Solution Story · AI Document Intelligence

2 of every 3 vendor documents, reconciled without a human

When you buy from 500+ vendors, every purchase order confirmation arrives in a different PDF. Someone has to check that what you're billed matches what you ordered — line by line, price by price. That someone is now mostly software.

The problem

Reconciliation nobody has time to do — and everybody pays for skipping

A distributor buys constantly: siding from one vendor, windows from another, fasteners from a third. Each purchase generates a confirmation PDF that should be checked against the original order — quantities, units of measure, line items, prices, freight. Done manually, it's minutes per document across hundreds of thousands of documents. Skipped, it's silent margin leakage: you pay what the vendor billed, not what you agreed.

The choice used to be: pay people to check everything, or trust everyone and check nothing.
We built the third option.
What we built

Two parsing tracks, one rules engine

The hard part isn't reading a PDF — it's reading 500 different kinds of PDF, reliably, at volume. We solved it with two complementary tracks.

Custom parsers

for vendors with consistent formats

Most high-volume vendors send structurally consistent PDFs. For each, we built a dedicated parser that extracts line items, quantities, units, prices, and freight with precision. Fast, deterministic, cheap to run.

AI document models

for vendors whose documents vary

For vendors whose layouts shift — different templates, scanned documents, format changes — we trained models on Azure AI Document Intelligence. The models generalize where hand-built parsers would break. This is production AI: hundreds of thousands of real documents, not a demo.

Vendor PDFs 500+ vendors 500+ formats Custom parsers consistent formats AI document models varying formats · trained models Reconciliation rules PO · line items · UOM quantity · price · freight Auto-confirmed 389,434 no human touch Review queue 213,353 exceptions, worked by experts exception types routed by cause: line item · unit of measure · quantity · price · freight
Documents flow through the right parser, hit the rules engine, and split: routine confirmations close themselves, exceptions reach humans with the cause already identified.

Exceptions arrive pre-diagnosed

The review queue doesn't say "check this document" — it says "quantity rule violated on 2 lines." Reviewers start at the problem, not at page one.

Rules are configuration

Tolerance thresholds and vendor-specific rules are maintained as settings, not code. The business tunes its own automation rate.

The archive is an asset

683,656 processed documents form a searchable history of every vendor confirmation — pricing disputes get settled with evidence, not memory.

Results

Production AI, measured honestly

683k
documents processed to date
~65%
auto-confirmed, zero human touch
500+
vendors covered and growing
100%
of documents checked — none skipped
The goal was never 100% automation. It was 100% coverage — with humans spending their time only where judgment is actually needed.
Why the review queue is a feature, not a failure