Automating Document Processing with AI: Invoices, Contracts, Forms
Modern LLM extraction reads messy scans, PDFs, and email attachments without a template per supplier — the detail that killed classic OCR projects. In our projects, 70–95% of documents process automatically; the rest routes to review via per-field confidence scores. At 500 invoices a month and 4 minutes each, that's roughly 33 hours back every month.
Document processing is the unglamorous AI win: nobody demos it on stage, and it pays back faster than almost anything else we build.
Why old OCR failed and LLM extraction doesn't
Classic tools needed a layout template per supplier: field coordinates, keyword anchors, endless maintenance. The supplier redesigns their invoice, extraction silently breaks. LLMs read a document the way a person does — from context, not coordinates. A new supplier's invoice works on day one, a skewed scan usually still parses, and German and English documents run through the same pipeline.
The architecture that holds up in production
Three stages. Intake: a watched mailbox, an upload endpoint, a scanner folder. Extraction: the model returns structured JSON with a confidence score per field. Decision: above your threshold, the document books automatically into the ERP or accounting system; below it, it lands in a review queue where a human corrects fields in seconds.
Two details make or break it: deterministic validation (line items must sum to the total, IBANs must pass their checksum — cheap rules catch what the model misses) and feeding every human correction back into the pipeline as a test case.
How much can you actually automate?
| Document type | Typical automation rate | What stays manual |
|---|---|---|
| Invoices | 85–95% | Disputes, exotic layouts |
| Delivery notes | 80–90% | Damaged scans, partial deliveries |
| Contracts | 70–85% | Clause interpretation, negotiation points |
| Application forms | 75–90% | Incomplete or contradictory entries |
Ranges from our projects. The last 5–15% should stay manual — that's where the judgment calls live, and forcing them through automation is how you manufacture expensive mistakes.
What's the ROI in real numbers?
Take 500 invoices a month at 4 minutes each: about 33 hours of processing every month. At €45 internal cost per hour, that's roughly €1,500 a month or €18,000 a year — before counting fewer booking errors and faster payment runs. A pipeline like this is typically a 4–8 week build, so the math usually works from a few hundred documents per month upward.
The DSGVO part you can't skip
Invoices and application forms are full of personal data. Process in the EU: EU-region endpoints or EU-hosted models, a signed DPA with every provider in the chain, and no customer documents in consumer AI tools. We've written up the full picture in our guide to GDPR-compliant AI.
Document pipelines are core process automation work for us — usually 4–8 weeks from shared inbox to auto-booking.
Olio on LinkedIn
