Most document automation demos show you the happy path. The failure that costs money is different: the model returns a plausible wrong answer with full confidence, nobody notices, and it lands in a document that has already gone to a customer.

So here are numbers from a pipeline I built. 32 line items across three different document layouts. 28 matched automatically, 4 escalated to human review, 0 wrong matches.

The four escalated lines are the design working, not failing. One item existed in two variants scoring identically, so it stopped and showed both instead of picking. One was not in the source catalogue at all, and nothing was invented to fill the gap. One had a correct code but a unit conflict between the request and the catalogue, so the price was withheld and the quantity flagged. The rule throughout: a tie is not a winner, a tie is a question.

What you get: PDFs in, by upload, email, webhook or messenger, and structured rows out into Excel, Google Sheets or a database. Rate-limit handling and retries, because the provider will throttle you and a burst turns into a run that is half complete and looks finished. An explicit path for documents the pipeline could not read, so they get named in the summary rather than silently skipped or silently wrong. And the workflow JSON, readable code, and a handover note, so the pipeline is yours and not mine.

I will tell you before you order if your documents are scanned rather than text-based. That needs OCR first, it is a separate step, and finding out afterwards is how these projects go wrong.