Turn a large reconciliation problem into a small review queue.

The ERP can tell you which invoices are open. It cannot tell you which of them are open only because nobody posted the cash.

Answering that needs five sources, not two: the invoices open at the start of the period, the invoices raised during it, the bank file, the cash receipts journal, and the aged receivables report the business is actually reading. The backlog is the gap between what the bank received and what finance recorded, and no single system holds both sides.

Separating the populations is what makes the queue workable. Only 747 items are genuine anomalies carrying $54,969 of variance. The open receivables belong in an aged debt report and cash in transit needs nothing at all.

One population earns its place by what it stops. Where a customer has cash the engine could not allocate, that customer’s open invoices are held rather than chased, because the check may already have settled them. The effect is measured against the true position rather than asserted, in both directions: 95.2% of the collections queue is genuinely unpaid, and 93.4% of the invoices that should be chased are on it. Without the hold precision was 76%, and a collections team would have called customers about $1.29M they had already paid — the exact failure this project exists to prevent, recreated by the tool meant to fix it. Recall is the figure that flatters nobody: $268,548 of genuinely unpaid invoices are not on the list, most of them because a grouped match wrongly closed them.

The hold itself is the least accurate thing here, and it is published that way: 73.4% precision and 76.9% recall, $215,851 held that did not need holding against $184,999 that did. Guessing which invoices a check settled is exactly what the engine has said it cannot do confidently, so the control a supervisor should work from is the account, not the invoice.

For the CFO the output is narrower: the receivables report can be trusted again, $1,311,231 is posted rather than chased, and the two clerks spend their week on the items that need a person.

Python, pandas, and rapidfuzz. The pipeline starts at a file, not a dataframe: extracts are written as CSV with the mess a real export carries — a byte order mark, amounts as text with thousands separators, two date formats in one column — and read back through an intake stage that coerces them, checks control totals against what was declared, validates the schema, and reports everything it had to change. Nothing is cleaned silently.