Gold purchases entered into SAP without retyping anything

The starting point
Precious metals, and a pile of documents with every purchase
For sixty years, Boucledor has made buckles and clasps for the great high-luxury watch brands. The house buys its gold and silver from several suppliers.
Each purchase produces an order confirmation, then a delivery note, an invoice, an assay certificate and a metal statement accurate to the gram. They arrive as PDFs or on paper, each with its own layout, and all must end up in the same form: a line in SAP, the house's management software.
What it cost
Ten-digit identifiers, copied by hand
Two people handled it, each on one type of document, on top of their usual work. For each document, they had to open it, read it, retype five to ten fields into SAP, then check everything again.
A typo gives no warning: one swapped digit in a reference would be enough to enter the wrong weight of gold into the accounts.
The solution
A table ready to import, whatever the layout
Each person drags their files into their page, in batches and unsorted: sixty at once if need be. An AI recognises each one's sender and type, and picks out the values; the system also checks that nothing is missing from the file.
The import into SAP is still done by hand, by downloading the table: automating it is the next step.
Avant / après
Avant
- Every file opened and read, one by one
- Five to ten fields retyped into SAP
- Reading everything again to spot a typo
- A missing document only showed up when you looked for it
Après
- Everything dropped in batches, unsorted
- No field to retype
- No more typos possible
- A missing document flagged on screen
Key decision
The AI copies, and code does the calculating
Depending on the sender, an order number is written with four or five digits, sometimes followed by a suffix Boucledor doesn't want. We first asked the AI to keep only the useful part: it started inventing numbers.
The rule became absolute: the AI copies what is printed, deciding nothing. Trimming a number or adding up metal weights is then done by code, which always gives the same result. On the hundred or so real documents provided for testing, the checks pass at 100%.
Key decision
Why not use an off-the-shelf extraction service?
These papers carry what the house protects most: its purchase prices, its partners, the quantities of metal moving through it. Entrusting them to an off-the-shelf service meant sending them outside Switzerland.
So everything runs at a Swiss host in Geneva, AI models included. And no standard product could read these industry-specific formats.
Key learning
A new supplier mustn't become a new project
The list of senders was going to change, and so were their layouts: a tool cut for today's formats would have been obsolete at the first change. So the system is made of one module per document type.
For a new format, we describe it and test it on a few real examples. That's what made it possible to widen the scope along the way, without reopening the project.
Testimonial
« What we're trying to tackle with Exponent is waste. Spotting operational tasks that don't add enough value, and saving the teams time every day. We'll quickly be looking at 200, 300 hours a year saved. »
Supplier documents costing you hours of data entry? Let's talk.
We start with a conversation, not a quote.
Discuss a similar project