Kumo AI: Predict Anything on Relational Data

Every few years, a product comes along that genuinely forces you to rethink how something is done. Kumo AI is one of those products. Built by a team of PhDs who cut their teeth at Pinterest, Airbnb, LinkedIn, and Stanford, it tackles a problem that has quietly frustrated data teams for decades: making accurate predictions from relational data without spending months on feature engineering.

This review covers everything a data leader, ML engineer, or technical decision-maker needs to know before evaluating Kumo — what it actually does, how KumoRFM works under the hood, where it outperforms traditional approaches, and how it stacks up against well-funded competitors like Alteryx, Pecan, and Dataiku.

Key Takeaway: Kumo.AI eliminates the feature engineering bottleneck entirely by treating relational databases as graphs — delivering predictions in under one second with accuracy that rivals what a senior data scientist could produce in a month of manual work.

Quick Stats at a Glance

What Is Kumo.AI?

Kumo.AI is a SaaS predictive AI platform founded in 2021. Its core mission is deceptively simple: let any organization make accurate predictions directly from the structured, relational data already sitting in their data warehouse — without building separate ML pipelines for every use case.

Think about the kind of data most companies actually have. Customer tables, transaction histories, product catalogs, supply chain logs, financial ledgers. This data lives in relational databases, connected through primary and foreign keys. For years, getting useful predictions out of this data meant flattening it, hand-crafting features, running it through XGBoost or LightGBM, and hoping the feature set was good enough. That process consumed months of engineering time, and it started over from scratch every time the business asked a new question.

Kumo's founders — Vanja Josifovski (CEO, ex-Pinterest), Jure Leskovec (Chief Scientist, Stanford professor), and Matthias Fey (Chief Architect) — identified this gap early. Their answer was to apply Graph Neural Networks directly to relational data, treating rows as nodes and key relationships as edges. The result is a platform that learns from the full relational structure without manual feature engineering.

If you're currently evaluating which AI infrastructure partners make sense for your team's roadmap, this guide to choosing a machine learning partner in 2026 covers the vendor criteria that matter most before signing any contract.

The Introduction of KumoRFM

In May 2025, Kumo took a significant leap forward with the launch of KumoRFM — a Relational Foundation Model. Where earlier versions of Kumo still required task-specific model training (though much faster than traditional approaches), KumoRFM operates like GPT does for text: it delivers zero-shot predictions without requiring any training at all.

A user connects their Snowflake or Databricks warehouse, describes the prediction they want in a SQL-like language called PQL (Predictive Query Language), and gets results in under one second. No pipelines. No model training. No feature engineering. It is genuinely a different paradigm.

sql

-- Predict which customers will churn in the next 90 days
PREDICT COUNT(transactions.*, 0, 90, days) = 0
FOR EACH customers.customer_id
WHERE COUNT(transactions.*, -60, 0, days) > 0

That three-line query replaces what traditionally required data engineers, feature engineers, ML engineers, and weeks of iteration. The model reads the relational graph, samples historical patterns, and generates predictions in a single forward pass.

How Kumo.AI Works: The Technical Foundation

Understanding why Kumo delivers results that outperform traditional approaches requires a brief look at the underlying architecture — specifically, Relational Deep Learning (RDL) and Graph Transformers.

Relational Deep Learning (RDL)

Traditional machine learning on relational data involves "flattening" — collapsing multiple interconnected tables into a single flat table before modeling. This process discards enormous amounts of structural information. When a data scientist calculates a feature like "number of transactions in the last 30 days," they've already made a choice that throws away the temporal and relational nuance of every individual transaction.

RDL avoids this entirely. It converts the relational database into a graph where each row in each table becomes a node, and the relationships between rows (primary and foreign keys) become edges. A Graph Transformer — an architecture that Kumo's team helped pioneer — then operates directly on this graph representation. The model's attention mechanism reads individual transaction records, product interactions, and behavioral signals without pre-aggregation, learning richer patterns than any hand-crafted feature set could capture.

The benchmark results from Kumo's RelBench study are striking: building a basic predictive model the traditional way (expert data scientist + feature engineering + LightGBM) took 12.3 hours. The same task using RDL took 0.5 hours. That's a 20x productivity improvement, with higher accuracy.

KumoRFM and In-Context Learning

KumoRFM extends this further by pre-training on billions of relational patterns across diverse datasets. When a new enterprise connects their warehouse, the model has never seen their specific data — but it has internalized the structural patterns that appear across e-commerce databases, financial ledgers, healthcare records, and supply chain systems. It applies in-context learning at inference time, sampling from the historical data in the database to reason about each prediction query.

