Short answer: today's best AI models read roughly 90 to 96% of invoice data correctly, depending on document quality, but a person still has to catch the rest. So AI document processing doesn't replace the accountant or the clerk; it replaces typing: people look at the exceptions, not every line. It pays off most where there are many PDFs, paper documents and foreign invoices. Domestic B2B invoices, meanwhile, are moving to structured e-invoices across the EU, and where data arrives as data, there's nothing left to extract.
What AI document processing is
The older approach extracted data from incoming documents with optical character recognition (OCR) and predefined templates. That worked as long as every supplier sent the same format. When a new supplier came along or an existing one changed its invoice layout, the template had to be set up again.
Today's solutions, built on large language and vision models, "understand" the document instead: without templates, they recognise the invoice number, date, VAT number, net and gross amounts, and line items. This is often called intelligent document processing (IDP).
How accurate is it?
A 2025 research paper on invoice processing compared several large models. The best multimodal model, reading the document image directly, reached:
- 96.5% accuracy on clean, digital invoices,
- 92.7% on scanned invoices,
- 87.5% on scanned receipts.
In the same study, pipelines that first converted the document to text and then analysed it performed much worse. The lesson: models that read the image directly are more reliable today.
90 to 96% sounds good, but translate it. If an invoice has twenty fields, 95% accuracy means one field wrong on average. So a good system doesn't promise never to make mistakes; it promises to know where it's unsure.
Confidence scores: how people and machines work together
Microsoft's guidance for its document processing service explains the core idea well. The model gives every extracted field a confidence score: 0.95 means it's likely correct 19 times out of 20. The guidance recommends aiming for at least 80% model accuracy, close to 100% for financial data, and adding a human review step in critical workflows. It also notes that not every field returns a confidence score.
In practice:
- high-confidence fields: the system accepts them automatically;
- low-confidence fields: a person checks them, side by side with the document image;
- business rules: the system checks that net plus VAT equals gross, the supplier exists, and there's a matching purchase order.
So a person doesn't check every invoice, only the fields where the machine is unsure or something doesn't match. Set the threshold strictly in the first weeks and only relax it when measurements show high-confidence fields really are error-free. That way the system proves itself on your documents, not on the vendor's promise.
How much does it save?
Ardent Partners' surveys give a reference point for invoice processing cost, in two 2025 editions with different samples:
- One found best-in-class organisations process an invoice for $2.65, against an average of $12.42 (based on 204 finance leaders).
- The other found an average cost of $9.40 per invoice, with best-in-class cycle times of 3.1 days instead of 17.4, and exception rates of 9% instead of 22%.
The numbers differ between editions, but the direction is clear: the most automated companies process an invoice for a fraction of the cost, in hours instead of days. In an SME this shows up less in dollars and more in hours: less typing, fewer errors, faster approvals.
Where it delivers most
- Foreign invoices and receipts that arrive as PDFs or on paper.
- Delivery notes and purchase orders that must be matched to invoices.
- Contracts and quotes, where key data must be extracted: deadlines, amounts, termination terms.
- Field data collection, where paper forms need to be digitised.
- Expense claims, where staff send photos of receipts.
Where you won't need it
Domestic B2B invoices are changing. Under the EU's ViDA package, adopted in 2025, e-invoices based on the EN 16931 standard become the default, cross-border B2B digital reporting becomes mandatory from July 2030, and Member States can now mandate domestic e-invoicing. A structured e-invoice can be read, checked and booked by accounting software with little or no human intervention. Where data arrives as data, there's nothing to extract.
That doesn't mean you should wait. Many countries haven't set dates yet, and foreign and non-invoice documents aren't going away. But align the payback period of any AI project built only on domestic incoming invoices with your country's e-invoicing timeline.
Process matters more than the tool
The most common mistake is buying an AI extraction tool while the rest of the process stays the same: someone copies the extracted data into the business system by hand, approvals go by email, and matching against the purchase order happens in someone's head. Then the savings are small.
In a well-designed process:
- documents arrive in one place (a shared mailbox or upload);
- AI extracts the data, with confidence scores;
- the system matches it to the purchase order and goods receipt;
- discrepancies and uncertain fields go to a person;
- approved items flow automatically into the business or accounting system.
