Every document goes through two, from different vendors. Two runs of one model make correlated mistakes — a second opinion is only worth having if it can be wrong differently.
Pages this period
17
Indicative reading cost
€1.70
both readers, this volume
Structured, never read
3
arrived as e-invoices
Screens ask for a capability, never for a vendor. This table is the resolver.
document.ocr
Text and layout
document.classify
Document type
document.extract
Field extraction
document.tables
Line items
einvoice.parse
e-Invoice parsing
The default first reader. Strong on printed invoices with a defined layout.
https://documentai.googleapis.com/v1
Invoice parser
invoice-parser
€0.09
~4s
Returns line items as a table, which is what lets the arithmetic close a document.
Document OCR
ocr-processor
€0.00
~2s
Second opinion on every document. Holds up better on photographs than on scans.
https://{resource}.cognitiveservices.azure.com/documentintelligence
Prebuilt invoice
prebuilt-invoice
€0.01
~5s
The default second reader. A different vendor to the first on purpose — two runs of one model agree on the same wrong digit.
Read
prebuilt-read
€0.00
~2s
Alternative second reader, or a third when the first two cannot agree.
https://textract.eu-central-1.amazonaws.com
Analyze Expense
AnalyzeExpense
€0.01
~4s
Built for receipts and invoices specifically. A useful third opinion on the crumpled ones.
Cheap enough to read everything, including the documents nobody would have scanned.
https://api.mistral.ai/v1
Mistral OCR
mistral-ocr-latest
€0.00
~3s
European provider and an order of magnitude cheaper per page — the volume option for a shop taking forty delivery notes a day.
Runs on your own hardware. No per-page cost and nothing leaves the building — the option for a client who will not send invoices to a cloud.
Tesseract 5 (hrv)
tesseract-5-hrv
no per-page cost
~6s
Needs the Croatian language data installed, and reads text rather than understanding documents — it never produces named fields, so it cannot be one of the two readers.
Structured invoices off the wire. The share of documents arriving this way is the single best predictor of how little work a month will be.
UBL 2.1
ubl-2.1
no per-page cost
~0.2s
Not a reader. The invoice arrives as structured XML, so there is nothing to read and nothing to settle — every field is exact and both readers are skipped. Every document that arrives this way is one that can never be misread.
Prices are indicative and move often; they exist so an estimate before a thousand-page month is the right order of magnitude, not so anyone bills against them. Keys are the workspace's own — nothing here is resold.