How many sample documents do we need to get started?

Enough to cover the variety you actually receive. For invoices, 30 to 50 samples from different suppliers usually show where a prebuilt model struggles. For a custom model, vendors differ: Google's Document AI custom extractor can predict from up to 5 labelled documents with its foundation-model option and be fine-tuned with more than 10. Include your worst scans, not just clean examples.

Can AI read handwritten forms?

Often, but less reliably than printed text, and results depend on the handwriting, the scan and the tool. Test with your own forms before planning around it. Where you control the form, printed fields, boxes for each character and checkboxes instead of free text make extraction much easier. Expect more of these fields to go to human review.

Where does our document data go when we use a cloud service?

To the provider's servers in the region you choose, under that provider's data-handling terms. Check whether the service stores documents, for how long, and whether it uses them to train models. For contracts and personal data, pick a region that matches your obligations and agree this with whoever owns data protection before sending real documents.

Should extracted invoices post straight into the accounting system?

Not at first. Write them as drafts or pending bills so a person approves the posting, at least until the error rate on your own documents is known and low. Even later, keep approval for new suppliers, large amounts and anything that fails a validation rule. Posting errors are slower to fix than review queues.

Is this different from RPA?

Yes, but they work together. Robotic process automation (RPA) repeats fixed steps in software, such as copying a value from one screen to another. Document AI reads the document and decides which value is which. A common setup uses document AI to extract the fields and a workflow or RPA tool to move them into systems that have no API.

Reading Invoices, Contracts and Forms With AI: Document Processing Without Retyping

AI reads invoices, forms and contracts reliably when you pair an extraction model with validation rules and send low-confidence fields to a person.

Dithered illustration seen from directly above of two marble hands feeding a jumbled page of characters into a pair of press rollers, with neat sorted rows coming out the other side along a crimson line.

No headings found on page

AI can extract data from invoices, contracts and forms reliably when three things are in place: an extraction model suited to the document type, validation rules that check the result against data you already trust, and a person who reviews only the fields the model is unsure about. Without the checks and the review step, errors reach your systems quietly.

Picture a distribution company that receives supplier invoices as PDF attachments, scanned delivery notes from the warehouse and signed customer contracts by email. Two people in the finance team retype vendor names, invoice numbers, totals and due dates into the accounting system, and someone else copies renewal dates from contracts into a spreadsheet. The work is steady, slow and error-prone at month-end. Most of it can be automated. The question is which parts, with which tool, and how you catch the mistakes.

What is intelligent document processing, and how is it different from OCR?

Intelligent document processing (IDP) is software that reads a document and returns named fields, such as vendor name, total and due date, rather than just the text on the page. Optical character recognition (OCR) is the step underneath it that turns an image of text into characters.

OCR alone gives you a block of text. You still need something to decide that "12/11/2026" next to "Due" is the due date and that the larger of two totals includes tax. IDP adds that layer: it finds key-value pairs, tables and line items, and maps them to fields your systems understand.

Two kinds of models now do this work:

Purpose-built extraction models are trained on a document type, such as invoices or receipts, and return a fixed set of fields with a confidence score for each.
General language models with document input read the page as text and image and can return whatever fields you ask for, which suits documents with no fixed layout, such as contracts.

Which documents can AI read reliably today?

Documents with a common structure and a known set of fields, such as invoices, receipts and standard forms, are the most reliable. Contracts and free-form letters work for specific clauses and dates, but need more review because the answer depends on reading meaning, not position.

A practical way to sort your documents:

Invoices and receipts: Structure: similar across suppliers. Fields: vendor, number, dates, totals, tax, line items. Fit: prebuilt models handle most of them.
Your own forms: Structure: fixed, because you designed them. Fields: whatever the form asks. Fit: form parsers or a custom model trained on a few samples.
Delivery notes and purchase orders: Structure: varies by supplier. Fields: order number, items, quantities. Fit: prebuilt or custom models, checked against the order in your system.
Contracts: Structure: free text. Fields: parties, start and end dates, notice period, renewal terms, governing law. Fit: language models, with a person confirming each extracted clause.
Handwritten or poor scans: Structure: unpredictable. Fit: test before committing, and plan for more manual review.

Quality of input matters as much as the model. A clean PDF exported from an accounting system is easier than a phone photo of a crumpled delivery note.

Which tools can extract data from invoices and forms?

The major cloud providers each offer document extraction services with prebuilt models for common documents and options to train on your own, and general AI models can now read PDFs directly. Choose by document type, where your data may be stored, and what your team already runs.

