A Czech wholesale distributor processing 800+ invoices per month replaced expensive per-document SaaS fees with a custom GPT-5.3 vision system, cutting costs by 87% and keeping all data on their own servers. The AI pipeline handles full invoice extraction including line items, automatic cost categorization, and direct posting into ABRA Flexi. What used to take 67 hours of manual data entry each month now runs with minimal human oversight, freeing the accounting team to focus on analysis and financial planning.
Wholesale distributionThe client is a mid-sized Czech wholesale distributor supplying industrial materials to businesses across Central Europe. They process over 800 supplier invoices every month in a mix of formats: PDF attachments, scanned paper documents, and photographed receipts. Their accounting runs on ABRA Flexi, which serves as the central system for bookkeeping, accounts payable, and financial reporting. Before reaching out to Bitvea, the company had tried automated invoice processing through a commercial SaaS OCR platform but found the per-document pricing unsustainable and the accuracy insufficient for their accounts payable automation needs.
The company received 800+ supplier invoices monthly in mixed formats: PDFs, scanned paper, and email attachments. Each invoice passed through up to 8 people before it was booked: an assistant scanned and logged it, managers verified and approved, a junior accountant entered the data, and a senior accountant reviewed everything. The accounting team spent roughly 67 hours per month just on invoice data entry, equivalent to nearly one full-time employee doing nothing but copying numbers from paper into the system. On top of that, they were paying a commercial SaaS OCR platform per document, with 85% extraction accuracy that still required manual review of every sixth invoice. The lack of direct integration with ABRA Flexi meant data had to be exported as CSV files and imported manually, introducing another source of errors and delays in the accounts payable workflow.
We built a custom AI invoice processing pipeline powered by the GPT-5.3 vision model. Invoices arrive by email or upload, and the AI extracts all fields: supplier details, amounts, line items, VAT, due dates, and payment information. Validated data flows directly into ABRA Flexi via API with no manual steps in between. The system also handles automatic cost categorization and duplicate detection, so the accounting team only reviews flagged exceptions. The entire system runs on the client's infrastructure, and the per-invoice API cost is approximately $0.04 including line item extraction, compared to $0.21–$0.42 with the previous SaaS platform.
The AI reads invoices in any format (PDF, scans, photos) and extracts all fields including individual line items, with 95%+ accuracy on the client's specific layouts. The invoice OCR model was fine-tuned on real samples from the client's top 50 suppliers, which account for over 90% of monthly volume. New supplier formats are learned automatically after a few manual corrections.
The AI assigns each invoice to the correct cost center, spending category, and GL account based on learned patterns from historical bookings. This eliminates manual classification entirely for recurring suppliers. For new or ambiguous cases, the system suggests the most likely category and flags it for quick human confirmation.
Validated invoice data is pushed directly into ABRA Flexi via the ABRA Flexi API, with no CSV exports or manual imports needed. The integration handles supplier matching, creates new supplier records when needed, and posts invoices with full line item detail. This closed the gap that previously required manual data transfer between systems.
Built-in checks run against ARES (Czech business register) and VIES (EU VAT validation) to verify supplier details automatically. The system also performs duplicate detection, VAT calculation verification, and cross-references supplier data against existing records in ABRA Flexi. Any discrepancies are flagged for human review rather than silently passed through.
The entire system runs on the client's servers using Docker containers. No invoice data leaves their infrastructure, which was a firm requirement from the client's management. Updates and model improvements are deployed through a standard CI/CD pipeline without disrupting daily processing.
The project started with a two-day workflow audit where we mapped the entire invoice lifecycle, from receipt to final posting in ABRA Flexi. Based on that, we designed a modular AI pipeline with separate stages for document intake, extraction, validation, and ERP integration. We trained the extraction model on 200+ real invoices from the client's most frequent suppliers, iterating on accuracy until it consistently hit 95%+ on their specific formats. The final stage was integrating directly with the ABRA Flexi API and running two weeks of parallel processing alongside the existing manual workflow to verify results before going live.
Timeline: 6 weeks from kickoff to production
Commercial platforms charge you per document and process your data on their servers. Here's how a custom AI pipeline compares to Rossum and Digitoo.
Cumulative cost over 5 years: SaaS per-document fees vs. a custom GPT-5.3 vision pipeline
Pricing as of March 2026. Rossum AI Starter from $18,000/yr · Digitoo 5–10 CZK/invoice · GPT-5.3 vision $1.75/$14 per 1M tokens
Custom solution from ~$2,000. Ongoing: ~$35/mo API costs (800 invoices at ~$0.04 each) + ~$50/mo hosting & maintenance.
Invoice processing is not just about OCR accuracy. The real value is in the hours your team gets back when they stop copying numbers from paper into the system.
Manual data entry, verification, and classification per invoice, eliminated by AI extraction and auto-categorization.
For 800 invoices per month, that's nearly one full-time employee redirected from data entry to analysis and advisory work.
Without automation, each invoice passes through assistants, managers, junior accountants, and senior reviewers before it's booked.
Time savings data based on Digitoo automation study and industry analysis of manual invoice processing workflows.
Training the AI model on real client invoices rather than generic samples was the single biggest factor in reaching 95%+ accuracy. Generic OCR models struggled with the variety of Czech invoice formats.
Running the new system in parallel with the existing manual process for two weeks built trust with the accounting team and caught edge cases before they became production issues.
Direct API integration with ABRA Flexi eliminated an entire category of errors that came from manual CSV imports. The time saved on error correction alone justified the integration effort.
On-premise deployment added complexity to the initial setup, but it removed ongoing concerns about data privacy and gave the client full control over their invoice data.
“We used to have one person spending most of their week just typing invoice data into the system. Now the AI handles it in seconds and our accountants only step in when something is flagged. The cost savings were obvious within the first month.”
Turn paper invoices into structured data, automatically.
Learn more