Are my invoices sent to the OpenAI cloud?
+
During PDF OCR import, the PDF contents are sent to the OpenAI vision API (GPT-4o) for field extraction. OpenAI declares GDPR compliance and does not store API inputs for model training (Data Processing Addendum, as of 2026-05). Extraction never runs on its own — it only runs when you attach a file to an expense, so for sensitive documents you can simply fill the fields in manually with no attachment.
Which AI models does 1FAKTURA use?
+
For OCR extraction from PDFs and images we use OpenAI GPT-4o (we update the model family alongside OpenAI releases). For expense categorization and line-item suggestions we use smaller and faster models (typically GPT-4o-mini) which give sufficient accuracy at a fraction of the cost. AI calls are always server-side from our infrastructure; we never hand the OpenAI key to the client.
How accurate is AI extraction?
+
On standard Slovak invoices (Pohoda, Money S3, KROS exports, or PDFs from cloud invoicing tools) accuracy is 90 – 95 % on all core fields (IČO, amounts, rates, date). On hand-written or unusual formats it drops to 70 – 85 %. As a user you always review before saving — AI only pre-fills, never automatically commits final values.
What if AI fills a field wrong?
+
When pre-filled fields show, you simply edit them manually and save — your value is what gets stored; the AI never overwrites it. When the VAT arithmetic, the IBAN or the IČO on the document don't check out, 1FAKTURA lowers the extraction confidence and routes the document to manual review, so you see the problem before saving. For documents where you'd rather not use extraction at all, fill the expense in manually with no attachment.
Are AI features included in the plan price?
+
Yes, AI features are included in the Lite plan and higher with no extra charge for AI calls. Internally we measure AI call costs per customer and at 10 × average usage we negotiate a custom enterprise rate. For most sole traders and SMBs the cost of AI calls is negligible.
What's the difference between AI and classic OCR (Tesseract)?
+
Classic OCR (Tesseract) extracts text from an image — raw text without structure. AI vision models (GPT-4o) understand context and return structured fields (IČO = X, amount = Y) directly. For Slovak invoices with varied layouts, the AI approach is 5 – 10× more accurate than Tesseract + regex rules combined. For internal debug we keep Tesseract as a fallback when an AI call fails.
Are you planning AI generation of the entire invoice?
+
Not in the short term. AI is a copilot for us, not an autopilot — the final decision (which VAT rate, what amount to invoice, what due date) stays with the user. AI can suggest, pre-fill, categorize and flag anomalies, but invoicing is a legal act and responsibility for its correctness lies with the entrepreneur. This stance stays in 2027.
Does AI work for cross-border invoices?
+
Yes, extraction handles invoices issued in Czech, English or German — the model is not limited by language. The VAT ID field, however, expects the Slovak format (SK + 10 digits) and Slovak rates, so a foreign VAT ID and a foreign rate are things you fill in yourself. The tax regime — § 15, § 43 or reverse charge — is not suggested by the AI; that decision stays with you.