A French foodtech brand recommends the right next product, customer by customer. From segment-based campaigns to individual predictions.

A leading foodtech brand runs a high-frequency consumer model, where customers come back week after week and the catalog evolves constantly. In that kind of business, the difference between a flat customer and a growing one is whether the brand can recommend the right next purchase at the right moment.

The company had a strong product, a deep data foundation, and an engaged customer base. What it didn't have was a way to move from broad segmentation to individual recommendations without building a data science team.

The CRM team was running cross-sell and upsell campaigns on broad customer segments. "High-value customers", "new buyers", "lapsed users", each segment received the same product push, regardless of what the individual customer was actually likely to buy next.

The result was a treadmill. Conversion stayed flat, the team spent more time maintaining segmentation rules than optimizing campaigns, and every test ran on the same assumption: that a segment behaves like a single customer. It doesn't.

Moving to individual predictions meant predictive modeling at scale, refreshed as behavior shifts, deployed into the CRM and email tools the team already used. The kind of project that usually takes a year and a dedicated team.

Matr gave the brand a way to put individual next-product predictions in production without growing the team or running a parallel data program. The data team owned the work, the modeling complexity moved to Matr, and the CRM team got to use the output where campaigns are actually built.

What sold the team was the combination: a model trained on their warehouse data, predictions written back where the CRM team already works, and an explainability layer so every recommendation came with the reason behind it, ready to be challenged, adjusted, or shipped.

The brand's data team connected Matr to their warehouse and built the first next-product model in 2 days, with no dedicated data engineering effort. Each customer gets an individual predicted next purchase, refreshed weekly as their behavior on the platform evolves.

From there, the data team operates Matr the way they'd operate a data science team that never sleeps: defining the prediction objective, validating the modeling approach, monitoring drift, retraining when behavior shifts. The model stays accurate over time, with no parallel infrastructure to maintain.

This is what made predictive recommendations useful rather than just accurate. The CRM team doesn't open Matr or interrogate model internals. They consume the output where they already build campaigns.

For every customer, the model surfaces the next product most likely to convert and the reason behind the prediction. The CRM team uses both: the who to target the right customers, the why to write the right message. A customer flagged for category expansion gets a discovery campaign. A customer flagged for repeat purchase gets a replenishment push. A customer flagged for upgrade signals gets an upsell offer.

The result isn't a one-off recommendation engine. It's a working capability: a data team running individual predictions in production, and a CRM team building every campaign on them.

• +22% conversion on personalized recommendation campaigns
• 2 days to build and deploy the first model in production
• 3 million next-product predictions refreshed weekly
• Used by CRM and marketing teams to power every retention and growth campaign, with no manual segmentation maintenance

Today, the brand runs next-product prediction alongside other Matr models on the same platform, with the same data and the same explainability layer.