Case study: AI catalogue matching for an engineering systems distributor

Context

Orders arrive in any shape: a free-form list, a photo of a sheet from site, a message forwarded from chat, or a question about what would fit. They rarely carry an exact article number. Every line is looked up by hand in a catalogue of hundreds of thousands of items, and that is the daily work of over three hundred people.

Task

Make the matching automatic so the manager is left checking the result rather than searching for it.

Implementation

• Incoming order parsing: text, a photo of a document and a forwarded message are brought to one structure
• Search runs three ways at once: by meaning, by exact text and by structured fields such as type, diameter, material and brand. The candidates are merged and re-ranked
• Rules on top of search: the brand substitution table is maintained explicitly rather than guessed by a model. Pack sizes are applied automatically, so an order for 31 units in packs of 10 becomes 40 with a note to the manager
• Every match carries a confidence score that decides what happens next: an exact match goes straight to the basket, several close options are picked by the manager in one click, and when there is not enough data a clarifying question is asked instead of a guess
• A live catalogue: scheduled synchronisation with the accounting system, so new and discontinued items reach the index without manual work
• The finished basket is transferred into the accounting system with no manual entry

One catalogue and one set of matching rules across every branch.

Results

• The manager checks what was matched instead of looking up every line by hand
• An order of any shape reaches the same single flow
• Doubtful lines are surfaced separately: nothing the system is unsure about goes into an order