Next Best Offer — Rexel

The Signal

Rexel had a powerful ML algorithm predicting what products a customer was likely to buy. It was accurate. It was ignored.

Sales reps at the counter didn't trust it, didn't understand it, and actively worked around it. The algorithm was outputting product IDs into a backend system that technically functioned — but had never been designed for humans.

The brief: "Make sales reps use the recommendations."

Context

Rexel is a global electrical distribution company with 30,000+ employees and hundreds of B2B stores. Each day, thousands of sales interactions happen at counters — moments where a well-timed suggestion could meaningfully change the basket.

The existing recommendation system was built by a data science team and had been running for 18 months. It had never shipped to users. No one had designed the interface. The data team handed over a CSV. Someone built a very basic admin panel. Nobody used it.

Constraints: existing POS system (can't replace), 8-second interaction window, reps speak 4 languages across pilot markets.

The Work

Phase 1 — Discovery (2 months)

22 in-store interviews across France, Germany, Spain, UK. Observed 60+ counter interactions, timed the decision window, mapped the critical moment: the 8 seconds after a customer hands over a product.

Key finding: reps weren't lazy or resistant. They were busy and didn't trust something they couldn't explain. "I can't recommend something when I don't know why I'm recommending it."

Phase 2 — Design & Co-creation

Rebuilt the recommendation surface as a single-line "nudge" inside the existing POS — not a separate tool, not a new tab. Each suggestion included a reason written in plain language: "This customer bought this before" or "Teams in their industry typically need this."

Added a thumb up/down feedback mechanism to train the model over time.

3 rounds of co-creation workshops with reps from each market.

Phase 3 — Pilot & Iteration

8-week pilot in 12 stores across France and Germany. Weekly syncs with store managers. Iterated on copy, reasoning tone, and positioning of the nudge.

Key Insight

The reps didn't need better recommendations. They needed a reason to speak up.

The copy on each suggestion was as important as the algorithm. Trust was built through transparency, not accuracy.

Outcomes

• +12% average cart size during pilot
• +9 NPS points from customers
• −17% return store visits
• −4% churn in pilot stores