will continue to operate as an independent entity, with 麻豆原创 having previously committed to invest more than 鈧�1 billion to scale it into a globally leading frontier AI lab for the structured data that underpins the world鈥檚 businesses.

Visit the . Get 麻豆原创 news via  and .

Media Contact:
Alex Vaught, +1 (206) 678-5712, alex.vaught@sap.com, PST 
; press@sap.com

This document contains forward-looking statements, which are predictions, projections, or other statements about future events. These statements are based on current expectations, forecasts, and assumptions that are subject to risks and uncertainties that could cause actual results and outcomes to materially differ.鈥� Additional information regarding these risks and uncertainties may be found in our filings with the Securities and Exchange Commission, including but not limited to the risk factors section of 麻豆原创鈥檚 2025 Annual Report on Form 20-F. 
漏 2026 麻豆原创 SE. All rights reserved.
麻豆原创 and other 麻豆原创 products and services mentioned herein as well as their respective logos are trademarks or registered trademarks of 麻豆原创 SE in Germany and other countries. Please see  for additional trademark information and notices.

The top team鈥檚 rowers focused entirely on maximizing individual power. If the boat slowed, they rowed harder in isolation, disrupting the oars鈥� synchronized rhythm and creating water drag. Meanwhile, the junior rowers knew they were individually weaker, so they rowed in harmony.

Enterprises have faced countless variations of this problem: implementing systems that maximize productivity at the individual or team level but actively hinder the wider enterprise. Many organizations are experiencing something similar with AI today.

AI and sub-optimization

Sub-optimization is a systemic failure that occurs when the performance of a specific part of a system is maximized, inadvertently hampering the performance of the entire system. There are three intertwined themes: intensity, context, and prediction, which, taken together, explain how AI can sub-optimize an organization by making individuals and local systems stronger while straining the broader organization.

Research reinforces this disconnect between individual or even company-wide AI adoption and the value it delivers. shows near-universal enterprise AI adoption: 89% of respondents report regular AI use in at least one business function, but only 37% report an earnings before interest and taxes (EBIT) impact from AI at the enterprise level. Even worse: only six percent of companies can be categorized as high performers that already capture significant organization-wide value from AI.

What makes it so hard to move from AI adoption to measurable value capture? I believe there are three themes that influence a company鈥檚 ability to benefit its entire organization.

Intensity

AI tools often don鈥檛 reduce work; they intensify it. A found that employees who heavily use AI worked faster, took on a broader range of tasks, and worked longer hours, often without being asked. So, what appears to be higher productivity in the short run is actually silent workload creep and mounting pressure as employees manage new AI workflows and do more with less. that the most mentally taxing form of AI engagement was oversight; AI tools that require direct monitoring increased feelings of being overwhelmed by the volume of information at work.

It is easy to see why AI can feel intense: tasks that once required days can now be prompted into existence almost immediately. People start more things; they do more analysis and write more memos. However, like an eight-lane highway that suddenly narrows to a single-lane toll booth, individuals must still consume all this output. This bottleneck only compounds at the organizational level, as all employees produce more than ever, leaving both individuals and the organization as a whole struggling to keep up. Creation has scaled. Absorption has not.

The solution isn鈥檛 necessarily to use less AI, but to change where and how AI shows up. AI should understand user intent and surface the insights needed to answer the question, rather than generating static assets or requiring you to switch between different apps and systems.

If your question creates more things, it鈥檚 not helping absorption. 麻豆原创鈥檚 answer is , a central workspace across 麻豆原创 and non-麻豆原创 systems that uses AI agents to handle tasks.

Ask, 鈥淲hich stores run out of 65鈥慽nch TVs in the next 72 hours, and where is stock I can move?鈥� It will pull data across systems and orchestrate agents to act on the user鈥檚 behalf. In this case, 麻豆原创鈥檚 answer is autonomous action combined with a highly individual user experience for that specific situation, not more assets to be absorbed. If employees can avoid juggling systems and consuming assets, they can spend more time exercising judgment on actions that matter. This is how AI can alleviate intensity.

