Kumo's New AI Shatters Benchmarks by Rethinking Enterprise Data

MOUNTAIN VIEW, CA – April 14, 2026 – Predictive AI leader Kumo today launched KumoRFM-2, a foundation model that marks a significant departure from how artificial intelligence interacts with corporate data. The company claims it is the first foundation model to outperform traditional, human-intensive machine learning methods on the complex, multi-table relational data that forms the backbone of modern enterprises.

The new model promises to replace months of specialized data science work with simple, plain-English queries, scaling to handle databases with over 500 billion rows. For businesses drowning in data but struggling to extract predictive insights, the implications are profound. Tasks that once required a team of PhDs, extensive data preparation, and custom-built models can now, according to Kumo, be performed instantly by business users across an organization.

Democratizing the Data Crystal Ball

For years, the power of predictive AI has been locked away, accessible only to companies with deep pockets and elite data science teams. The primary bottleneck has been "feature engineering"—a laborious process where data scientists manually select, clean, and transform data to create features that a machine learning model can understand. KumoRFM-2 aims to demolish this barrier.

"Enterprise data - customer records, transactions, product catalogs - holds enormous untapped revenue potential," said Dr. Vanja Josifovski, Co-Founder and CEO at Kumo. "Until now, using that data to generate business predictions required months of feature engineering and deep data science expertise, putting it out of reach for most teams."

KumoRFM-2 automates this entire pipeline. It requires no task-specific training and allows users to simply ask predictive questions in natural language, such as "Which customers are most likely to churn in the next 90 days?" or "Which sales leads have the highest probability of converting this quarter?"

The real-world impact of this approach is already being validated by major industry players. Databricks, a leading data and AI company, has been using Kumo's platform to refine its lead scoring process.

"Kumo.ai has transformed how we approach lead scoring at Databricks," said Anoop Muraleedharan, Sr Director Data & Analytics, Databricks. "We've seen conversion rates from leads to opportunities improve from 1.2x to 6x, and we've doubled the volume of high-intent, quality leads entering our pipeline. The impact on our marketing performance has been substantial."

This shift empowers a new class of "citizen data scientists," allowing marketing, sales, and operations teams to directly leverage AI for forecasting and decision-making without waiting in a long queue for the data science department.

Beyond Flattening: A New Architecture for Relational Data

The key to KumoRFM-2's performance lies in its novel architecture, which addresses a fundamental flaw in virtually all other predictive models, including large language models (LLMs) and popular algorithms like XGBoost. These tools require that complex, multi-table databases (like customer tables, order tables, and product catalogs) be "flattened" into a single, massive spreadsheet before analysis can begin.

Kumo argues this flattening process destroys the most valuable predictive signals: the intricate relationships between the tables. "For years, AI has been constrained by a fundamental limitation of not being able to reason over structured enterprise data. A database is not a document, it is a graph of relationships," explained Dr. Jure Leskovec, Kumo's Co-Founder and a Stanford professor who pioneered the underlying technology. "KumoRFM-2 is the first model that sees the full graph."

Built on a new Relational Graph Transformer architecture, the model treats a database as an interconnected graph, where rows are nodes and the links between tables (foreign keys) are edges. This allows the AI to "attend to any datapoint," preserving the complete structure and uncovering patterns that are lost during flattening. This approach effectively eliminates what some in the industry call the "flattening tax"—a loss of accuracy incurred when relational context is discarded.

This new architecture, set for publication at the prestigious ICLR 2026 conference, replaces older Graph Neural Networks (GNNs) with a transformer-based approach that can see across the entire data graph, avoiding the information bottlenecks of previous methods. The underlying engine is built for speed, processing data at 5 gigabytes per second and enabling low-latency predictions at production scale.

Setting a New Performance Standard

Kumo has backed its ambitious claims with a slate of impressive benchmark results across 41 predictive tasks. On the SAP SALT enterprise benchmark, which uses real-world ERP data, a fine-tuned KumoRFM-2 achieved an 89% accuracy score (measured in Mean Reciprocal Rank), significantly surpassing established tabular model ensembles like AutoGluon (77%).

On the Stanford RelBenchV1 benchmark, the model outperformed the strongest supervised machine learning model by 5%. Critically, it achieves these state-of-the-art results with extraordinary data efficiency, using as little as 0.2% of the labeled data required by traditional supervised methods.

The model also demonstrates remarkable resilience. Kumo's internal studies show it maintains high accuracy even when 75% of the relational links between tables are removed or when faced with high levels of noisy or missing data. By aggregating information from across the entire data graph, it can effectively "fill in the blanks" where other models would fail. This robustness is crucial for real-world enterprise environments, where data is rarely perfect.

The Race for a Competitive Edge in Enterprise AI

With its unique architecture and proven performance, Kumo.AI is positioning itself as a key player in the next generation of enterprise AI. The company, founded by a powerhouse team of former AI leaders from Airbnb, Pinterest, and LinkedIn, is backed by Sequoia Capital and an all-star list of investors and advisors, including former Snowflake CEO Frank Slootman and Databricks CTO Matei Zaharia.

The model integrates directly with major cloud data warehouses like Snowflake, Databricks, and Spark, allowing it to work where the data already lives rather than requiring costly and complex data migration. This enterprise readiness has already led to production deployments at companies like DoorDash, Coinbase, and Sainsbury's.

The competitive landscape for enterprise AI is fierce, with giants like Google, AWS, and Microsoft offering their own platforms. However, Kumo's focus on natively handling relational data provides a powerful differentiator. The industry appears to be taking notice, with enterprise software giant SAP recently announcing a partnership to integrate KumoRFM into its ecosystem, signaling a broader market shift toward relational-native AI.

By solving the difficult, long-standing problem of making sense of interconnected business data, KumoRFM-2 is not just offering a faster tool; it is proposing a fundamental change in how businesses unlock value from their most critical asset. This new capability allows organizations to move beyond asking what happened and start accurately predicting what will happen next, providing a powerful new lever for gaining a competitive advantage.