How AI Document Extraction Works

1. Preparing the image

Extraction quality is set long before any model runs. De-skewing, de-noising, contrast normalisation, page rotation and resolution checks determine what the downstream models can see. Quality scoring at this stage lets you reject or re-request a document rather than pushing a bad scan through the pipeline.

Bundles matter too: a single PDF often contains several logically separate documents that must be split before classification.

2. Classification

Each split document is classified — invoice, delivery note, claim form, identity document. Classification selects the extraction schema, the validation rules and the workflow that follow, so misclassification is expensive and is usually given its own confidence threshold.

3. Reading text, layout and tables

OCR and handwriting recognition produce text with positional coordinates. Layout models use those positions to reconstruct reading order, columns, key-value pairs and table structure. This is what allows an amount in the bottom-right of a page to be understood as a total rather than a stray number.

4. Understanding with language models

Language models interpret the extracted text in context: resolving which of three dates is the due date, normalising supplier names, reading free-text clauses, and reasoning across two documents that reference each other. This is the step that removes template maintenance.

5. Confidence and validation

Every field carries a confidence score. Fields below threshold, failed arithmetic, missing mandatory values and mismatches against reference data become exceptions. Validation is where extraction becomes trustworthy — a correct-looking number that disagrees with the purchase order is still a problem.

• Format rules — dates, tax identifiers, IBANs, HS codes
• Arithmetic — line items reconcile to subtotal, tax and total
• Cross-document — invoice against purchase order and goods receipt
• System-of-record — supplier, contract or policy exists and is active

6. Review and learning

Reviewer corrections are the most valuable training signal available. Capturing them in a structured way, with the original document and the model's original answer, is what makes accuracy improve over a deployment's life rather than drift.