Top Reps Multi-Thread 2.4x More. A Context Layer Can Now Prove It Predicts Wins.

Nate Nasralla draws a hard line between two things vendors both call “context” for AI sales agents: a context graph, which retrieves and stores facts, and a context layer, which trains ML models on historical closed-won and closed-lost outcomes to produce statistically validated judgment. Naming Glean, Sybill, and Clearskies as graph-only tools, his Fluint playbook argues that retrieval alone can tell an agent what happened in a deal, but only a trained model can tell it what predicts whether a deal closes.

The build process he lays out is deliberately unglamorous: capture GTM data as immutable, timestamped event logs via change-data-capture rather than CRM snapshots; engineer features in a point-in-time-correct store to avoid leakage; train regression, gradient-boosted tree, or KNN models against actual revenue outcomes; validate with holdout tests and p-values; and serve the resulting scores through push notifications and MCP into existing workflows. He backs the approach with specific numbers — top performers multi-thread 2.4x more than average reps, deals with a shared business case before the third meeting show 40% higher win rates, and executive-level champions close 3.2x more often than manager-level ones — while noting 150-200 closed deals are typically needed before the model is statistically reliable.

That last constraint is exactly what separates Nasralla’s approach from the “custom agent” pattern Gong pushed three weeks earlier: no-code agents that reason over a graph of mapped revenue interactions, built by describing a trigger and a desired output in plain language. Gong’s version ships faster and needs no model training; Nasralla’s version needs a real sample size and a permanent MLOps function, but produces a scored prediction instead of a retrieved summary.

Both are AI-stack decisions worth separating clearly: one is context retrieval with a chat interface bolted on, the other is a trained model with retrieval bolted on. Reps evaluating either should ask which one they’re actually buying.