# Graph Neural Networks Just Solved Enterprise AI?

Jure Leskovec, Stanford University and Kumo.AI | MLOps Podcast | Episode 348 | 49:01
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=Tq613N0dyI0
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/graph-neural-networks-just-solved-enterprise-ai
Published: 2026-01-16
Tags: agents, feature-engineering, forecasting, recommender-systems

## TL;DR
- Relational Foundation Models learn directly from connected enterprise tables, which removes much of the manual feature engineering used in predictive models.
- Graph-based models can combine attributes and interactions to improve recommendations, fraud detection, advertising predictions, and related time-series forecasts.
- Predictive models can give autonomous agents the data-based probabilities they need to make business decisions such as retention offers, fraud checks, and customer support actions.

## Summary
Jure Leskovec argues that enterprise AI has overlooked its most valuable data: structured records and the relationships between them. Traditional predictive models usually join tables, create features, and then train a model, a process he says can take months and creates ongoing maintenance problems. His approach, called relational deep learning, represents connected tables as a graph and lets a transformer learn from the raw relational data. A pre-trained Relational Foundation Model can answer predictive queries across a database, while task-specific fine-tuning can improve performance on high-value use cases. Leskovec describes examples from Pinterest, DoorDash, and Reddit, where interactions between users, products, restaurants, and ads provide signals that sequence models or manually written features miss. He also connects these models to autonomous agents, which need predictions about churn, lifetime value, fraud, or the next best offer before they can make useful decisions. The approach still requires clean schema relationships, data and engineering work, and a system for acting on predictions.

## Key ideas
### Enterprise predictive models are built on relational data that is still treated like a single table
[00:26](https://www.youtube.com/watch?v=Tq613N0dyI0&t=26s)
Jure Leskovec describes churn, recommendation, risk, and fraud models built from databases with users, products, orders, and behavior records. Teams usually join those tables, create features, keep the features current, handle time-travel issues, and then put the model into production. He contrasts this with computer vision and language modeling, where neural networks learned directly from pixels or token sequences. His claim is that predictive modeling has stayed tied to manually featurized data even though the raw business data already contains the structure needed for learning.

### Relational deep learning lets one model learn across connected tables
[03:28](https://www.youtube.com/watch?v=Tq613N0dyI0&t=208s)
Leskovec says his models use a graph representation of database data, so they can learn over multiple tables at once. A relational graph transformer attends over the tables in a database schema rather than over tokens in a normal language model. This removes much of the feature engineering step. The model also sees finer-grained signals because it can learn how to combine individual transactions, interactions, and other records instead of relying on statistics chosen in advance by a data scientist.

### Enterprise databases need models trained for structured data rather than textified prompts
[08:04](https://www.youtube.com/watch?v=Tq613N0dyI0&t=484s)
Leskovec separates Relational Foundation Models from using a large language model for enterprise prediction. An LLM may have general common sense, but it does not know the fraud probability of a particular transaction or the likely behavior of a specific user. A pre-trained relational graph transformer can answer predictive questions across a database without training a separate model for every question. He compares this with the need for DNA and protein foundation models in biology, where the data has properties that natural language models do not capture.

### The workflow needs schema relationships and column types, then supports pre-trained or fine-tuned models
[10:41](https://www.youtube.com/watch?v=Tq613N0dyI0&t=641s)
In practice, users select the tables they want to use, identify primary and foreign key relationships, and provide semantic types for columns. Leskovec says this preparation is relatively small compared with manual feature engineering. A pre-trained model can answer an ad hoc predictive query over the collection of tables. Teams can also fine-tune the model for a particular database and task when they need the highest performance, such as in fraud detection. The model does not require a fully developed ontology or knowledge graph, and it can often infer links from column names.

### Attention can replace arbitrary feature counters with learned combinations of raw events
[20:52](https://www.youtube.com/watch?v=Tq613N0dyI0&t=1252s)
Leskovec explains that traditional fraud features might count transactions over a chosen period, define morning hours, or calculate spending during a manually selected window. Different people then keep adding or changing those SQL-derived features. A relational model can attend to individual transactions and learn combinations that may be too fine-grained or unintuitive for a human to specify. He compares this with computer vision, where engineers no longer hand-code every edge or color gradient because the neural network learns how to combine pixels.