Context

Most AI systems understand the world, but not the enterprise in which they operate. There is a difference between a system of record鈥攖ransactions, master data, process logic鈥攁nd tacit knowledge鈥攅mails, chats, unwritten rules. And enterprises run on both. If AI only sees the system of record, its answers might be technically correct but contextually wrong because they don鈥檛 reflect the organization’s lived practice.

Even the most ostensibly basic questions require company context. Asking 鈥淲hich suppliers can I source coconuts from?鈥� requires knowledge of an organization鈥檚 process landscape across procurement, supply chain, compliance, finance, and other domains. This type of enterprise knowledge is usually scattered across process models, policies, chats, spreadsheets, and applications, so it鈥檚 tough to maintain. And even if they find it, agents cannot turn it into action without procedural knowledge of the involved people鈥攖he unwritten rules, decisions, and steps鈥攖hat make a process executable.

preview continuously captures institutional knowledge and makes it usable for both people and agents. It turns written inputs, chat inputs, process knowledge, policy guidance, and application logic into reusable building blocks that AI agents can consume. Blocks are captured once, governed centrally, and reused across the company. So when someone asks Joule Work about coconuts, the answer is driven by the company鈥檚 memory and reflects actual rules the process owners agreed upon鈥攆or example: “Only source from Brazil; others require formal exception approval.”

麻豆原创 Company Memory is not a one鈥憈ime implementation; it鈥檚 continuous. In this way, company knowledge behaves like infrastructure, ensuring agents act contextually, not just correctly, as policies and teams change.

Prediction

Business decisions are fundamentally prediction problems that rely on structured data. Most organizations use LLMs, which are great at unstructured data like text but for architectural reasons not so great at working with and generating the structured numerical data that underpins good predictions. Delay prediction, forecasting, anomaly detection, stock optimization, and credit risk are everyday operating questions that depend on structured, tabular data and forward-looking judgment.

Asking LLMs for reliable forecasts on enterprise tables is simply the wrong tool for the job. At the same time, traditional custom machine learning approaches are too slow for many real-time questions: after extracting data, sending it to specialists, and waiting weeks, the question often has changed by the time the answer comes back. This combination means predictive capabilities are either restricted to specialists or rendered inaccurate by generic LLMs; in either case, the organization鈥檚 decision-making is weakened.

Reliable forecasting and risk assessment should be a system property, not an individual hack. 麻豆原创-RPT-1.5 and TabPFN 3 are models that excel with tabular data and will integrate with Joule Work and 麻豆原创 Business Data Cloud for forward-looking questions directly on live tables.

麻豆原创-RPT-1.5 for 麻豆原创 data and TabPFN 3 for any tabular data are specialized prediction engines for structured data. They enable decision-makers working in the core systems to ask, 鈥淪hould I reroute volume? What鈥檚 the probability of on鈥憈ime delivery? What鈥檚 the cost delta across scenarios?鈥� and get answers grounded in real enterprise data.

Availability across the organization eliminates specialist bottlenecks and better equips the enterprise to handle uncertainty through prediction. This enables informed top-level decisions that really move the needle for a company.

AI for the benefit of the whole organization

AI has already proven it can make people more capable, but that does not automatically help the wider organization. AI shouldn鈥檛 be about optimizing isolated tasks; it should be about reshaping how work, knowledge, and decisions flow through the company.

Joule Work, 麻豆原创 Company Memory, and 麻豆原创-RPT-1.5/TabPFN 3 are great examples of how 麻豆原创 designs system-level capabilities. They offer a unified engagement layer, a living institutional memory, and a prediction engine for structured business data that elevate AI from individual-level hacks into a collective benefit for the enterprise.

This is AI that bridges the individual-to-institutional value gap, moving from simply getting AI into the company to generating value throughout the company.

Florian Kunzke is global director of AI Strategy at 麻豆原创.