AI invoice processing: from PDF to verified data

What does AI actually do when processing an invoice?

It identifies details and their meaning: supplier, document number, due date, currency and table rows, for example. Downstream software verifies that the values are consistent and the document belongs in the expected workflow.

The distinction matters: a recognised amount is not automatically the correct amount to post. A document can contain an advance payment, discount, multiple currencies or a previous-period figure. The model must also supply the value’s type and location in the source. The accountant can then examine a specific discrepancy instead of rereading the entire attachment.

Start with the output format. Define required fields, data types and what an empty value means. Zero, missing and unreadable are three different cases. If the import cannot distinguish them, errors move into accounting, where they are harder to find.

How do OCR, AI and ordinary rules differ?

OCR reads characters, AI helps identify meaning and rules perform exact checks. They complement one another. A language model need not calculate every total, and standard document recognition need not be reinvented.

The current Azure Document Intelligence documentation describes a prebuilt invoice model including line items and extraction confidence. A confidence score is not an accounting check, however. A correctly read value can still be commercially wrong.

What do real-world results demonstrate?

For automation provider Ellby, AWS reports an increase in invoice automation from under 60% to over 94%, savings of over 300 hours of manual maintenance per month and an onboarding reduction of over 55%. This is a vendor case study. Maintenance time saved is not the same as accountants’ time per invoice, and the automation rate is not a declared accuracy rate for every field. AWS and Ellby case study.

“We've made our automation smarter and easier to scale.” — Steven Chenery, Ellby CEO, shortened quotation from the case study.

The study notably lists detailed line-item extraction as future development. Ask exactly what any quoted percentage covers. An invoice with its number and total extracted may still require substantial work on individual items.

What do practitioners find most difficult?

Document variability and exceptions. The author of a hospitality project on Reddit discusses varied PDFs and credit notes. The post seeks a collaborator, rather than independently testing a tool. It still illustrates the gap between a clean demo invoice and a real inbox.

In the n8n Community in April 2026 a finance-workflow author discusses duplicates, currencies and uncertain data. Our practical takeaway is simple: the test set must include more than the supplier’s prepared demo documents.

Include a phone photo, a multi-page document, a corrected invoice, a credit note, two attachments with the same number and an invoice with changed bank details. Define the expected result for each in advance. That makes a more useful comparison than claims that a model “understands accounting”.

How can we prevent duplicates and incorrect imports?

Record every receipt as an event and identify documents using a combination of fields. A filename alone is unreliable. The same invoice can arrive in two files, while different companies may use the same invoice number.

Retain the original document, its fingerprint, extracted data, corrections and accounting-system identifier. If an import connection fails, check whether the record already exists before retrying. A blind retry can create a duplicate while the user sees only an error message.

A supplier bank-account change is not an ordinary text update. Verify it separately using the established business process. Attachments also must not change agent permissions: a sentence in a PDF is not an instruction from the system administrator.

How can we tell whether the automation pays off?

Measure completed documents correctly accepted by accounting, not just files read. Track review time, returned documents, post-import corrections and operating costs including support.

Illustrative calculation: 800 invoices a month at four minutes of transcription each total roughly 53 hours. One minute of remaining review per document uses another 13 hours. Add exceptions and administration. The difference does not automatically reduce staff costs; it can free capacity for closing or earlier processing.

Evaluate accuracy by field and consequence. A wrong due date has different effects from a small typo in an item description. A single average score hides those differences.

How can we start without rebuilding the accounting system?

Choose one receiving mailbox and one document set. Run processing alongside the current workflow and compare results first. Then enable controlled import for unambiguous cases. Unclear documents need a separate queue and a responsible person.

Before going live, ask to see corrections, repeated receipt, failed imports and retrieval of the original document. A functioning pilot covers these situations too. Once extraction is proven, you can extend it with AI for accounting: matching, closing documents and explaining discrepancies.