OCR invoice processing: what really ends manual data entry

For finance and AP leads at 50 to 500-person companies who run OCR, or are about to buy it, and still see someone typing or fixing invoice fields. Not for developers after a free OCR library.

OCR invoice processing does not end manual data entry: it replaces typing with checking. Every field that optical character recognition extracts comes back with a confidence score, and the low scores go to a person for review. Only structured e-invoices, exchanged in a format such as UBL over a network such as Peppol, remove both the typing and the checking, and only for the suppliers who send them. No capture tool adds what the invoice does not say, like the GL account, the cost center or the approver, so the fair test of any invoice tool is the share of invoices that reach your ERP with no keystroke at all.

What does "eliminating manual data entry" actually mean?

Eliminating manual data entry means nobody types, confirms or completes a single field of the invoice before it is posted. That definition is stricter than the one most tools are sold on, and the gap between the two is where disappointed buyers usually end up.

Manual data entry on a supplier invoice is really three jobs stacked on top of each other:

• Transcribe. Copy what is printed on the invoice into the accounting system: supplier, invoice number, dates, totals, VAT, and often the lines.
• Correct. Fix what a machine read wrong or could not find. This job only exists once a machine does the reading.
• Complete. Add what is not on the invoice at all: the GL account, the cost center, the purchase order it belongs to, the person who approves it.

The market's promise is about the first job. Stripe's guide to OCR invoice processing says OCR "eliminates the need for employees to manually type data" from invoices. That is accurate about typing. It says nothing about correcting or completing, which is why teams that bought OCR often feel the work moved rather than disappeared.

The first job is the one everybody recognises. At La Nouvelle Garde, a brasserie group, the teams spent nearly 70% of their time between Gmail and the accounting system before automating, and the supplier inbox alone took two full days a week for one person, according to the case study published in September 2025. Those two days are transcription. OCR takes most of them away. What it hands back in their place is the subject of the next section.

Why does 95% accurate OCR still leave work on every invoice?

Because accuracy is quoted per field, and an invoice is many fields. A tool that reads each field correctly 95% of the time will usually get about one field wrong on a 20-field invoice, and it cannot be certain which one.

The comparison guides that rank for OCR invoice processing quote accuracy anywhere from 95% in Solvexia's words, "under optimal conditions", to above 99% in Klippa's software roundup. Of the five guides we reviewed on 26 September 2026, only Klippa's says its figure is counted per field, and none turns it into the share of invoices that go through untouched. Without the unit, the numbers cannot be compared, and without the conversion they cannot tell you how many invoices will need a person.

The vendors of the data extraction engines themselves are clearer than the resellers. Microsoft's documentation for Document Intelligence, updated in April 2026, defines field confidence as the estimated probability that an extracted value is correct, and gives the example itself: a confidence of 0.95 means the prediction is "likely correct 19 out of 20 times." The same page recommends using that score to decide whether to accept a value automatically or flag it for human review.

Amazon builds the same assumption into its product. Amazon Textract can be paired with Amazon Augmented AI, which sends a document to human reviewers when a key's confidence falls inside a range you set, or when an expected key is missing. Human review is not a failure of OCR. It is the operating model the engine vendors document.

Multiply per-field confidence across the fields of one invoice and the effect becomes visible.

Illustration computed by Phacet on 26 September 2026, assuming errors are independent from one field to the next.

In practice errors cluster: a blurred scanned invoice fails on many fields at once, a clean PDF on almost none. The direction holds either way. Posting an invoice untouched requires every field to be right, which is how one accountant put it in a thread on automating AP invoices: an invoice "needs every field right to post automatically."

Line items push the count up fast. Header fields are few and predictable. Descriptions, quantities and unit prices multiply the number of fields and are where layouts vary most, so a tool with excellent header accuracy can still send a large share of invoices to review once you need line-level data for matching or cost tracking.

Which invoice tools remove which part of the data entry?

Each category of invoice tool removes a different layer. Invoice capture by scanning and OCR removes transcription and creates correction. Structured e-invoicing removes both, for the suppliers who use it. None of them removes completion, which depends on rules and data the invoice does not carry.

Phacet analysis, 26 September 2026, based on the vendor documentation cited in this article.

The table makes one point hard to miss. The only line where typing and checking both disappear is the one where the invoice arrives as invoice data instead of as an image. A structured file has labelled fields: there is nothing to read, so there is nothing to misread.

That line is no longer theoretical. Since 1 January 2026, Belgium has required structured e-invoices between VAT-registered businesses, with Peppol as the main channel, and the European Commission's summary of the Belgian rule puts it plainly: sending "a PDF invoice by e-mail or via a platform will no longer be enough." In France, every VAT-registered business has had to be able to receive e-invoices since 1 September 2026.

Whatever the channel, the data has to land somewhere it can be checked and completed. Phacet's Structure layer turns invoices and other finance documents into auditable tables, which is where capture ends and the rest of the work starts.

Is OCR obsolete now that e-invoicing exists?

No. OCR stays necessary for every supplier who still sends a PDF or a paper invoice, and for most mid-sized companies that is a large share of the supplier base for years to come. What changes is its role: from the main way invoices enter the system to the fallback for those that do not arrive as data.

