AI invoice processing automation: what works and what doesn’t
Which accounts payable steps AI can handle on its own, where people still matter, and which metrics tell you whether the automation actually paid off.
If several hundred invoices reach you every month by email, as PDFs and as scans, somebody is retyping them. AI can now take over most of that work – but not all of it, and not without rules. Success is not decided by a promise of “100% automation” but by a single metric: how many documents get through without a person touching them.
Why invoices are usually the first process to automate
An incoming invoice has everything a good first project needs. High, regular volume. A clear start (the document arrives) and a clear end (it is posted and paid). A result you can measure in hours and euros. And with the right controls, the cost of an error is small – a misread amount is caught at matching or approval, not at the bank.
The environment is changing too. In Slovakia, mandatory electronic invoicing applies from 1 January 2027 (Act No. 385/2025 Coll., amending the VAT Act): a domestic VAT payer supplying goods or services to another domestic taxable person must issue an e-invoice, and the customer must be able to receive it. The invoice becomes a structured file under the European standard EN 16931 in the Peppol BIS format, delivered through a certified delivery service provider. From 1 July 2030 the obligation extends to cross-border supplies within the EU and to further taxable persons, electronic reporting of invoice data to the tax authority is added, and the VAT control statement is abolished. At EU level this builds on the VAT in the Digital Age (ViDA) package, adopted by the Council on 11 March 2025, which introduces digital reporting for cross-border B2B transactions from 1 July 2030.
A question to start with: how many documents pass through your team’s hands each month, and how many minutes does each one take?
What actually works today
Extracting data from PDFs, scans and emails
Document-processing models now reliably read the supplier, company registration number and VAT ID, invoice number, issue and due dates, net amount, VAT, total, IBAN, payment reference and line items. Old OCR needed a template for every invoice layout – when a supplier redesigned their invoice, it stopped working. Today’s models understand document layout: they know the figure next to the word “Total” is the total even on an invoice they have never seen before. Microsoft, for instance, states that its prebuilt invoice model accepts digital PDFs, scans and phone photos and supports 27 languages.
Where accuracy drops: handwritten documents, dark or skewed scans, small print and unusual layouts (an invoice combined with a delivery note on one page, for example). A good solution does not process these cases “somehow” – it flags them as low-confidence and routes them to a person.
Matching against the purchase order and goods receipt
Extracted data is only the beginning. The real saving comes from three-way matching: invoice – purchase order – goods receipt. The system compares supplier, items, quantities and prices. If everything matches within the agreed tolerance (a difference of up to 1% or €5, say), the document moves on automatically. If not, an exception is raised with a reason: different price, missing goods receipt, quantity above the order.
Set tolerances according to what a manual check costs you: too tight and you drown in exceptions, too loose and you overpay.
Posting to the ledger and the approval workflow
The third step is posting according to rules: account, cost centre, project, VAT deduction. Then routing to the responsible approver, a reminder before the due date and the payment order. This is where AI meets ordinary process automation: the model extracts and proposes, rules decide, people approve the exceptions.
A quick check: for each of the three steps, can you say who does it today and how long it takes them?
Where people are still needed
The human’s job is not retyping. It is exceptions and control. The following should land on your accountant’s desk:
- the first invoices from a new supplier – until you have confirmed the system reads them correctly,
- amount mismatches above tolerance and items with no purchase order,
- suspected duplicates (same supplier, same amount, same number in a different format),
- documents with handwritten notes, corrections or attachments that change their meaning,
- decisions in a dispute with a supplier – a complaint, a credit note, a set-off,
- every change of a supplier’s bank account.
The last point matters most. Fraud based on fake changes to payment details, known as business email compromise (BEC) – the attacker impersonates a supplier and sends “updated” bank details – is among the most expensive cybercrimes. The FBI’s Internet Crime Complaint Center (IC3) recorded 24,768 BEC complaints in 2025, with losses of 3.05 billion US dollars. An automated system should always stop an IBAN change and require verification by phone – on a number you have from previous correspondence, not from the email announcing the change.
How to measure success: agree the metrics up front
Without a “before” measurement you will never know whether the project paid off. Two weeks are enough: every document gets a timestamp on receipt and on posting, and your accountant notes how many they had to correct or chase.
| Metric | What it means | Target after the pilot |
|---|---|---|
| Straight-through rate | Share of documents processed with no human intervention | Set it after the first measurement; it rises with every exception type you resolve |
| Field accuracy | Share of key fields extracted correctly (amount, IBAN, number, dates) | Track separately for PDFs and for scans |
| Exception rate and causes | How many documents end up with a person, and why | A list of causes ranked by frequency |
| Receipt-to-posting time | Mean and median in days | Shorter than “before” |
| Cost per document | Hours × rate + licences, divided by the number of documents | Lower than before the pilot |
| Duplicates and overpayments caught | Count and amount per month | A number you did not have before |
| Payment timeliness | Share of invoices paid on time, early-payment discounts taken, late-payment penalties paid | Fewer penalties, more discounts |
The straight-through rate is the metric that decides. If it rises and the exception rate falls, the project is working – even if it never reaches 100%.
