AI Document Processing Desk

Finance and operations teams retype supplier documents into sheets and accounting tools. This build starts by measuring that manual work, then automates what the numbers justify.

[

Challenges

]

Documents arrive by email, WhatsApp and upload in many layouts. Manual entry is slow, errors reach the books, and duplicates and mismatched totals are found late.

[

Solutions

]

A pipeline classifies each document, extracts structured fields with a confidence score, checks totals, dates, vendors and duplicates, and approves clean documents automatically. Exceptions go to a review screen with the document beside the data, and every change is logged.

What the system does

• Starts with a short automation audit: the manual process is timed on a sample of documents, giving a measured baseline.

• Receives documents by email, WhatsApp and upload, and detects duplicates by file fingerprint.

• Classifies each document as an invoice, purchase order, receipt or other.

• Extracts vendor, document number, dates, line items, tax, totals and currency, with a confidence score for every field.

• Validates that totals match line items plus tax, that dates make sense, that the vendor matches a master list, and that the document is not a duplicate.

• Approves documents automatically when every rule passes and confidence is high, and sends the rest to a review queue.

• Shows the document beside the extracted data, highlights low-confidence fields and saves every correction as an example to improve later extraction.

• Exports to Google Sheets or a webhook, and keeps an audit trail of who changed what.

Stack

n8n · Supabase (Storage, Postgres, Auth) · Next.js review interface · a vision-capable language model · Google Sheets API · Docker on ETHIOR’s own server.

Status and results

Designed, built and tested by ETHIOR. This is a reference build, not a client engagement. It is currently in development. The figures above are design targets, not results. They will be replaced with measured results, with sample sizes and dates, once the system is running.

Limits

Accuracy will be measured on synthetic and redacted documents, and real supplier layouts will need tuning. Human review is part of the design on purpose.