We auto-classify Sinsta Presents' overseas (English) corporate card transactions with a deterministic rule engine, and surface only the ambiguous "hold" cases to staff — with grounding from card role, past history, and Google Drive approval-request matching. Card numbers are never stored on the server.
English merchant-name transactions were especially hands-on for the settlement staff.
Because merchant names were in English, staff had to determine the purpose and account for each one — and the same merchant split by purpose across transactions (the same Amazon could be parts = materials cost vs. a business-trip purchase = travel cost).
The record of what purpose, by whom, and under which approval a transaction was handled was scattered, making consistency, reproducibility, and audit trails hard to secure.
Card detail files contained the full card number, raising exposure and security (PCI scope) concerns when uploading them to an automation tool.
Staff upload the card statement as-is; clear-cut cases are organized automatically, and only the ambiguous ones are confirmed with grounding.
Upload the original card statement Excel (.xls/.xlsx) as-is to /labeling/.
Truncates the card number to the last 4 digits and extracts only English transactions.
Auto-classifies by deterministic rules, splitting off only ambiguous ones as "hold."
Recommends a purpose with approval, card-role, and history grounding.
After confirmation, download the labeled Excel (card number masked).
Classifies with priority rules instead of an LLM — reproducible, auditable, and with no call cost.
On upload it immediately truncates the card number to the last 4 digits and discards the full PAN column. The principle: "not storing it is the safest."
**** **** **** 0863For ambiguous transactions, it attaches a recommendation and grounding badges by priority (approval amount match > card role > approval brand > brand history > heuristic).
Snapshots and parses approval-request PDFs in Drive, matches overseas commerce transactions by brand, amount, and date, and links back to the original approval request.
Clear-cut cases are organized automatically, and staff only confirm the ambiguous hold cases while looking at the grounding.
Reproduced with actual card statements, most English transactions were auto-processed without misclassification.
In real-data reproduction, 104 of 119 English transactions were auto-classified, and the match rate for auto-classified cases was 100% (0 misclassifications). Staff handled only the 15 hold cases.
The full card number is never left anywhere in the system — only the last 4 digits are kept. Output is masked and Drive is accessed read-only, minimizing exposure risk and PCI scope.
Classification by a deterministic rule engine, with only approval-request parsing by LLM. Reproducibility and cost, both handled.