Ardent Partners' State of ePayables 2025, published in January 2026, finds that a majority of suppliers, 57%, can now send invoices electronically. Read the other way, 43% cannot. The report does not say whether an e-mailed PDF counts as electronic, and if it does, the share of truly structured invoices is lower still.

Mandates also stop at the border. Belgium's rule covers Belgian VAT-registered businesses, so a supplier in the United States or the United Kingdom keeps sending PDFs. France routes transactions with foreign operators to e-reporting rather than to e-invoicing, so a French buyer's invoices from abroad keep arriving as documents.

The realistic picture for a company with a few hundred suppliers is a mixed flow: structured files from large domestic suppliers, PDFs from small and foreign ones, and a handful of paper invoices. The tool question is not OCR or e-invoicing. It is whether one process takes both and gives the team a single queue of what still needs a person.

What is still typed after capture, and who should type it?

The data that is not on the invoice. Which GL account and cost center the cost goes to, which purchase order and delivery it matches, and who approves it. Capture tools cannot type this because there is nothing to read. Rules, supplier master data and history can, and a person only has to decide the cases where they disagree.

The completion layer typically covers:

• GL account and cost center, often by supplier, by site or by line description.
• The purchase order reference, when the supplier left it off or used a different number.
• The approver and the approval workflow, set by amount, entity or budget owner.
• VAT treatment for lines that do not follow the supplier's usual pattern.
• A new supplier record, including bank details that a person should confirm before any payment.

Once capture is automated, it becomes clear where the remaining time sits. At Jinchan, an F&B group, automating the finance inbox saved around ten to fifteen minutes a day, while price control and verification saved several hours a week, according to the case study published in January 2026 with Alban Cacace, the group's co-founder and COO. Getting data in was the small part. Checking and completing it was the large one.

That is the layer Phacet's Inbox agent takes on. It picks up each supplier invoice from the mailbox, validates it against what you already know about the supplier and the order, routes it, and sends a person only the exceptions, with its reasoning shown and every step recorded in an audit trail. For account codes and categories, the reclassification agent applies your chart and rules consistently across suppliers and transactions.

The person stays in the loop. Their job changes from keying fields to deciding the few that the system could not settle on its own.

How to test a vendor's data entry claim before you sign

Ask for one number on your own invoices: the share that reached the ERP with zero edits. Everything else a vendor shows you is an input to that number.

1. Pilot on your own invoices. Use last month's real flow, not a demo set, and include your messiest suppliers, multi-page invoices and at least one handwritten one.
2. Measure the touchless rate per invoice. Count invoices posted with no field typed, corrected or added, and divide by all invoices received.
3. Ask how accuracy is counted. Per character, per field or per invoice. A percentage without its unit tells you nothing about the review queue.
4. Ask what happens below the confidence threshold. Who handles the validation of that field, and is the correction logged? A correction that leaves no trace is invisible to an auditor.
5. Test line items separately. If you match or allocate costs at line level, header accuracy is the wrong benchmark.
6. Check structured invoices are read as data. A UBL, CII or Factur-X file should be parsed directly, not printed to an image and read again by OCR.
7. Count what the tool completes, not only what it reads. GL account, cost center and PO link are where the remaining hours usually hide.

For scale, Ardent Partners measures an average exception rate of 18.4% across AP departments, and a rate 47% lower for the best-in-class, which puts the leaders close to 10%. A vendor promising zero exceptions on a mixed flow is describing a different company from yours.

The question is worth asking precisely because the buying guides skip it. Across the five comparison pages ranking for this query on 26 September 2026, 12,968 words in total, the word touchless appears zero times, while accuracy appears 89 times. Nicolas Marchais, Phacet's CEO, told Les Echos that repetitive tasks taking two hours a day are a good place to start with AI. Invoice entry qualifies, as long as you measure what is left once the typing stops.

FAQ

Is OCR available for free?

Yes. Open-source engines such as Tesseract read printed text at no licence cost. They return raw characters, though, not invoice fields. Mapping that text to supplier, date, totals and line items, routing uncertain fields to a person and posting to an ERP is where the cost of an invoice tool sits, and where the remaining manual work is decided.

Can ChatGPT do OCR?

Yes. ChatGPT can read text from an invoice image or PDF and return it as a table. It does not know your suppliers, price lists or chart of accounts, it does not connect to your mailbox or ERP, and it keeps no audit trail of what it changed. It replaces a copy and paste, not an accounts payable process.

What percentage of invoices still need manual handling after automation?

About one in five on average. Ardent Partners' State of ePayables 2025 puts the average invoice exception rate at 18.4%, and best-in-class departments run 47% lower, close to 10%. The rate depends mostly on the share of structured invoices, purchase order coverage and supplier master data quality, not on the extraction engine alone.

Do you still retype supplier invoices if you use OCR?

Rarely the whole invoice, but rarely nothing either. OCR removes most of the typing, then asks a person to confirm low-confidence fields and to add what the invoice does not contain, such as the GL account or approver. Phacet's Inbox agent handles that second layer with rules and history, so the team only reviews the exceptions.

Last updated: 26 September 2026. Figures verified against Microsoft Learn, Amazon Web Services documentation, Ardent Partners and the European Commission on that date.