Attentive Sends 75 Billion Personalized Messages for Over 640 Million Shoppers

The text arrives when you’re daydreaming about the shoes in your shopping cart or the sale that’s about to expire. There’s a good chance Attentive is behind that nudge, powering the SMS, email, RCS and push notifications that connect over 8,000 brands, including Steve Madden, Urban Outfitters and Crate & Barrel, with over 640 million subscribers. True to its name, Attentive drives 25% higher online revenue per customer year over year.

From when to reach out to a customer to what to recommend, the intelligence behind these messages runs on ML models, powered by structured data signals called features. These features lived inside a six-system stack built around Tecton, Attentive’s former feature store, requiring Snowflake, EMR, Jupyter, Metaflow and Anyscale to keep it running. Shipping a single feature meant coordinating all of them.

It’s no surprise that more than 30% of the ML platform team’s time went to operational maintenance and queuing issues instead of model development. Not to mention the licensing and compute costs of running six systems at once.

Tecton’s limitations made the data less accurate, too. It silently dropped late-arriving events, leading to models trained on incomplete data and train-serve skew that caused the pipeline to miss daily targets. The result was a personalization engine powering billions of messages operating on less complete data.

Attentive turned to Snowflake ML and replaced years of custom-built infrastructure with Snowflake’s natively integrated Feature Store. The complex calculations that determined when someone was likely to buy, churn or unsubscribe, previously built on custom code across multiple systems, became managed, automated and consistent. Most importantly, late-arriving data is captured accurately in Snowflake Feature Store, so models train on a complete picture.

By the time migration began, Attentive’s newer models had already moved off the old system and only five needed to migrate. The team went from proof of concept to production in six months, with Snowflake Professional Services supporting throughout.

Today, 2,500 features across 136 feature views feed five migrated models serving more than 640 million of Attentive’s subscribers, all on a single platform. Tecton and EMR are being retired entirely, and every new model Attentive builds will follow the same path.

On top of Snowflake Feature Store, the team also built the Attentive Feature Store SDK to standardize how engineers define and ship new signals to their models. What once required weeks of coordination across six systems now takes a configuration file and two commands. The same setup runs in a lightweight test environment and scales to full production without changes.

Training Attentive’s product recommendation model on 3.2 billion rows now takes two and a half hours instead of eight. Scoring subscribers for campaigns (inference) dropped from three hours to under two, so messages go out sooner with fresher, more relevant predictions. The data refresh jobs that feed those models, which used to queue for hours and cause missed delivery windows, now frequently finish in five minutes. Infrastructure costs are down approximately 30%.

The five models migrated from the old system, covering product recommendations, send-time optimization, opt-out prediction and more, also run on more accurate data. Every model now trains on a more complete picture of each subscriber.

For Attentive’s over 8,000 brand customers and the more than 640 million subscribers they reach, better models mean messages that feel right instead of accelerating unsubscribes. Next, Attentive is moving from scheduled refreshes to near real-time signals. Soon, the moment a subscriber lingers on a cute pair of slippers or the perfect birthday gift at 11 p.m., Attentive will make sure that signal shapes what they hear next, staying true to its name.