Build Agentic Ecommerce with KumoRFM

The useful idea here isn’t a chatbot. It’s a workflow: keep LLMs out of heavy, structured data and let them orchestrate specialized predictive systems. Kumo’s Relational Foundation Model (a GNN-based predictive engine) becomes a tool the agent calls via a query language (PQL). The LLM handles intent parsing, tool selection, and composing queries; Kumo returns probabilities (purchase, churn, next-best-product) quickly and at scale.

This shifts ecommerce “AI” from FAQ chatbots to decision support and automation in marketing, merchandising, and retention. Examples shown: predicting a customer’s 30-day purchase probability, top-N likely SKUs, and churn risk; then auto-generating highly targeted emails. The important change is architectural: the agent doesn’t ingest large tables; it asks a purpose-built model to compute exactly what’s needed, reducing token bloat and hallucinations on structured data.

Tradeoffs are clear. The LLM still mis-writes queries and needs a strict guide prompt and step limits. Long IDs inflate token costs/latency; mapping/aliasing is required. Kumo RFM is a vendor product with its own PQL; expect lock-in, calibration questions, and governance work (explainability, auditability, and data access controls). The demo uses a curated sample dataset; no evidence of uplift, calibration, or real-time performance under production loads.

Operationally, you’ll need schema governance (semantic types, primary/time keys), predictable tool-call costs, and guardrails against unbounded iterative calls. This pattern also forces clarity on data contracts between your warehouse and the predictive service.

Most will fixate on the slick email personalization. The real lever is a repeatable pattern to add predictive primitives (propensity, NBO, risk) behind an agent and expose them safely to internal users or customer-facing flows.