Recovering Retail Deductions at 70% Win Rate With LLM Document Intelligence

Deductions were running at a few points of wholesale revenue and climbing as the brand added retail accounts. Each one came as a different artifact: an 812 with a numeric reason code, a scanned PDF with claim backup, a spreadsheet exported from a retailer portal. To dispute one, someone had to read it, decode the retailer's reason code, dig the proof out of the order-to-cash system (was the ASN actually late? was it really short-shipped?), and write a packet. At volume, that was impossible, so AR disputed the easy few and absorbed the rest.

The money was recoverable. The reading was the bottleneck.

The governing rule: every deduction is read, classified, and scored for validity against the order-to-cash record, and a human approves before anything is disputed. The LLM does the reading and the assembly; it does not file claims on its own.

Two decisions shaped it. First, normalize each retailer's reason-code dialect into one canonical deduction taxonomy, so a "shortage" from one retailer and a "carton compliance" from another land in the same model. Second, ground every extraction in its source: the model returns the exact page and quote it pulled each figure from, so a disputer (and the retailer) can see where the claim came from. We did not let the model invent recoverability. It scores, a person decides.

What we deliberately did not do: no auto-submission of disputes, no touching the ERP's books directly, and no offshore claims team. The system makes the work tractable; the AR lead still owns the call.

Deduction documents land from EDI 812 feeds, retailer portals, and email, and are stored in S3. Claude Haiku 4.5 reads each one (using vision for the scanned PDFs), extracts the line items, and classifies the reason against the canonical taxonomy, returning the source page for every value. A match step joins each deduction to the order-to-cash record from the EDI build (ASN timestamps, proof of delivery, invoice, shipment) and scores validity. For the disputable ones, Claude Opus 4.8 drafts an evidence-backed dispute packet. Everything surfaces in a TypeScript review console where AR approves, edits, or rejects, and Temporal runs the dispute workflow: submit, await the retailer's response, escalate, and record the outcome.

It reuses the order-to-cash data the EDI build already produces, so every deduction is scored against real shipment evidence rather than a guess.

Every deduction is now triaged instead of a sampled few, and disputes that used to never get written go out with the evidence attached. Win rate on disputed claims runs around 70%, recovering deductions worth roughly two to three points of margin that had been leaking away. Turnaround is about three times faster because the reading and packet assembly are automated. And because every deduction is classified, the recurring root causes surface, for example a pattern of ASN-timing chargebacks from one retailer, which feeds straight back into fixing the order-to-cash process so the chargeback stops happening.

The deduction data is a map of where the operation leaks money. Feeding the recurring root causes back into the EDI and routing build (the same brand's order-to-cash system) turns recovery into prevention: the chargebacks that keep winning disputes are the ones worth engineering out entirely.