AI can read invoices, forms and PDFs and return the fields you care about as structured data, such as supplier name, due date and total. It works best when OCR (optical character recognition) first turns the page into text and positions, and a model then decides which text means what. It is not a set-and-forget process: every extraction needs validation rules and a person for the cases the system is unsure about.
What “document AI” actually does
Most document tools run in layers, and it helps to know which layer you are buying.
- Text recognition. OCR reads typed (and in some products handwritten) text from scans, photos and PDFs. Amazon describes Textract as detecting typed and handwritten text in documents such as financial reports, medical records and tax forms.
- Layout and structure. The tool works out which words form lines, tables and form fields.
- Field extraction. A trained model labels the content. Microsoft’s prebuilt invoice model analyses invoice text and returns customer name, billing address, due date and amount due as structured JSON, along with line items.
- Flexible questions. Some services let you ask for a specific value in plain language. Textract calls this Queries, and it sits alongside separate operations for forms, tables, and for invoices and receipts.
- Your business rules. This layer is yours: checking totals, matching purchase orders and routing exceptions.
A general-purpose language model can also read extracted text and answer questions about it. It is flexible, but it can produce a confident answer that is not on the page. For money and compliance fields, prefer a tool that returns values tied to a location on the document, so a reviewer can see where each value came from.
Where accuracy breaks down
Extraction quality depends heavily on the input. The Microsoft documentation gives concrete requirements that illustrate the point: provide one clear photo or high-quality scan per document, remove password locks from PDFs before submitting, and make sure text is large enough (a minimum of 12 pixels high on a 1024 by 768 image, which it equates to roughly 8-point text at 150 dots per inch).
Common failure points, beyond those stated limits:
- Poor scans: skew, shadows, folded pages and fax-quality images.
- Unusual layouts: a supplier whose invoice puts the total somewhere unexpected.
- Ambiguous labels: “Date” might be an issue date, a service date or a due date.
- Multi-page tables: line items that continue across pages.
- Handwriting and stamps: these overlap printed text and are harder to read reliably.
Treat any quoted accuracy figure from a vendor as a claim to test on your own documents, not a promise. No public figure can tell you how your suppliers’ paperwork will behave.
Validation: the step that makes it trustworthy
Validation is where a demo becomes a dependable process. Build checks that do not rely on the AI being right.
| Check | Example |
|---|---|
| Arithmetic | Line items add up to the subtotal; subtotal plus tax equals the total |
| Format | Dates parse; tax and bank identifiers have the expected shape |
| Cross-reference | Supplier exists in your records; purchase order number is open |
| Duplicates | Same supplier, number and amount already received |
| Confidence | Low-confidence or missing fields go to a person |
The Microsoft invoice output includes confidence information for extracted tables and cells, and key-value pairs are an optional return that is off by default. Check what your chosen tool returns per field, and set thresholds you can adjust after reviewing real results.
A safe rollout in four steps
- Pick one document type. Start with a single high-volume, low-risk type such as supplier invoices.
- Collect a test set. Use a few dozen real, varied examples (with sensitive details handled appropriately) and write down the correct values by hand.
- Run in shadow mode. Let the system extract while people continue the normal process, then compare. Note which fields fail and why.
- Go live with review. Auto-accept only fields that pass validation and confidence thresholds, and send the rest to a review queue. Keep sampling accepted items so drift is noticed.
This is the pattern described in human-in-the-loop workflows, and it fits the wider approach in how AI automation can improve business operations.
Build or buy
A managed document service is usually the quickest route for standard documents like invoices and receipts. Custom extraction models or your own pipeline make sense when your documents are unusual or very high volume. The trade-offs are covered in build vs buy for AI tools. Either way, you will be sending documents to a third party, so read how the vendor stores and uses your data. If the documents contain personal information, Australian privacy obligations apply to covered organisations.
Where Radiance Impact fits
Radiance Impact’s Technology division builds workflow automation, including ready-made automation kits, and can help you scope a document workflow with validation and review steps. You can see the style of workflow logic in the automation demos, explore the Technology division, or get in touch to discuss a document process.
Sources
- Invoice data extraction, Document Intelligence, Microsoft Learn
- What is Amazon Textract?, Amazon Web Services
