Use cases by function9 use cases

Document processing that knows which fields it is unsure about

The enterprise name for this is intelligent document processing. The older generation of tools worked from templates: one layout, one template, and a new supplier meant a new template. Model-based extraction reads layouts it has never seen, which solves the template problem and introduces a new one. Models make confident errors, and a confident error in an invoice total or a court deadline is worse than a blank field.

So every build here is designed around uncertainty. Each field carries a confidence and a location on the page, validation rules check what can be checked (totals, dates, VAT IDs, master data), and a review screen shows a person the document region next to the extracted value for anything that fails. Corrections are stored and become the test set, so the pipeline is measured on your documents rather than on a benchmark.

The documents differ by industry, and so do the rules around them: delivery notes and forms into spreadsheets, contracts against your own playbook, commercial leases into critical dates, customs declarations for a broker to sign, clinic referrals and edited bank statements caught before they cost money.

Document processing systems I can build

What separates a document pipeline you can trust

  • Citations for every field

    Each extracted value points to the page and region it came from. A reviewer checks in seconds, and an auditor can trace any number back to its source.

  • Validation beats confidence scores

    A model's own confidence is a weak signal. Line items that add up to the total, a VAT ID that exists, a supplier that matches the master data: those are strong ones.

  • The review screen is the product

    Whether a team saves time depends on how fast a person can confirm or correct a flagged field. That screen gets as much design attention as the extraction.

  • Documents can attack the model

    Hidden text in a PDF can carry instructions aimed at an AI reviewer. Extraction runs without tools that can act, and nothing in a document is treated as an instruction.

  • Some outputs need a licensed signature

    Customs entries, court deadlines and credit files are prepared by the system and signed by the broker, lawyer or underwriter. The design makes that review fast rather than skippable.

  • Where the data may go

    Health, legal and financial documents often cannot leave your infrastructure. Self-hosted OCR and open models are now good enough for many of these pipelines.

Frequently asked questions

Can AI read PDFs and extract data into Excel?

Yes. Current models read scanned and digital PDFs, including tables and handwriting of reasonable quality, and can fill a spreadsheet or a system directly. The part that decides whether it works in practice is validation and review: which fields are checked automatically, and how quickly a person can confirm the ones that are not.

What is intelligent document processing?

Intelligent document processing (IDP) is the category of software that classifies documents, extracts structured data from them and validates it, usually with a human review step. Older IDP relied on templates per layout; newer systems use language and vision models that handle unfamiliar layouts, with validation rules to catch their mistakes.

How accurate is AI document extraction?

It depends on the documents, so measure it on yours. A few hundred real documents with checked values give you field-level accuracy before anything goes live. Clean digital invoices are extracted very reliably; faded scans and handwriting less so, which is why uncertain fields are routed to review rather than trusted.

Can the documents stay inside the EU?

Yes. The major model providers offer EU processing options, and self-hosted OCR and open-weight models can keep documents entirely on your own infrastructure. For health data under GDPR Article 9, or client files at a law firm, that is often the deciding factor in the design.

Also browse: Finance and accounting · Law firms · Insurance · Excel and Google Sheets · all use cases