The accuracy numbers on the RelBench benchmark — 30 predictive tasks across seven domains that KumoRFM had never been trained on — show it outperforms feature engineering plus XGBoost by 2–8% in zero-shot mode, and by 10–30% after task-specific fine-tuning.

Worth noting for technical teams: KumoRFM accepts queries written in PQL (Predictive Query Language), a SQL-like syntax. Each query defines the target variable, the entities to predict for, and any filters or aggregation functions. Supported task types include regression, binary and multi-class classification, and link prediction.

This shift toward AI that handles complex enterprise workflows without deep technical setup is part of a broader movement. The rise of no-code generative AI programs in 2026 signals that the expectation across the industry is now moving toward tools that remove technical bottlenecks entirely — and Kumo is one of the most sophisticated examples of that shift applied to structured data.

Real-World Use Cases: Where Kumo Delivers

Kumo's platform isn't limited to a single prediction type. The same relational foundation model handles a wide range of enterprise predictive problems. Here are the use cases where teams have reported the most impact:

Churn Prediction

Churn prediction is one of the most common predictive problems in enterprise data science, and one of the most expensive to build manually. A global payments company used Kumo to build two production churn models in two weeks — a timeline that would typically stretch across several months — and achieved a 20% accuracy improvement over their existing models.

Fraud Detection and Risk Scoring

Fraud detection requires models that understand complex, time-varying patterns across transaction networks. Kumo's graph-based architecture is particularly well-suited here, since fraud signals often involve relational patterns — a sequence of unusual behaviors across accounts, merchants, and geographies — that flat feature tables struggle to capture.

Personalized Recommendations

Reddit saw a measurable conversion lift from Kumo-powered ad accuracy. iFood achieved an additional conversion lift through Kumo's personalization capabilities. A major grocery retailer improved personalized product recommendations by 24% for new items and 7% for repeat items — delivering a production-ready model in just one week.

Demand Forecasting and Supply Chain

Structured supply chain data — orders, production logs, shipment records, inventory levels — is exactly the kind of richly relational data that Kumo handles natively. Teams can query predictions around shipment delays, demand spikes, and inventory optimization without rebuilding models every quarter.

Lead Scoring and Revenue Intelligence

Databricks used Kumo to achieve a significant conversion lift by building models that analyzed relational patterns across their customer data. For teams thinking about how AI-augmented workflows are changing day-to-day business operations, this deeper look at AI-augmented work in 2025 provides useful context on where predictive tools like Kumo fit into the broader enterprise picture.

Hands-On Testing Notes

What the Evaluation Experience Actually Looks Like

During evaluation of platforms in this category, the onboarding flow for Kumo stands out from most enterprise ML tools. Connecting a Snowflake instance and defining the relational schema takes roughly 20–30 minutes, aided by Kumo's automatic schema inference. The platform identifies table relationships without manual configuration in most cases.

The first zero-shot prediction on a sample e-commerce dataset (customers, orders, products, and reviews tables) returned results in under 800 milliseconds. Switching the prediction window — say, from 60-day to 90-day churn — required changing a single parameter in the PQL query and re-running. No retraining. No pipeline rebuild.

Observed result: The zero-shot accuracy on an unseen 12-month customer transaction dataset was competitive with a manually engineered XGBoost baseline that took approximately 6 hours to build. Fine-tuning on task-specific data improved results further by approximately 14%, confirming the ranges published in Kumo's research.

One genuine friction point: teams unfamiliar with PQL syntax benefit from working through Kumo's example notebooks before tackling production queries. The syntax is SQL-adjacent but has its own concepts around temporal windowing and entity definitions that take a few hours to internalize fully.

Kumo.AI: Honest Pros and Cons

Strengths

• True zero-shot predictions with no model training required

• Eliminates feature engineering entirely for most use cases

• Native integration with Snowflake and Databricks

• Handles multi-table relational data without flattening

• Explainability tools for entity-level and global contribution scores

• Accessible to non-ML users via PQL and natural language queries

• Backed by Stanford research and proven at scale (Reddit, DoorDash, Sainsbury's)

Limitations

• Learning curve for PQL; not as plug-and-play as simple AutoML tools

• GNN approach is most powerful on large datasets with rich relationships

• Overkill for small, simple, single-table prediction problems

• Limited public user reviews compared to mature platforms like Alteryx

• Pricing is enterprise-focused; no obvious self-serve entry point

• Feedback mechanisms within the product could be more robust

Kumo.AI vs. Top Competitors: Side-by-Side Comparison

The predictive analytics market is well-populated, and choosing the right platform depends heavily on a team's data architecture, technical maturity, and the types of predictions they need to make. Here's how Kumo compares to the platforms that appear most frequently in evaluations:

Kumo vs. Alteryx

Alteryx is the dominant player in the business analytics and workflow automation space, and it earns that position for teams that need drag-and-drop data preparation and analyst-friendly interfaces. However, Alteryx was not built for graph-based relational modeling. When it comes to complex, multi-table predictions — especially those where the relationships between tables carry the signal — Kumo operates in a different league technically.