What a pilot looks like in practice
Picture a trading company with around 800 incoming invoices a month and two accountants. Weeks 1–2: measure the “before” state without changing the process. Weeks 3–6: a pilot on one supplier segment – stock goods, say, where purchase orders and goods receipts exist and matching has something to work with.
What typically emerges: exceptions are not spread evenly. They cluster around a handful of suppliers who send low-quality scans or invoices without a purchase order number. At that point the company changes the process, not the software – it asks those suppliers for PDFs with a text layer or for an e-invoice outright, and adds the purchase order number to its terms and conditions. The accountants’ work shifts from retyping to reviewing the exception queue.
What mandatory e-invoicing changes
A structured e-invoice under EN 16931 (in practice the Peppol BIS Billing 3.0 profile in UBL syntax) removes PDF extraction for domestic suppliers who are VAT payers. The data arrives as data, not as a picture. That is good news – but it solves only the first of the three steps.
Matching against purchase orders, tolerance checks, duplicate detection, IBAN-change verification, posting and approval all remain. And extraction does not disappear entirely: foreign suppliers until 2030, businesses not registered for VAT, cash receipts and expenses from business trips will keep arriving as PDFs and scans. It therefore pays to build the automation now so that PDF extraction is just one input – and from January 2027, receipt via Peppol plugs in alongside it.
What to watch when choosing a solution
- Where the data is processed. Invoices contain personal data (names, signatures, contact details). If a cloud service processes them, it is a processor under Article 28 GDPR – you need a contract that sets out the subject matter, purpose and duration of processing. Prefer processing within the EU.
- Whether your documents are used to train models. Get it in writing.
- What handling an exception looks like. One click on one screen, or opening three systems? This decides whether your accountant accepts the tool.
- Integration with your accounting software. Two-way: the document goes in, the payment status comes back. A solution built as custom software pays off when off-the-shelf products do not talk to your system.
- Archiving. In Slovakia, Section 35 of Act No. 431/2002 Coll. on Accounting requires accounting records to be kept for ten years following the year they relate to. The original document must remain accessible even after your contract with the software vendor ends.
- Audit trail. For every document, a record of what the model extracted, what a person changed and who approved what.
Key takeaway: Invoice processing automation is not about replacing your accountant but about changing their work from retyping to reviewing exceptions. The one metric that decides is the share of documents processed without human intervention – measure it before the pilot and after. Mandatory e-invoicing from 2027 removes PDF extraction for domestic VAT payers, but matching, controls and approvals remain yours. Never let a change of a supplier’s bank account go through automatically.
If you want to know what share of your invoices would get through untouched today, start with a free process audit – we will walk through two weeks of measurement with you and propose the scope of a pilot. The first step is a free consultation – at Easytrade, every automation project starts with one.
Sources
- Financial Administration of the Slovak Republic – e-invoice (mandatory electronic invoicing from 1 January 2027, Act No. 385/2025 Coll., EN 16931 / Peppol BIS format, certified providers) – https://www.financnasprava.sk/sk/podnikatelia/dane/dan-z-pridanej-hodnoty/e-faktura
- Financial Directorate of the Slovak Republic – Information 1/DPH/2026/I on Act No. 385/2025 Coll. (scope of the obligation from 1 January 2027 and 1 July 2030, abolition of the VAT control statement) – https://www.financnasprava.sk/_img/pfsedit/Dokumenty_PFS/Zverejnovanie_dok/Dane/Novinky_leg/Nepriame_dane/2026/2026.01.14_001_DPH_2026_I.pdf
- European Commission – VAT in the Digital Age (ViDA): adopted 11 March 2025, digital reporting from 1 July 2030 – https://taxation-customs.ec.europa.eu/taxation/vat/vat-digital-age-vida_en
- OpenPeppol – Peppol BIS Billing 3.0 (EN 16931 model in UBL syntax) – https://docs.peppol.eu/poacc/billing/3.0/
- FBI IC3 – 2025 Internet Crime Report (BEC: 24,768 complaints, losses of 3.05 billion USD) – https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf
- Act No. 431/2002 Coll. on Accounting, Section 35(3) (consolidated text published by the Financial Administration of the Slovak Republic) – https://www.financnasprava.sk/_img/pfsedit/Dokumenty_PFS/Zverejnovanie_dok/Sprievodca/Postupy_uct/2022/2022.01.10_002_DZPaU_2022_SD_Zak_uctov.pdf
- Regulation (EU) 2016/679 (GDPR), Article 28 – Processor – https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng
- Microsoft Learn – Azure AI Document Intelligence, prebuilt invoice model (supported inputs, 27 languages) – https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/prebuilt/invoice
A software and AI company from Košice, Slovakia – since 2011 we have helped companies integrate AI agents, automate processes and build custom software.
