AI transparency (draft)
How Faktuwa uses an AI model to read data from invoices, and what its limits are.
Last updated: 14 October 2026
Draft version. This text is under legal review, and square brackets mark details still to be completed. This English version is a courtesy translation of the Polish text, provided for information. Once approved, the Polish version will prevail if the English and Polish versions differ.
What the AI does
After launch, the converter is planned to use an artificial intelligence model to read the invoice (PDF or photo) and propose values for the fields needed for the XML file: numbers, dates, seller and buyer details, items and amounts.
What the person does
Each proposed field is shown on the "Review and correct" screen. Before the XML file is downloaded, you must review and approve the data. You can correct any field. Fields with low reading confidence are marked.
The XML file is not generated without your approval.
Limitations
• The model may make mistakes, especially with a poor-quality scan, an unusual invoice layout, handwritten notes or an unreadable font.
• The Service does not check whether the invoice is genuine, whether the transaction took place, or whether the VAT rates are correct for your sales. These decisions are yours.
• Technical validation (conformance with the FA(3) schema, NIP checksum, net/VAT/gross calculations) catches some errors, but not all of them.
• We do not give an accuracy rate. The results of testing on a set of invoices will be published only after they have been measured.
Model provider
Invoice reading is carried out by a model provided by Anthropic. Parts of the invoice content are sent to that provider for the purpose of reading.
Details of data processing are in the Privacy policy.
Reporting errors
If you notice an error in the proposed data, correct it on the review screen. If an error appears in the file after download, report it to [E-MAIL KONTAKTOWY].
Date
Draft of 14 October 2026. Last updated: [DATE].