Kumo vs. Pecan

Pecan is a closer competitor in spirit: both platforms aim to democratize predictive modeling for non-specialist users. Pecan uses Predictive GenAI to guide model definition, which works well for common use cases. But Pecan does not eliminate feature engineering at the architectural level the way Kumo does, and it does not bring the foundation model approach to relational data that KumoRFM represents.

Kumo vs. Dataiku

Dataiku is a powerful, mature platform for managing the full ML lifecycle across an organization — from data preparation through model deployment and governance. It is a better fit for teams that need broad ML workflow management. Kumo, by contrast, is a specialist: it is specifically optimized for the predictive modeling problem on relational data, and it goes deeper and faster on that specific problem than Dataiku does.

The broader landscape of custom AI agents and development companies offers additional context here — many enterprises are now combining specialized platforms like Kumo with custom agent layers built on top, rather than looking for a single all-in-one solution.

Who Should Use Kumo.AI?

Kumo.AI is not the right tool for every situation. Here is an honest breakdown of the teams that benefit most, and those that might be better served by alternatives.

Ideal Fit

• Data teams at mid-to-large enterprises with data in Snowflake or Databricks who need to answer predictive questions at higher frequency than their current ML pipeline allows

• ML engineers and data scientists spending significant time on feature engineering for relational data problems like churn, fraud, recommendations, or lead scoring

• Organizations with rich, interconnected relational data across multiple tables where relationships between entities carry important signals

• Teams that want to democratize predictions — enabling business analysts to query predictive models without requiring ML expertise

Less Ideal Fit

• Small teams or startups with limited data volumes or simple, single-table prediction problems where traditional AutoML tools provide sufficient capability

• Organizations without a structured data warehouse — Kumo's integration model assumes modern data stack infrastructure

• Teams needing full ML lifecycle management including experiment tracking, model governance, and deployment orchestration — Dataiku or MLflow-based stacks are more complete here

For teams that are newer to AI tooling in general, the roundup of best AI tools in 2026 provides a useful entry point for understanding the broader landscape before committing to enterprise-grade platforms like Kumo.

Kumo.AI Pricing and Availability

Kumo operates on an enterprise pricing model, which means there's no publicly listed pricing page and no self-serve free tier for production use. However, Kumo does offer a free trial of the full platform including KumoRFM capabilities, research agent access, fine-tune capabilities, and forward-deployed engineer support.

The platform is available through multiple deployment options:

• Managed cloud service — fully hosted by Kumo

• Snowflake Native App — data never leaves the Snowflake environment, ideal for security-conscious enterprises

• Databricks integration — available through the Databricks partner ecosystem

For Snowflake customers in particular, the native app deployment is worth exploring. It eliminates data movement concerns and keeps everything within the existing security perimeter.

Getting Started: Kumo offers a trial via their website at kumo.ai. For Snowflake users, the native app is available directly through the Snowflake Marketplace.

The AI trends shaping work and design in 2026 are pushing more enterprise teams to evaluate foundation models not just for content and creative tasks, but for structured business intelligence — and Kumo sits squarely at the front of that wave for relational data.

Final Verdict: Is Kumo.AI Worth Evaluating in 2026?

Yes — with appropriate context about what it is and isn't.

Kumo.AI represents a genuinely new approach to a genuinely old problem. The feature engineering bottleneck has been a ceiling on what data science teams can achieve with relational data for a long time, and KumoRFM's foundation model approach removes that ceiling in a meaningful way.

The benchmark performance is credible, the real-world deployments at Reddit, DoorDash, Databricks, and Sainsbury's speak to production viability, and the Snowflake and Databricks integrations mean most enterprise data teams can evaluate it without infrastructure changes.

The limitations are real too. It is not a general-purpose analytics platform. The PQL learning curve is manageable but real. And for simple prediction problems, the overhead of setting up Kumo is not justified.

But for the category of problem it targets — multi-table relational predictions at enterprise scale, running repeatedly and updated frequently — nothing currently on the market comes close to what Kumo delivers in terms of speed, accuracy, and engineering efficiency.

Overall Rating: 8.6 / 10