Document AI that reads 1 300 supplier invoices a day

The problem was not “recognise text” but “never post a wrong amount”. Classic OCR reads characters; it does not understand that “Total” on one supplier’s layout sits in the bottom-right table while on another it is a line in the footer. Template-based extraction breaks every time a supplier changes its format — which is why Gartner (2023) describes the market moving from template OCR to intelligent document processing that combines layout models and language models. For the finance team the real requirement made confidence estimation and human-review design as important as the extraction model itself.

The operator sees the scan and the extracted fields side by side. Green fields are above the auto-post threshold; amber ones need a click. Line items are reconciled against the totals before anything reaches the ERP.

It was built as a queue-based pipeline: intake and image normalisation, a multimodal model that returns typed fields with a per-field confidence, reconciliation against master data, a human review queue for anything below threshold, and posting to the ERP with the scan attached. Human-in-the-loop is by design, not a fallback: Deloitte’s global intelligent-automation survey (2022) found that most organisations pursuing automation still combine it with human checks, and the finance team’s tolerance for a wrong posting was zero.

• Intake. Documents arrive from network scanners, a mailbox and a shared folder; a queue normalises images (deskew, dewarp for phone photos, page splitting).
• Understanding, not just OCR. A multimodal model reads the document as a whole and returns counterparty, VAT ID, number, date, net/VAT/gross and line items as structured JSON with a per-field confidence and the source region on the page.
• Reconciliation. Extracted values are matched against the supplier master, contracts and purchase orders; line items must sum to the totals. A document that fails any check is routed to review with the failing rule highlighted.
• Review queue. The scan and pre-filled fields side by side; the operator confirms or corrects. Every correction is stored as training data.
• ERP posting. Confirmed documents are posted through the ERP API with the source scan attached; each document keeps a full audit trail from intake to posting.

The table shows why the project did not start from template-based OCR. Templates work when layouts are few and stable; here there were a dozen layouts per supplier and they changed. The multimodal model reads the page as a whole, so a new layout is a new example, not a new template — and it returns a confidence per field, which is what makes a review threshold possible.

The straight-through rate — documents posted with no human touch — started at 58% on launch and rose month by month as the model was retrained on operator corrections. Field accuracy is measured on a monthly audit sample against the ERP.

Twelve percent of documents still reach a human — but a review now takes 38 seconds instead of two minutes of typing, because the operator confirms rather than transcribes.

Python workers on RabbitMQ; a multimodal LLM for document understanding with a fine-tuned layout model for region detection; PostgreSQL for documents, fields and the audit trail; a Vue review interface; ERP integration through the vendor API. Runs in Docker on the client's infrastructure; documents are processed inside the corporate perimeter.

This approach is for any finance, logistics or insurance team with a high flow of documents in diverse layouts: invoices, delivery notes, customs declarations, insurance claims, loan applications. The pattern is always the same — extract with confidence, reconcile against what you already know, and let people handle only the exceptions. The regulatory direction is the same: structured e-invoicing has been mandatory in EU public procurement since Directive 2014/55/EU, and the VAT in the Digital Age package adopted in 2025 extends digital reporting to cross-border B2B invoices over the coming years, so a pipeline that already produces typed, reconciled data is a step towards that rather than a stopgap.