Azure Document Intelligence (Microsoft): Prebuilt: the prebuilt invoice model extracts fields such as customer name, billing address, due date, amount due and line items, and Microsoft lists invoices, utility bills, sales orders and purchase orders as supported, in 27 languages. Confidence: returned for words, key-value pairs, selection marks, regions and signatures, though Microsoft notes that not all document fields return one.
Amazon Textract (AWS): Prebuilt: AnalyzeExpense extracts invoice or receipt date and number, item prices, total amount and payment terms, plus line items. Confidence: a score for each piece of data. Other features: AnalyzeDocument reads forms, tables, signatures and layout, and accepts queries, where you ask a plain question such as which customer the document names.
Google Cloud Document AI: Options: the Form Parser extracts key-value pairs, tables and checkboxes without predefined fields, and the custom extractor pulls fields you define in a schema. Training: Google recommends starting with its foundation-model option, which predicts with up to 5 labelled documents and can be fine-tuned with more than 10.
General AI models with PDF input: Example: Anthropic's PDF support for Claude converts each page to an image and provides the extracted text alongside it, so the model can read tables and charts. Limits: up to 600 pages per request (100 when the context window is under 1M tokens), and each page typically uses 1,500 to 3,000 tokens. Fit: contracts and varied documents where you define the fields in the prompt.

Prices are per page and change; check each vendor's pricing page on the day you compare.

How do you decide which fields need a human check?

Send a field to a person when the model's confidence is below a threshold you set, when a validation rule fails, or when the field is high-risk enough that you always want a second look. Everything else can flow straight through.

Microsoft's guide to accuracy and confidence scores describes field confidence as an estimated probability between 0 and 1 that the prediction is correct, and says it "can be used to determine whether to automatically accept the prediction or flag it for human review." It gives no universal cut-off, so set yours from your own test results.

A marble face behind a brass loupe studying one crimson coin set apart from a tray of coins sliding on to a strongbox

Confidence alone is not enough. A model can be confident and wrong, for example when it reads the subtotal as the total. Validation rules catch those cases.

Validation rules worth adding first

Totals add up: line items plus tax equal the invoice total.
The vendor exists: the extracted vendor matches a supplier in your accounting system, by tax number where possible.
The order matches: the purchase order number exists and the amounts are within tolerance.
No duplicates: the same vendor and invoice number haven't been processed before.
Dates make sense: the due date is after the invoice date, and contract end dates are after start dates.

For contracts, treat every extracted renewal date and notice period as a field to confirm. A missed notice period costs more than the minute it takes to check one.

On AWS, Amazon Augmented AI (A2I) was built for this step and integrates with Textract's AnalyzeDocument operation, starting reviews by confidence range, missing keys or random sample. AWS's A2I page now notes that it is no longer open to new customers, so new projects usually build the review queue in their own workflow tool.

How do you set up a document processing workflow step by step?

Start with one document type and one destination system, measure it on real samples, then widen. A small pipeline that runs every day beats a broad one that nobody trusts.

Collect 30 to 50 real samples of one document type from different senders, including the messy ones.
Write down the fields you need and where each goes in the destination system.
Run the samples through two candidate tools and compare field by field against what a person would type.
Add validation rules that check the extracted values against your own records.
Set review thresholds so low-confidence or failed fields go to a review queue with the document side by side.
Connect the destination through its API, writing drafts or pending records rather than posting directly.
Log every document, its extracted fields, any corrections and who approved it.
Run in parallel with the manual process for a few weeks before switching the manual step off.

A marble hand holding a bronze mason's square against a stack of carved tablets, one tablet jutting out with a crimson gap of light

The intake step is often the hardest part. Documents arrive by email, upload, shared drive and scanner. Route them to one inbox or folder before extraction, and keep the original file linked to every record it creates.

How do you know the extraction is working?

Measure field-level accuracy on a fixed test set, the share of documents that pass with no human edits, and the corrections reviewers make each week. Corrections are the most useful signal, because they show exactly which fields and senders cause trouble.

Keep the test set stable so you can compare when a vendor updates its model or you change a prompt. Add new failure cases to it as you find them. When one supplier's layout keeps failing, a small custom model or a supplier-specific rule usually fixes it faster than switching tools.

Watch for silent drift too. A supplier changes its invoice template, and a field that was always right starts going to review. A weekly look at review volume by sender catches this early.

How we approach it at Vizio AI

We start with the documents, not the tool. We gather real samples, list the fields each team actually retypes, and trace where each value goes. Then we test two or three extraction options on the same samples and pick by measured field accuracy, data location and cost per page.

The pipeline we build keeps a person in the loop for low-confidence fields and failed checks, writes to the destination system as drafts, and logs every correction so the rules improve over time. For finance teams, it connects with the invoice matching and close work we describe in AI in the Finance Back Office. When the goal is answering questions from documents rather than extracting fields, Chat With Your Company Documents covers that pattern, and our earlier post on RPA in document management covers the rule-based side of contract workflows.

This is part of the operations work of our DX Studio.