How to choose a tool
There are three kinds of solutions: the document processing services of the big cloud providers, products specialised in invoice processing, and general AI assistants that can extract data when built into a workflow. A few questions narrow the field quickly:
- Does it handle your documents? Test it on fifty real documents of mixed quality, not the vendor's samples.
- Does it give per-field confidence scores, and can you set the threshold?
- How does it connect to your business or accounting system? A ready connector, or development work?
- Where does it process the data? Invoices contain personal and business data, so this is a data protection question.
- What does it cost per document, and how does the price change as volume grows?
Data protection and retention
Invoices, contracts and expense claims contain names, addresses and bank account numbers. If an external AI service processes them, that's data processing under the GDPR: you need a data processing agreement and must know where data is stored and for how long. In business plans, the major providers typically commit not to train their models on uploaded documents, but check it in the contract.
Accounting records are also subject to statutory retention rules. Even with AI processing, keep the original documents, retrievable, for as long as your national rules require.
Three common mistakes
Testing on the vendor's samples. Everything is perfect in the demo. Your suppliers' crumpled delivery notes and oddly formatted invoices show what it can really do.
Dropping human review. One wrongly extracted bank account number or amount costs more than the automation saves in a year. Always have someone check uncertain fields.
Forgetting exceptions. Credit notes, partial deliveries and foreign-currency invoices are rarer, but they take the most time. Plan where they go.
What good looks like after three months
Three months in, a well-run document automation usually shows a few clear signs. Most incoming documents of the chosen type go through without anyone retyping them. The people who used to enter data now spend their time on exceptions: mismatched quantities, missing purchase orders, unknown suppliers. The share of documents that need human review has fallen as the threshold was tuned on real data. And approvals happen in days rather than weeks, because nothing waits in someone's inbox.
If those signs aren't there, look at the process before the tool. In most cases the bottleneck is a manual step that was left in place, such as copying data into the accounting system or chasing approvals by email, rather than the accuracy of the AI itself.
What we've seen
On an earlier project at an international animal health company, audits on customers' farms were done on paper and typed into spreadsheets afterwards. The solution there wasn't AI scanning of paper but replacing the paper: a web and mobile audit system where data is created digitally from the start. An important lesson: sometimes the best document processing is not creating the document at all.
On another project, at the company's Brazilian subsidiary, regulatory documents and procedures moved from a file server into a document management system with approval workflows, versioning and expiry tracking. Again, the point wasn't extraction, but making sure every document has an owner, a status and a deadline.
What to do now
- Count incoming documents by type: domestic invoices, foreign invoices, delivery notes, others.
- Measure how long recording and approving a document takes today.
- Pick the type with the most manual entry, where e-invoicing won't arrive soon.
- Design the whole process, not just extraction: matching, approval, booking.
- Set a confidence threshold and keep human review for uncertain fields.
- Ask where the document is created: if it's created by you, can a digital form replace it?
- Measure again after a month: time, errors, and how many documents went through without human review.
If you'd like us to look at which document types are worth AI processing in your company, and how it fits your accounting, have a look at our process automation and AI solutions.
FAQ
How accurate is AI at invoice processing?
In a 2025 study, the best model reached 96.5% on clean invoices, 92.7% on scanned ones and 87.5% on receipts. That's why uncertain fields should be checked by a person.
Will AI replace our accountant?
No. It replaces typing, not expertise. The accountant focuses on discrepancies, review and advice instead of routine data entry.
Is AI worth it if e-invoicing is coming?
For domestic B2B invoices, e-invoicing will replace extraction in the long run, but many countries haven't set dates yet. For foreign invoices, receipts, delivery notes and contracts, AI stays useful.
What is a confidence score?
For every extracted field, the model states how sure it is. A score of 0.95 means it's likely correct 19 times out of 20. Uncertain fields should go to human review.
How much does processing an invoice cost?
Ardent Partners' 2025 data puts the average at $9 to $12 per invoice and best-in-class at $2.65. In an SME, savings show up mainly as hours saved and fewer errors.
Sources
- Berghaus et al. – Multi-Modal Vision vs. Text-Based Parsing: Benchmarking LLM Strategies for Invoice Processing (arXiv 2509.04469)
- Microsoft Learn – Interpret and improve model accuracy and confidence scores (Document Intelligence)
- Ardent Partners – The State of ePayables 2025
- Thomson Reuters – 3 AP metrics that matter
- European Commission – Adoption of the VAT in the Digital Age package
