---
title: "Why propensity models stall in production | Snowplow Signals"
description: "Propensity models stall between a notebook and production: labeled data, a feature pipeline, validation, serving and train-serve skew before any score ships."
source: https://signals.snowplow.io/learning/intent-and-propensity/why-propensity-models-stall
---

[Learning](https://signals.snowplow.io/learning) / [Intent and propensity](https://signals.snowplow.io/learning/intent-and-propensity) 

# Why propensity models stall in production

Propensity models often stall before they change what a customer sees. The gap is between a working notebook and production, where labeled data, a feature pipeline, offline validation, a serving path and train-serve skew each have to be solved before the first score ships.

By [Alex Dean](https://signals.snowplow.io/blog/authors/alex-dean) Published Oct 26, 2026 

## The gap between a notebook and production

A propensity model is often convincing in a notebook long before it changes anything a customer sees. The hard part is not fitting the model; it is everything between a good offline result and a score that runs on live traffic and drives a decision. That gap is where most of these projects sit.

## Where it stalls

Getting there means labeled training data, a feature pipeline, offline validation, and a path that serves the same features in production. Then the values in production have to match the values the model trained on, or the predictions drift without anyone noticing. Each step has an owner and a queue, and the result is a launch date that keeps moving while the model waits behind the data work.

## A lighter path to acting on intent

When the goal is to act on what a visitor is doing now, there is a shorter route than a full model. A typed question over the live state of the session sorts the visitor into one of a few named states, with no training set to collect, because the states are the definition. Signals computes and serves that live state; your code acts on the answer.

When you do need the model, the serving path and the skew are the steps Signals takes off the list: one definition builds the training set and serves the live value, so production reads what the model learned from.

## Put it to work

* [Sort visitors by intent without a propensity model →](https://signals.snowplow.io/lp/intent-buckets)
* [Real-time attributes →](https://signals.snowplow.io/products/real-time-attributes)

## See it on your own traffic 

Add the SDK, define one attribute, and see an attribute updating against your own traffic in the first session. The free tier runs on the real engine: no card, no sales call.

[Start free](https://signals.snowplow.io/start) [Talk to an engineer](https://signals.snowplow.io/start?intent=engineer) 

More explainers in the [Learning Center →](https://signals.snowplow.io/learning)
