AI document processing for invoices, forms and contracts

How does AI document processing work?

A pipeline receives a file by email, upload or SFTP drop. If the file is a scan it goes through OCR. A language model or a layout-aware extraction model then pulls named fields and returns them as JSON. Code validates the JSON, and valid records are written to your ERP, CRM or database.

The validation layer decides whether the project works. Extraction without validation produces confident wrong numbers.

Choose between a general vision-capable language model and a specialized extraction service by testing both on your own documents. General models handle odd layouts well and are flexible about new fields. Specialized services can be cheaper per page at high volume. Vascoh builds the pipeline so either can sit behind the same interface.

Where the time goes today

Manual handling is slow even when it is accurate. Ardent Partners data cited by Bottomline puts the average time to process one invoice at 17.4 days for businesses without automation, and the average cost at $12.88. Invoices are only one document type. Delivery notes, certificates of insurance, purchase orders, lease pages and inspection forms follow the same pattern.

Classification comes before extraction. An inbox might receive invoices, statements, certificates and random attachments in the same stream. A first step identifies the document type, and each type then follows its own schema and validation rules. Unknown types go to a person instead of being forced into the closest template.

What validation looks like

Good validation is specific to the document type. Totals must equal line items plus tax. A vendor name must match a vendor record. A date cannot be in the future. A purchase order number must exist and be open. When a check fails, the record goes to a review screen showing the source image next to the extracted fields.

Store a confidence score or a validation status with each field, along with the final value. When a reviewer corrects a field, save the correction next to the original. After a few hundred corrections you have a labeled set that shows which vendors or form types cause most of the errors.

• Arithmetic checks on totals, tax and currency
• Master data matching against vendors, customers or part numbers
• Duplicate detection on document number plus amount
• Confidence thresholds per field with a human review queue

Why layouts cause trouble

Templates break. A supplier changes a logo, a scan comes in rotated, a table runs onto page two, a handwritten note covers a line. Models cope better than template rules, but multi-page tables and low-resolution phone photos still produce errors. The pipeline should measure its own accuracy by field so you can see which document sources need extra attention.

Multi-document cases need extra logic. A closing file, a claim or a shipment may contain ten documents that must agree with each other. Cross-document checks, such as matching a name and an amount between two files, catch problems that single-document extraction misses.

Connecting the output

Extraction ends when data lands where work happens. That might be a QuickBooks Online bill, a NetSuite vendor bill, a property management record or a row in a custom database. Write calls need idempotency so a retry does not create a duplicate, and the original file should be stored and linked to the record.

Retention matters as well. Keep the source file, the extracted JSON and the posted record linked by one ID, and apply the same retention period that your accountant or compliance contact requires for the original paper.