Illustrative case: Accounting firm (18 FTE), Eindhoven · 2026

AI agent for document processing and file assembly

From 9 hours of sorting a day to 14 reviews a day.

Results

Indicative figures

  • 142documents per day processed automatically
  • −9 hmanual sorting per day
  • 96.8%accuracy after the shadow period
  • 10 wksfrom sprint to full rollout

The story

The story behind AI agent for document processing and file assembly

An accounting firm with 18 employees processed hundreds of documents by hand every day. An AI agent now recognizes 142 documents a day, books 90% automatically in Exact Online and queues the rest for review, with a log that satisfies the regulator.

The challenge

Every morning started with sorting: email, client portal, scans and bank exports. Two assistants spent 9 hours a day on it together before an accountant could do anything. Invoices landed in the wrong file, receipts sat untouched and the end of every quarter meant overtime.

From the video: Queue Incoming invoice from Kantoorhuis Brabant with the recognized fields highlighted

An earlier attempt with a standard OCR tool failed on variety: 300 clients, 300 ways to make an invoice, and no connection to Exact Online or the case system. The partners also wanted certainty about GDPR and about who could see the data.

Our approach

We started with a 2-week Automation Sprint on one stream: purchase invoices. An agent based on Claude, hosted via Azure in the EU, reads the document, determines client, supplier, amount, VAT and ledger account, and books in Exact Online through the official API. Below 92% confidence the document goes to a review queue.

From the video: Review Review screen with confidence per field, the AI's reasoning and the proposed Exact Online entry

After a shadow run with 2,000 real documents, accuracy was 96.8%. We then extended to bank statements, sales invoices and loose file documents, with a knowledge base (RAG) on the firm’s own manuals for ledger selection.

Governance got as much attention as the technology: data minimization, processing agreements, no training on client data, a log per decision and a monthly sample check by an accountant.

The solution

An agent that reads the inbox, portal and scan folder 24/7, classifies documents, books or queues them, and completes files. A review screen shows the extracted values next to the original per document; one click approves.

From the video: Overview Receipt with two VAT rates, split by the agent across two ledger accounts

A dashboard shows daily volume, accuracy, lead time and hours saved. Model cost: about €140 per month. The firm spends the freed hours on advisory work.

Watch it in action0:21 · 3 scenes

Project facts

About this project

Something similar in mind? Book a call
Client
Accounting firm (18 FTE), Eindhoven
Year
2026

Delivered

  • AI agent for document processing
  • Exact Online and case system integration
  • Review queue
  • Knowledge base on the firm’s manuals
  • Log and GDPR documentation
  • Dashboard with savings
  • Training and sampling protocol

Technology

  • Claude API
  • Azure OpenAI (EU)
  • Node.js
  • PostgreSQL + pgvector
  • Exact Online API
  • Basecone
  • Microsoft Graph
  • n8n

Illustrative case: composed from comparable engagements. Figures are indicative, not the measured result of one client.

Let’s build something.

Tell us about your project. You will hear back within one business day and receive a concrete proposal within a week.

info@rabzi.nl