Case Study

Blind, then proven.

A revenue-recovery services firm.

Blind, then proven


Problem
2.3 billion rows of flattened, partially orphaned transaction data, with no document boundaries and no foreign keys.
Data
Transaction data across three silos: purchase orders, invoices, remittances, disputes, claims, contracts, and X12 EDI.
Stack
SDO
Time to live
10 business days from deal signed to data flowing.
Outcome
1.5 million candidates scored blind on the firm's own operating definitions (one silo, measured end to end). Parity against the firm's claim history: $209M agreed, $206M identified that was never claimed, $4M the firm claimed that our rules do not raise today. The proof of value became a complete backend replacement, and the firm is proceeding with a full platform build.
Partner read
When a client's data looks unfixable, that's an implementation project, not a subscription.

The test we set up against ourselves

We set out to prove one thing: can independent analysis of raw client data, with no access to the firm's dispute outcomes, find valid opportunities they missed? We designed the test to prove us wrong. Stage one was blind: work only from purchase orders, invoices, shipments, remittances, claims, and contracts, and never look at what the firm had already disputed. Stage two lined our findings up against their actual claim history and measured the delta. Their claims data was held back as a validation target: if our independently reconstructed documents reproduce the stored claim amounts, the reconstruction is correct.

The scope claim stays honest: one retailer silo's vendor data, run on the firm's own operating definitions, their mandate windows per client and their outcome vocabulary. That scope was measured end to end, and the work is complete.

What we were handed

We received 2.3 billion rows of transaction data across three silos. The document layer had been destroyed systematically when the lake was built. Structured transaction documents were flattened into header and line tables, and the boundaries between documents were lost. Foreign keys were removed by design: roughly 430 million invoice lines carry no invoice number. Seven independent systems share no keys. Eight tables were delivered empty at source, including the invoice header and the only PO-level table. We proved the emptiness was at the source, filed a gap report, and requested re-export instead of papering over it.

What we rebuilt

We rebuilt the documents themselves: purchase orders, invoices, receiving records, claims, disputes, and remittances, reassembled per vendor. That includes multi-hop PO-to-invoice matching, which re-attributes the orphaned line rows by walking back through item, UPC, quantity, and cost attributes. Next came the cross-system linkage layer: millions of document headers tying invoices and claims to payments, deductions, EDI rejections, and dispute records across five previously unlinked systems.

On top of that runs the operation on the firm's own definitions: their mandate windows from the contract terms, the purchase-order balance or parent-invoice chain as the unit of dispute, their code definitions. Every candidate carries its source files and the full scoring logic, so anyone can retrace the work. Money types are never summed across categories, and every step of the scoring is open to inspection.

What the numbers say

Parity, on the disputable unit, matched on normalized claim keys (one silo measured):

Parity bucketShare of rowsValueWhat it means
Both found (agreement)37%$209MWe flagged it and the firm claimed it. Replication proof.
Ours only62%$206MWe flagged it and would dispute it; the firm never claimed it. The delta.
Theirs only1%$4MThe firm disputed it; our rules do not raise it today. Calibration targets.

The ours-only pile is large for three reasons, and the composition shows the operation working as designed. A handful of large post-audit billing charges dominate the dollars: nearly half the ours-only value sits in compliance, markdown, and defect-billing charges. High-volume small charges do the rest: returns-defective rows averaging about $117 each, the grinding sweep most claim operations never finish. Last, the firm's claim history is selective, not exhaustive: just over a third of the comparable rows carry their claims. That is the delta thesis.

The theirs-only bucket was diagnosed, not hidden. A rule-library gap accounted for about $59M of claims; we registered 15 claim-evidence rules for those mechanisms, and the charges moved into agreement. What remains is the genuine residual, shown as calibration targets. The excluded rows are shown too: two cash-discount claims from 2020 of about $90M each, both denied, broken out so they cannot inflate the buckets.

Outside the mandate sit three lanes the claims channel never touches. Contractual billings: about $520M of charges that contradict or lack signed terms, worked through a terms audit against the signed agreements. On-time-in-full penalties: about $15M charged, with $1.4M already credited back, which proves the route returns money. Supplier-quality charges: about $2.2M that fail the retailer's own published fee schedule. Claim outcomes are contested by nature: in a sample of the firm's own disputes, 45% moved toward payment and 35% were denied.

What it became

The proof of value became a complete backend replacement, and the firm is proceeding with a full platform build. Nothing here is called recovered: that word is reserved for the firm's own operations, where the team confirms each finding and approved value is measured against these benchmarks. The engagement is complete: blind by design and proven on the firm's own history.

Talk to our engineers first.

Tell us what you integrate and who you've shipped for. Certification starts when your technical leads talk to ours.