Bringing AI Predictions to Agents - Announcing KumoRFM Support for Model Context Protocol (MCP)
September 3, 2025




Josh Przybylko, Jure Leskovec, Salli Liu, Matthias Fey, Blaž Stojanovič

Agents are redefining how we think about enterprise software. In fact, according to Capgemini[1], “82% of Organizations intend to integrate an AI Agent within 1-3 Years.”
Instead of rigid, rule-based applications with fixed workflows, AI agents bring adaptability and intelligence into everyday business tools. They can reason about context, call APIs, query databases, and take actions on behalf of users.
Agents need predictions
For agents to make truly informed decisions about what actions to take, they must go beyond simple data retrieval. They need the ability to tap into enterprise data and generate forecasts, recommendations, risk assessments, and other types of predictions that guide reasoning and actions.
The first generation of agents has mainly concentrated on information retrieval, commonly utilizing techniques such as Retrieval-Augmented Generation (RAG). While large language models (LLMs) excel at answering questions about existing knowledge, they are inherently retrospective. They surface and reason over information that already exists.
Predictions unlock a new dimension of value. For example:
- A customer services agent could predict a customer’s churn risk and automate outreach with preventative actions, such as providing discount or bundling offers.
- A logistics agent could predict delivery delays using traffic and weather data, then reroute shipments.
- A finance agent could predict payment default risk and adjust credit terms dynamically.
- A sales agent could forecast a customer’s likelihood to respond to outreach and estimate lifetime value to tailor offers accordingly.
The key is that for agents to make good decisions on which actions to take, they need to make forward looking estimations and forecasts, that is, the agents need to make accurate predictions.
However, for these predictive business tasks, companies today still rely on classic machine learning techniques that are both resource and time-intensive. Companies need specialized data science teams to perform extensive “feature engineering”, the manual crafting of predictive signals from raw data, and iterating through model training. Furthermore, when agents require multiple types of predictions, separate models must be trained, deployed, and maintained for each task, and in some cases for each customer or region.
If we believe in an agentic future, enterprises need a faster, more flexible approach to machine learning, one that reduces friction and brings prediction into the natural flow of agent reasoning.
Enter KumoRFM
KumoRFM is a Relational Foundation Model built for predictive analytics on structured data. Unlike traditional machine learning models, it leverages in-context learning, similar to large language models, to adapt to new datasets and prediction tasks at inference time. This entirely new approach to machine learning removes the need for extensive feature engineering or bespoke model training, while also delivering superior performance. On benchmarks such as RelBench, KumoRFM outperforms standard supervised learning methods by 2–8%, with an additional 10–30% boost when fine-tuned[2].
At its core, KumoRFM uses a Graph Transformer-based architecture for relational deep learning, pre-trained on volumes of public and synthetic relation data, enabling training-free predictions[2]. Kumo RFM fundamentally streamlines the process of generating AI predictions from relational data, making it simple to integrate those predictions directly into agentic workflows.
KumoRFM Introduces MCP Support
Today we’re excited to announce support for the Model Context Protocol (MCP), the open standard for connecting AI agents with external tools and data. By introducing support for MCP, it is now easier than ever to introduce data driven predictions into agentic workflows.
Unlike proprietary plugins or one off integrations, MCP provides a unified, secure, and consistent way for LLM based agent orchestrators, such as Claude and GPT, to interact with external systems. Additionally, MCP simplifies integration with popular agent frameworks like LangChain, Crew.ai, and the OpenAI Agents SDK, ensuring that developers can quickly compose and extend agentic workflows without reinventing the wheel. With this integration, KumoRFM can serve as a predictive engine within any agent workflow, working seamlessly alongside other tools that an agent may call upon, as depicted below.
Example: Warranty Renewal Agent
Alongside the launch of the KumoRFM MCP server, we have developed a cookbook demonstrating the development of a multi-step agent workflow leveraging Kumo RFM MCP. Consider the extended warranty market, valued at $187 billion in 2024 and projected to reach $426 billion by 2033[3]. This market has a notoriously high rate of churn. For example, in the USA, homeowners renew home warranties at a rate no higher than 50%, and most cancellations occur after the first year.
Pre-emptive churn prediction and personalized retention strategies are critical to profitability in this space.
Here’s how a predictive agent powered by KumoRFM + MCP could work:
- Identify customers with upcoming renewals
- Predict churn risk and flag at-risk customers
- Explain the contributing factors behind churn
- Recommend personalized discounts to maximize retention
- Generate tailored outreach and offers
- Automate this loop on a recurring basis
Today, constructing such a workflow is beyond the capabilities of LLMs alone. While powerful for language tasks, LLMs struggle to generate predictions from relational data without hallucination. Achieving such a workflow instead requires specialized teams to design, train, and maintain bespoke predictive models.
By contrast, this agent unifies two critical prediction tasks—churn classification and personalized recommendations—using a single relational data model input and a single KumoRFM instance. It delivers high accuracy, built-in prediction explainability, and confidence in results, all without the need for specialized training and the overhead of ongoing model maintenance.
Try KumoRFM MCP Server Today
By integrating KumoRFM into the MCP ecosystem, we’re unlocking agents that don’t just understand the present—they anticipate the future. With built-in AI predictions, workflows shift from reactive to proactive: anticipating customer needs, optimizing business operations, and driving more intelligent decisions.
The KumoRFM MCP server is live and free to try. Explore it now in the KumoRFM Free Trial, complete with cookbooks and examples to get started in minutes. Explore our demonstrational cookbooks to see KumoRFM MCP server in action.
For deeper dives, check out our research paper, coverage in VentureBeat, and our company updates.
References:
[1] Capgemini Research Institute (2025) Harnessing the value of generative AI. 2nd ed.
[2]Kumo.ai (2025) KumoRFM: A foundation model for in-context learning on relational data
[3] Astute Analytica (2025) Extended warranty market to worth over US$ 426.76 billion by 2033