### Relational models can use attributes for cold-start cases and interactions once they accumulate
[18:47](https://www.youtube.com/watch?v=Tq613N0dyI0&t=1127s)
For a new item or user with little interaction history, the model can use attributes such as a description, image, or other properties. As users buy, click, or otherwise interact with the item, the model can use its growing position in the graph. Leskovec says this lets the model shift between attribute-based and relational signals. He also applies the same idea to time series. Related product, store, or stock series can be represented as a graph so the model can borrow information from correlated series instead of forecasting each sequence in isolation.

### Operational costs fall when predictions are regenerated from fresh raw data
[24:19](https://www.youtube.com/watch?v=Tq613N0dyI0&t=1459s)
Leskovec says a production deployment can be simpler because refreshing the raw data is enough to generate new predictions. Teams do not have to maintain a separate feature store with counters that can become stale or accidentally include future information. He says schema changes can often be handled without breaking the model, while a major change in table composition may require retraining. This also helps with changing user behavior and fraud patterns, since the model can be retrained to learn newer signals instead of waiting for people to invent another feature.

### Predictive tools give autonomous agents the information needed for real decisions
[42:53](https://www.youtube.com/watch?v=Tq613N0dyI0&t=2573s)
Leskovec argues that agents need more than document retrieval if they are expected to act. Predictions support decisions such as whether to stop a transaction, retain a customer, discharge a patient, or approve a loan. A retention agent, for example, might need predictions about who will churn and which offer is most likely to work before it writes and sends a message. Relational Foundation Models could expose these predictions as tools through a structured predictive query, allowing an agent to ask for a user's future purchases, transaction value, churn probability, or lifetime value and use the result in its next action.

## Notable quotes
- Jure Leskovec: "The approach we developed it's called relational deep learning and it can take any database represented as a graph of relations between the entities in your database and now you just learn over that graph of relationships." (08:04)
- Jure Leskovec: "We cannot textify a database, throw that into an LLM and hope it will work because it does not." (09:53)
- Jure Leskovec: "You are here to build models to have impact on the business. You are not here to clean data and feature engineer." (16:00)
- Jure Leskovec: "Prediction is really a basis for decision-making." (42:53)
- Jure Leskovec: "A real agent is actually something that does something to you." (45:16)

## Tools & references mentioned
- Relational Foundation Models
- relational deep learning
- relational graph transformer
- Kumo.AI
- Stanford University
- Pinterest
- DoorDash
- Reddit
- Anthropic
- GPT
- Tecton
- Databricks
- Featureform
- Redis
- Yandex

## Who should watch
- You are building churn, fraud, recommendation, advertising, or risk models from several connected enterprise tables and want to reduce hand-built feature work.
- Your production models depend on feature pipelines, feature freshness, and time-travel controls that are expensive to maintain.
- You are designing an autonomous agent that needs data-based predictions before it can make decisions or take actions.

## Related talks

- [GraphBI: Expanding Analytics to All Data Through the Combination of GenAI, Graph, & Visual Analytics](https://mlopstalks.com/talks/graphbi-expanding-analytics-to-all-data-through-the-combination-of-genai-graph) (Paco Nathan, Senzing, 1:12:38)
- [Language, Graphs, and AI in Industry](https://mlopstalks.com/talks/language-graphs-and-ai-in-industry) (Paco Nathan, Derwen, Inc., 1:18:29)
- [The Intersection of Graphs and Large Language Models](https://mlopstalks.com/talks/the-intersection-of-graphs-and-large-language-models) (Anthony Alcaraz, Fribl, 15:08)
- [How AI is Transforming Business Data Forever](https://mlopstalks.com/talks/how-ai-is-transforming-business-data-forever) (Deepti Srivastava, Snow Leopard AI, 57:14)
- [How AI Agents Are Transforming Data Analytics](https://mlopstalks.com/talks/how-ai-agents-are-transforming-data-analytics) (Ines Chami, Numbers Station, 24:26)