AI can turn a scanned invoice, receipt or form into a record your accounting or ERP system can use, but the model reading the page is the easy part. What decides whether the project works is validation: arithmetic that must add up, check digits on tax numbers, lookups against your own supplier and order data, and a person who sees only the documents that fail. Measured on real invoices, pipelines built this way pass a large share of documents without a human touch while keeping errors inside a stated budget. In Greece there is a twist: since 1 October 2026 invoices between businesses are issued electronically, so the smart pipeline starts from structured data wherever it exists and keeps AI reading for what still arrives on paper or as a PDF.
What does AI document processing actually do?
AI document processing turns a document into a validated record in stages, and each stage owns one decision:
- Intake: store the original file unchanged, with a fingerprint that later detects duplicates.
- Classify: decide what the document is, an invoice, a credit note, a delivery note or a contract, because each type has its own fields and rules.
- Extract: read the fields into a fixed format, never into free text.
- Validate: check every value that can be checked.
- Route: pass verified records through and send the rest to a person.
- Commit and log: post once, even on a retry, and keep the file, the values, the model version and every human change.
Why the extraction step should return a fixed set of fields instead of prose is explained in structured outputs.
How accurate is AI at reading invoices?
AI reads clean, repeatable invoices very well and messy images much less well. On a set of Spanish electricity invoices, a frontier model reached 97.6% F1, and giving it a few examples beat asking with none by more than 19 points (electricity invoice study, 2026). Across 21 models reading structured data from document images, the best reached 67.2% exact accuracy on the images against 83.0% on the same kind of records given as text, while almost every output was perfectly formatted (Structured Output Benchmark). That last detail is the trap: the answer always looks right. And script matters: on business documents in Khmer, feeding the model text from a separate OCR engine beat giving it the page image by about ten points. Greek is also a non-Latin script, so test both routes on your own documents before assuming the model should read the image directly.
Why does validation matter more than the model?
Validation matters more because it is the only step that turns a plausible number into a trusted one. Five kinds of check, all written in code rather than in the prompt:
- Arithmetic: the lines add up to the subtotal, the VAT matches its rate, the total matches both.
- Check digits: a Greek ΑΦΜ has a modulus check and an IBAN has its own. A failed check is a failed extraction, not a low-confidence value. Reject an all-zero ΑΦΜ explicitly, because it passes the arithmetic check in a popular open-source validator.
- Lookups: the supplier exists in your supplier table, the purchase order exists and is open.
- Dates and money: 03/04/2026 is 3 April in Greece and 4 March in the United States, and both parse without an error, so parse with an explicit format and store amounts in cents.
- Business rules: the VAT rate is legal for that supplier and date, and totals above a limit carry an approval.
A field that nothing could check should be marked unverified, and nothing downstream should act on it silently. The same logic runs through why AI hallucinates.
Can the model tell you how sure it is?
No: asking the model how confident it is does not work for deciding what a person must check. Two 2026 studies found that the model's own confidence, whether stated in words or read from its token probabilities, collapses toward approving almost everything at useful thresholds. What works is combining signals from different places: whether two differently built extractors agree, whether the value passed validation, and how clean the page image and its OCR were. Combined that way, one study auto-approved 49 to 72% of fields while keeping errors in the approved group under its target (perception, layout and validation study), and another reached 99.1% accuracy on the 80% of invoice fields it was most sure about (multi-signal confidence study).
How do you measure success?
You measure success as the share of documents that go through with no human touch and turn out correct, always stated together with an error budget. "72% straight through" means nothing on its own. "72% straight through at under 1% field errors" is a claim someone can check. The fastest way to raise it is rarely a bigger model: add validations, add few-shot examples from your own corrections, clean the supplier master data and classify documents more finely, and only then upgrade the model. A person reviews what fails, and that review screen is part of the product, as described in human in the loop.
What does it cost to read a page?
Reading a page costs little on a small model and much more on a large one at high resolution. By my arithmetic at September 2026 list prices, a dense invoice page sent to Claude comes to about 3,500 input tokens, roughly 3.50 US dollars per thousand pages on its smallest current model and half that through the batch interface, against roughly 34 dollars per thousand pages on the largest model at its highest image resolution (Anthropic, PDF support). The output adds more for long line-item lists. The practical rule is to run the cheap tier by default and send only the pages it cannot settle to the expensive one.
What changes with e-invoicing in Greece?
E-invoicing changes where AI reading is needed. Since 1 October 2026, invoices between Greek businesses are issued electronically and their data reaches myDATA in structured form, so a pipeline should use that structured data wherever it exists and never re-read a PDF rendering of it. AI reading remains the right tool for what is still unstructured: foreign supplier invoices, retail receipts for expenses, contracts, forms, handwritten notes and archives. The EU is heading the same way, with cross-border digital reporting due from July 2030 (VAT in the Digital Age). How the books are reconciled against myDATA is covered in myDATA and e-invoicing.
Frequently asked questions
Can AI read Greek invoices and receipts?
Yes, but test it on your own documents. Printed Greek reads well on current models; faded thermal receipts, stamps and handwriting do not, and those should go to a person or a second extractor.
Do I still need OCR software?
Sometimes. For clean digital PDFs, the text layer is enough. For scans, compare a vision model with a dedicated OCR engine plus a text model on a sample, and consider running both, because their disagreements are the best signal of what to check.
How long must scanned documents be kept?
As long as tax law requires for the original, which in the EU is set by each member state and commonly runs for years. Keep the original file, the extracted values and every change together, in a format you can read without the vendor.
If invoices or forms still arrive on paper or as PDFs in your business, send me the document types and monthly volumes. Which of them can go straight through, and which will always need a person, usually shows from that list. The finance side is described in AI for finance and accounting.