Document parsing: how it works, methods, tools and use cases

Document parsing is the process of turning a document, such as a PDF, a scan, a Word file or an email, into structured data that software can use: named fields, tables and text in the right reading order. A parser reads the file, works out its layout, identifies the values you need, such as the invoice number, the policy dates or every transaction on a statement, and outputs them as JSON, CSV or records in another system.

This guide covers how document parsing works, the three main approaches, the tools, common use cases, and how to choose a parser.

What is document parsing?#

Parsing means analyzing input and giving it structure. For data formats like JSON, the structure is already in the data (see our guide to data parsing). Documents are harder, for three reasons:

• Digital PDFs have no structure

  A PDF stores characters at positions on a page, with no notion of "this is a table row" or "this is the total."

• Scans have no text

  A scan or photo is just an image; there's no text until OCR reads it.

• Layouts vary

  Every vendor's invoice and every bank's statement puts the same information in a different place.

A document parser has to rebuild structure from what's on the page.

How document parsing works#

1. Get the textRead the text layer from digital PDFs and Word files; run OCR on scans and images.
2. Analyze the layoutFind blocks, columns, headings, tables and reading order, and keep each word's position.
3. ClassifyIdentify the document type, and split files that contain several documents.
4. ExtractPull the fields and tables you need, as key-value pairs and rows.
5. NormalizeConvert dates, amounts and names to standard formats.
6. ValidateCheck required fields, formats and totals, and cross-check against other documents.
7. ReviewSend low-confidence values or failed checks to a person.
8. ExportDeliver JSON, CSV or Excel, or push the data to your systems by API or webhook.

Rule-based, AI and LLM parsers#

In practice, production platforms combine them: document AI models for known types, an LLM prompt for a new layout, and rules for validation.

Two different parsing jobs

Parsing for business data means specific values, checked and delivered to a system: the borrower's ending balance, the invoice total, the policy expiration date. Accuracy per field and validation matter most. Parsing for AI search and RAG means the whole document as clean text or markdown, with tables and headings intact, split into chunks for a language model to search. Faithful layout and reading order matter most.

Tools are often built for one or the other, so decide which job you're doing first:

Document parsing tools#

The tools fall into four groups. Here's what each one does, as described in its own documentation:

Docsumo's pre-trained models cover bank statements, pay stubs, tax forms, ACORD forms and invoices, among 250+ document types. See Docsumo Document AI. For Tesseract, see our Tesseract guide.

A minimal parser in Python

A minimal table read from a digital PDF with pdfplumber:

import pdfplumber

with pdfplumber.open("statement.pdf") as pdf:
    for page in pdf.pages:
        for table in page.extract_tables():
            for row in table:
                print(row)

This works on clean, digital PDFs with ruled tables. Scans, borderless tables and tables that break across pages need more work.

Document parsing use cases#

• Accounts payable

  Invoice headers and line items for PO matching. See AP automation.

• Lending

  Bank statements, pay stubs and tax returns for underwriting and income verification.

• Insurance

  ACORD forms, loss runs and certificates of insurance.

• Commercial real estate

  Rent rolls and operating statements.

• Legal and procurement

  Contract dates, parties, renewal terms and clauses.

• Healthcare

  Intake forms, referrals and claim forms.

• Knowledge bases and AI assistants

  Policies, manuals and reports parsed for search and RAG.

Build or buy?#

Build with libraries when you have a few stable, digital layouts and engineers to maintain the code. Buy a platform when you have many layouts or senders, scanned documents, tables that span pages, or values that drive decisions and need validation and an audit trail. The hidden cost of building is upkeep: every new layout and every format change is more rules.

• 99%99%field-level accuracy across 250+ document types
• 95%+95%+of documents processed straight through, without manual review

How to choose a document parser#

• Test on your own documentsInclude your worst scans and longest tables.
• Score field-level accuracy and straight-through rateNot just whether text came out.
• Check table handlingMulti-page, borderless and nested tables.
• Check validation and reviewCan you add rules and review exceptions on the source page?
• Check integrationREST API, webhooks and connectors to your systems. See integrations.
• Check securitySOC 2 Type 2, HIPAA and GDPR where your data needs them.

The bottom line#

Document parsing rebuilds structure from files that don't carry it. For a few fixed layouts, libraries and rules work. For varied business documents, use document AI with validation and review. For search and RAG, use a parser built to preserve layout as text. For the broader picture, read what is data extraction.

Book a demo to see Docsumo parse your own documents, or start a free trial.

Frequently asked questions#