Case study · Scout
Scout: classifying supplier products in a shared catalogue with AI
Scout connects property organisations, which manage properties made up of apartments or units, with product suppliers and service providers. For their teams to find items in a shared catalogue, each product must be placed in the right category, however its supplier describes it. We built several AI-assisted categorisation mechanisms into the platform: for a single product, for the classification step of the import and for a batch. In the import flow, each suggestion comes with a confidence level and distinguishes products to assign, to review, to reject or with no match.
Scout is a B2B procurement platform built with Medusa and Mercur. On one side, property organisations buy for the properties they manage. On the other, suppliers publish their products and service providers offer their services. Each supplier arrives with its own catalogue, its own item names and often its own way of grouping items. These files are heterogeneous, whereas Scout relies on a shared taxonomy: teams buying for a property must find equivalent products in the same place, whoever the seller. Matching each item to a category in this taxonomy represents a considerable amount of work when a supplier brings hundreds or thousands of items. In the supplier workspace, we therefore built tools that rely on AI to suggest this classification, making it readable and leaving room for uncertain cases.
A shared taxonomy for catalogues that do not look alike
How a product is classified determines how it will be found. A supplier may organise its items according to its own commercial logic, a manufacturing family or an internal label. On a marketplace that brings several suppliers together, these logics do not sit well side by side. As an illustration, the same mixer tap may appear under “Sanitaryware” at one seller and under “Taps and valves” at another. The team of a property organisation looking for this equipment for its apartments should not need to know both vocabularies.
Scout therefore imposes a shared organisation. The supplier can still choose a product record’s category itself from its dashboard, product by product. This remains simple for a single item. It becomes cumbersome when a complete catalogue arrives all at once, and it is at this scale that AI comes in.
Categories suggested from among those that already exist
The AI must not invent headings. If each product could be given a free-form label, the shared catalogue would immediately fragment. Classification therefore starts from Scout’s taxonomy. When categorisation is requested for a single product, the platform first searches this taxonomy for categories that could match it, based on what the product record contains. These candidate categories are then submitted to an OpenAI model, which chooses the best fit from among them. The classification step of the import flow and batch processing rely on separate mechanisms, which are not interchangeable.
In the classification step of the import, the result is not limited to a single answer. It presents the categories considered and a confidence level. The supplier thus knows whether the suggestion is clear-cut or whether several headings remained plausible. This information matters as much as the category itself: it is used to decide what can be applied as is and what deserves closer attention.
Four outcomes to separate the obvious from the uncertain
A real catalogue contains products that are easy to classify, ambiguous products and items that have no obvious place in the taxonomy. Treating all these cases in the same way would mean either checking everything by hand or accepting errors without seeing them. The categorisation step of the import flow, which handles at most one hundred products per request, therefore returns one of four outcomes for each product.
When the suggestion is sufficiently certain, the category can be assigned directly: these products do not go through prior human review. When there is genuine doubt, the product is marked for review and the supplier can make an informed decision. A suggestion can also be rejected, in which case the category considered is not kept. Finally, when no candidate category emerges from the taxonomy, the result says so clearly instead of forcing a match. The supplier thus knows where to focus its attention: the cases to review, the rejections and the products left without a match, rather than the entire file.
One product to correct or an entire catalogue to prepare
Needs do not always arise at the same scale. A supplier that edits a product record, or notices that an item is misclassified, wants a suggestion for that product alone without rerunning the processing of its entire range. Categorisation can therefore be requested for an individual product.
Conversely, the arrival of a complete catalogue calls for batch processing. This processing can take time, and the supplier should not be stuck in front of its screen while waiting. Categorisation batches are therefore tracked: the supplier views its processing jobs by type and status and sees their progress, in the same way as for its file imports. It knows whether the work is in progress, finished or needs attention.
A step in the import flow, through to review
In the supplier workspace, categorisation is part of the flow used to bring in a catalogue. The supplier uploads its CSV file, maps its columns to Scout’s, converts the values that need converting and validates the data. It then has its products categorised, and reconciles the result with its existing catalogue. Because its column mappings can be saved and reused, a supplier that regularly sends the same file format does not have to reconfigure everything. An interrupted import can also be resumed where it left off.
Once the products are classified, they still need to enter the marketplace catalogue. The supplier can submit all the already categorised products in one go, which moves this prepared catalogue to the review stage provided by the platform. It does not have to submit them one by one: the classification work leads directly to the next step.
With these tools, Scout’s suppliers no longer have to place each item by hand in a taxonomy they may not know. Suggestions remain within the shared organisation. In the import flow, they separate what can be assigned from what needs to be reviewed, rejected or left without a match. The classification is not presented as infallible: it flags the products on which a human decision remains useful.