Use case

Document processing

Invoices, forms and contracts are read, checked against your own records and posted to the right system, with low-confidence fields and failed checks sent to a person.

Discuss this use case How it works

The problem

How the work goes today

Documents arrive by email, upload portal and scanner in whatever layout the sender uses. Someone opens each one, works out what it is, and types the supplier, amounts, dates and line items into the ERP or a spreadsheet.

Checks happen by eye. Does the total add up, is there a purchase order, has this invoice been paid already, do the bank details match the vendor record. The checks that get skipped on a busy day are the ones that catch duplicates and fraud.

Month-end and audit then depend on finding the original. It sits in a shared inbox or a folder with a name only one person understands, and corrections made along the way were never recorded.

$2–3 vs $10+

Cost to process one invoice at the best-performing finance teams compared with the worst, from APQC benchmarking data. This comes from a third-party summary; the survey year and definitions vary.[1]

37%

Share of finance leaders using AI who named accounts payable automation as a use, in a Gartner survey of 183 finance leaders. Knowledge management (49%) and error or anomaly detection (34%) were the other top uses.[2]

How it works

Step by step

Highlighted steps are where a person decides. Everything else runs on its own and is logged.

Systems involved

ERP (SAP S/4HANA, Oracle, Microsoft Dynamics, NetSuite)
AP automation (Coupa, Basware, Tipalti)
Contract management (Ironclad, Icertis, DocuSign CLM)
Document storage (SharePoint, OpenText)
Extraction services (Azure AI Document Intelligence, Google Document AI, AWS Textract)
Email and file intake (Microsoft 365, Google Workspace, SFTP)

01System
Collect
Documents are picked up from shared inboxes, SFTP drops, portals and scanners into a single intake queue.
02Agent
Classify and split
The model identifies each document type and splits multi-document PDFs before anything is extracted.
03Agent
Extract fields
OCR and a layout-aware model pull each field with a confidence score and the location on the page it came from.
04System
Run business rules
Totals, tax arithmetic and date logic are checked, and duplicates are searched for in the ERP or AP system.
05System
Match to master data
The supplier, bank details, purchase order and goods receipt are matched against records in the ERP (SAP, NetSuite or Dynamics).
06Person
Review exceptions
Low-confidence fields, failed rules and invoices without a purchase order appear in a review screen next to the source image for an AP clerk to correct.
07Person
Verify bank-detail changes
Any new or changed bank account is held until someone confirms it by calling the supplier on a number already on file.
08System
Post
Approved records are written to the ERP, AP or CLM system by API, with a link back to the source document.
09System
Learn from corrections
Every human correction is stored as a labelled example and added to the test set used to check the next model or prompt change.

Measurement

What we measure

Agreed before anything is built, with a baseline taken from how the work runs today.

Straight-through rate: Share of documents extracted, validated and posted with no human touch, reported by document type and supplier rather than as one blended figure.
Field-level accuracy: Share of extracted fields that exactly match a hand-labelled answer key built from your own documents, reported per field type. Precision and recall are reported separately for optional fields.
Confidence calibration: Whether fields scored at 0.95 confidence are right about 95% of the time on the labelled set. This sets the threshold below which a person must review.
Exception rate by reason: Share of documents routed to review, broken down by reason code such as low confidence, failed match, missing purchase order or suspected duplicate.
Cycle time: Median time from a document arriving to being posted or approved, measured from intake timestamp to ERP posting timestamp.
Cost per document: Software, model and staff time including rework, divided by documents processed in the period.

Pitfalls

What usually goes wrong

Demo accuracy doesn't survive real documents: Build a labelled test set from your own documents first, including poor scans, handwriting, multi-page tables and the long tail of small suppliers, and judge every tool against it.
Wrong values posted without anyone noticing: Cross-check fields against each other and against master data, and never auto-post a low-confidence amount or bank account. A blank is easier to catch than a wrong number.
The model fills in a field that isn't there: Require the page location for every extracted value, allow a 'not present' answer, and include documents with missing fields in the test set.
A supplier changes its invoice layout: Track the exception rate per supplier so a jump shows up within days, not at month-end.
Spoofed invoices and bank-detail fraud: Keep fraud controls outside the model: call-back verification for bank changes and duplicate checks in the ERP itself.

Build or buy

When a ready-made tool is enough

Custom work isn't always the answer. Here's how we'd decide.

Cloud extraction APIs (Azure AI Document Intelligence, Google Document AI, AWS Textract), IDP platforms (ABBYY, Hyperscience, Rossum, UiPath Document Understanding) and AP suites with built-in capture (Coupa, Basware, Tipalti) cover standard invoices and forms well.
Vendor straight-through rates in case studies are self-reported. Test any product on a sample of your own documents before relying on its figures.
Custom work pays off for domain-specific documents such as trade finance packs, insurance claims, medical forms or contracts checked against your own playbook, or when validation needs master data and rules from several systems.
The usual result is a mix: an off-the-shelf extraction service, with custom validation, routing, review screen and ERP integration around it, run on a private model if documents can't leave your environment.

Industries

Where this applies

Sources

Figures are as reported by each source on its publication date. Examples are public industry cases, not AI Sewa clients.

Next step

Want to try document processing on your own data?

Start a conversation All use cases