QREQualification & recommendation
B2B distribution · Product marketing
21,000 products classified a year, at 82% first-proposal precision.
Manutan had to file thousands of new SKUs into its internal taxonomy. A three-level scoring engine with a HITL loop runs in production.
Key metrics
- Client
- Manutan
- Sector
- B2B distribution
- Function
- Product marketing
- Status
- Live in production
- Deployment
- Client platform
The context
Every year, thousands of new products from many suppliers must be filed into article sub-families — the taxonomy that structures offer, merchandising and search.
The problem
- Manual classification, heterogeneous and sometimes multilingual descriptions.
- Errors hitting findability and sales. No confidence score on manual assignments.
The solution
Multi-supplier ingestion, semantic analysis, embedding match, high / medium / low confidence scoring, HITL loop, capitalization on every correction.
NEXA components in scope
| Component | Role |
|---|---|
| Arbitration algorithm | Multi-criteria semantic assignment |
| Evidence Panel | Score, alternatives, justification |
| HITL | Merchandising validation / correction |
| AI Knowledge Vault | Validated patterns for the next wave |
| Cockpit | Green / amber / red volumes |
The impact
| Axis | Concrete impact |
|---|---|
| Precision | 82% correctly classified on the first proposal (high confidence). |
| Volume | ~21,000 products/year from 39 supplier files. |
| Adoption | Merchandising team equipped across the full flow. |
QREQualification & recommendation
« Classifying 21,000 SKUs a year by hand was no longer viable. The first proposal arrives with a score: we validate, we no longer reconstruct. »
82%
First-proposal precision
Your business ritual has the same potential.
Evidence is not our constraint. It is our product.