01 · OCR
A text dump is not a record
Optical character recognition returns characters. Finance still has to type vendor, order and amounts into the ERP. The scan did not reduce the posting work.
Intelligent document processing posts only after the ERP agrees
Intelligent document processing is the pipeline that classifies invoices, delivery notes and purchase orders, extracts the fields, and validates them against ERP master data before anything is posted. Optical character recognition only returns text. Pexon builds the validation and human-review loop into the ERP a mid-market European company already runs.
Four ways a readable document still fails to post
02 · Template
The next supplier breaks the layout
Template extractors die on the first rotated PDF or a line-item table that moved. Mid-market AP is a mix of PDFs, scans and e-mail bodies, not one form.
03 · Match
Extraction confidence is not a three-way match
A confident wrong vendor still posts. The check that matters is master data, an open purchase order, and a goods receipt — in the ERP, not in the model.
04 · Ops
Nobody owns the exception queue
Straight-through processing only exists if a named person clears what the threshold rejected. Without that owner, AP falls back to the shared mailbox.
The two answers on the table both fail, for different reasons
The first bad answer is better OCR. You get a searchable PDF and a CSV. Someone in accounts payable still keys the posting. Ardent Partners' 2025 AP metrics, summarised by Corpay on 31 July 2026, put the all-buyer average at USD 10.89 and 10.9 days per invoice against USD 2.78 and 3.1 days at the top. That gap is not an OCR problem.
The second bad answer is an intelligent-document-processing licence that never calls the ledger. You extract in a side product, then re-key or file-drop into SAP. SAP's own reference architecture for this work ends in a posting layer: outbound notifications, then an API such as A_SupplierInvoice, with blocked vendors and invalid accounts returned to an exception queue. Extraction without that loop is a demo with extra steps.
What intelligent document processing actually adds above OCR
01
Capture and classify
Invoices, delivery notes and purchase orders arrive from mailboxes, scans and uploads. The first job is to split mixed attachments and pick the schema. OCR runs here as a layer, not as the product.
- invoice
- delivery note
- PO
02
Extract and enrich
Named fields, line items, confidence per field. SAP publishes 85–95% accuracy for generative extraction of structured attributes — that is their range for extraction, retrieved 30 August 2026, not a posting rate. Enrichment looks up vendor and order in master data.
- fields
- confidence
- master data
03
Validate, then post
SAP's published pattern auto-confirms when every critical field is above a typical 90% threshold and sends the rest to a human workspace. Three-way match still has to pass before A_SupplierInvoice, or the equivalent API, creates the accounting document.
- 90% floor
- three-way match
- exception queue
The posting policy is the argument. The extractor is a line in it.
Illustrative of the SAP-published auto-confirm floor and posting API, not a connector we ran in a lab.
# posting policy — extractor is not the system of record
confirm_when:
critical_fields: ">= 0.90" # SAP typical auto-confirm floor
vendor: master_data_match
po_and_gr: three_way_match
post: A_SupplierInvoice # or the equivalent in your ERP
else: human_queueThree gates before the API. Invert that order and you buy a confident posting of a blocked vendor.
When the question becomes commercial
The extractor is a line. The product is invoice processing into the ERP a company of about 500 people in Europe already runs: match rules, the exception queue, and a posting API finance already trusts.
What intelligent document processing covers, and what you still have to build
| Layer | IDP stack | Still yours |
|---|---|---|
| Reading the page | OCR plus classification and field extraction | Scan quality, mixed attachments, language mix |
| Knowing the vendor | Enrichment against master data | Clean vendor records in the ERP |
| Permission to post | Confidence threshold and human workspace | Three-way match, tax, blocked-vendor rules |
| The ledger line | Outbound notification / posting API | Who clears exceptions the same day |
SAP's 85–95% extraction range and the typical 90% auto-confirm floor are from the SAP Architecture Center (last updated 3 June 2026, retrieved 30 August 2026). They are not Pexon measurements and they are not a posting SLA.
The breaks that surprise finance and IT together
- Most AP teams already bought tooling. Ardent's 2025 set, via Corpay, says about 75% of AP departments use some automation or AI. The top-performer exception rate is still 9% against a 22% average. The variable is how completely the posting loop was finished, not whether software was purchased.
- A 90% field is not a paid invoice. Auto-confirm is a queue filter. Three-way match, tax codes and blocked vendors still belong to the ERP. Skip them and you have automated a typing error into a payment.
- SAP Document AI is optional. It is a strong extractor with EU-only access as a published option. The loop we argue for works with that product, with another vendor, or with an open-weight model behind a gateway. Do not start by ripping out what finance already licensed.
Trade-off
Stay on OCR plus manual keying when invoice volume is small, almost every invoice is purchase-order-backed, and one clerk already clears the mailbox in a morning. Introduce intelligent document processing when you can name the break: non-PO invoices, mixed PDFs, or an exception rate you actually measure. If you cannot name the exception rate, you are buying a licence for a number you do not have.
Where this sits in the cluster
- invoice processing — the money page: book the invoice, do not licence another AP suite
- sap data layer — if vendor and order cannot be queried, this is a data problem first
- enterprise rag platform — the same brain, pointed at documents people ask questions of
- Company Brain hub — the cluster this post belongs to
Intelligent document processing questions
What is intelligent document processing?
Intelligent document processing is the pipeline that classifies a business document, extracts named fields, validates them against master data, and routes the result. Optical character recognition is one layer inside that pipeline. The output is a structured record that can be posted, not a text dump from a scan.
Is OCR enough for invoices and delivery notes?
No. OCR returns characters. It does not know whether the vendor exists, whether the purchase order is open, or whether the amounts match a goods receipt. Those checks live in the ERP. Posting on OCR output is how duplicates and blocked vendors enter the ledger.
What does a 90 percent confidence threshold actually mean?
SAP's published intelligent-document-processing architecture auto-confirms a document when every critical field is above a typical 90 percent threshold and sends the rest to a human workspace. That number is extraction confidence, not a licence to skip the three-way match. A confident wrong vendor is still wrong.
Do we have to buy SAP Document AI to do this?
No. SAP Document AI is one extractor with preconfigured invoice, purchase-order and delivery-note content. The posting loop — master-data match, three-way match, exception queue, A_SupplierInvoice or the equivalent API — is the part that has to sit in your landscape whether the extractor is SAP's, another vendor's, or an open-weight model behind a gateway.
What if we do not want to operate this ourselves?
Then do not stand up an extractor as a weekend project. The durable work is the posting policy, the exception queue, and the owner who clears it. Pexon sizes that loop in a two-week blueprint and will operate it only if you want that; the document is yours either way.
Next step
Map the posting loop before you buy another extractor
Two weeks, fixed price. We take one invoice population, write the match rules and the exception queue, and say whether the extractor you already have is enough. The plan is yours whether or not we build it.