The AP department that reads its own mail
A nationwide employee screening and compliance company receives invoices from hundreds of clinics and testing providers — every one in its own format, every line needing to be matched to a testing order before it can be paid. We built the intelligence layer that reads them: trained models per vendor, an LLM for the hard cases, and a review screen for the humans who stay in charge.
Every clinic invoices its own way — and every line must find its order
The client's business runs on a network of clinics and labs performing drug tests, physicals, and screenings nationwide. Each one bills differently: different layouts, different item descriptions, sometimes several invoices bundled in one PDF. AP clerks read each document, figured out which internal testing order every line belonged to (matching on donor name and date), translated the clinic's wording into internal item codes, and keyed the result into the ERP. Skilled work — spent almost entirely on reading and retyping.
The clerks weren't the bottleneck. The reading was.Judgment scales fine; transcription doesn't
Three layers of reading, from precise to resourceful
One model doesn't fit hundreds of invoice formats. The parsing stack degrades gracefully — precision first, flexibility as the fallback.
Vendor-trained document models
For the highest-volume clinics — 42 vendors trained so far — dedicated models know exactly where each field lives. Fast, accurate, and improving as volume grows. High-accuracy vendors are candidates for full auto-approval.
General fallback model
Vendors without a dedicated model hit a broad general parser. Lower accuracy by design — it exists so nothing goes unparsed while dedicated models are trained for vendors that earn them.
LLM parsing layer
For complex or shifting formats where structured models struggle, the extracted text goes to a large language model with a structured prompt — and comes back as clean, processable data. The safety net that handles what templates can't.
From inbox to ERP, with humans exactly where they belong
Domain rules, encoded
A "physical" can map to two item codes depending on regulated vs. non-regulated testing — resolved automatically from the order's test ID. One description, multiple codes, correct answer every time.
Built with the clerks, not at them
The clerk workflow was mapped hands-on with the AP team before we finalized screens. The dashboard mirrors their real day: not-reviewed, errors, aging, pending transmission.
The audit trail travels
Every exported line carries a hyperlink back to the exact invoice in the portal — an auditor in the ERP is one click from the source document. No duplicate PDF uploads, no hunting.
Going live now — designed to earn more automation over time
Start with humans reviewing everything. Let the accuracy data tell you which vendors earn auto-approval. That's how AI enters a finance department without anyone losing sleep.Trust is earned per vendor, not assumed globally
Client identity withheld under confidentiality; diagrams recreated for publication. System entering production July 2026 — production metrics to follow.
The same work, productized
Related client stories
2 of every 3 vendor documents, reconciled without a human
The same discipline at 683,000-document scale — with auto-confirmation already earned.
Read the story → AI Document IntelligenceFrom fax machine to front door
AI reading faxed government orders for a one-person fulfillment operation.
Read the story → Order AutomationMatched once, remembered forever
The cross-reference pattern that also powers this portal's item matching.
Read the story →How many hours does your AP team spend reading?
Invoices, statements, confirmations — the reading is automatable. The judgment stays human. We build exactly that split.
Book a Strategy Call