Document automation is the use of software to handle documents that would otherwise require manual processing — either creating documents from data, or extracting data from incoming documents.
Two categories#
Document generation means creating documents automatically from templates and data. Contracts, reports, invoices, letters. You have structured data; you need a formatted document. A template defines the layout; the software fills it in. This is well-understood, widely implemented, and mostly solved.
Document extraction means taking documents that arrive from external sources and pulling structured data out of them. Invoices from suppliers, lab reports from testing partners, claim forms from customers. You have a document; you need structured data. This is the harder problem, and for most businesses it’s where the volume and error rate are highest.
Most businesses need both categories eventually. They tend to start with extraction, because that’s where manual effort accumulates: a team processing hundreds of inbound invoices per week, re-keying data by hand into an ERP.
Why document extraction is the harder problem#
Generating a document from a template is deterministic. You control the template; you control the data; you know exactly what the output will look like.
Extracting data from an inbound document is not deterministic. You don’t control the source. Layouts vary between suppliers. The same supplier changes their template without telling you. Some documents are scanned at poor resolution. A required field appears in a different position depending on the document version. The data you want is there, but not where you expected it.
This is why document generation tools have been broadly available and reliable for years, while document extraction remains an active engineering problem. The variability of real-world documents is the challenge — and it doesn’t go away once you’ve handled your first batch of inputs.
For a deeper look at the broader category, see what is intelligent document processing.
What document extraction automation involves#
A working extraction system has four main stages, each with its own failure modes:
Ingestion — documents arrive in different formats: PDF, scanned image, email attachment, TIFF. Ingestion normalises them into something the extraction logic can work with. OCR happens here for scanned inputs. Corrupted or unreadable files are caught at this stage, not silently passed through.
Extraction — the system identifies and pulls out specific fields according to a defined schema. Invoice number, total amount, line items, dates. The extraction logic needs to handle layout variation across sources; a single rule set based on one supplier’s template will break the moment a second supplier is added.
Validation — extracted values are checked before anything goes downstream. Required fields present, data types correct, values within expected ranges, business rules satisfied. A document that fails validation generates a specific error and stops — it doesn’t silently produce wrong output.
Output — validated, structured data delivered to the downstream system: ERP, database, API, spreadsheet. The output format is consistent regardless of which input layout was processed.
See document extraction pipeline for a detailed breakdown of each stage.
Common use cases#
- Invoice processing and three-way matching
- Purchase order data extraction
- Contract data extraction (dates, parties, key terms)
- Lab report digitisation
- Customs and trade documentation
- Expense claim processing
- Loan application document review
The pattern is consistent across all of them: high-volume inbound documents, manually processed, with a downstream system that needs the data in structured form.
What it is not#
OCR alone. OCR converts a scanned image to text. That text still needs extraction logic to identify and pull out specific fields. OCR is an input to document automation, not the same thing.
RPA. Robotic process automation automates UI interactions — clicking through screens, copying data between applications. It can move data around, but it doesn’t understand document content. If the document layout changes, the RPA script breaks.
A one-time project. Documents change. Suppliers update templates. New document types are onboarded. The automation needs to change with them, which means it needs to be built for maintainability, not just initial accuracy.
The cost of manual document processing is well-documented; the less obvious cost is building extraction automation that can’t be maintained once the documents change.
Book a Diagnostic Session →