Invoice Processing with AI: What to Automate First and What to Keep Manual

Accounts payable is often the first place finance teams look when they consider AI. The work repeats every week, the input is documents, and mistakes are expensive enough that someone always notices. That makes invoices a good candidate, but only if you are clear about which steps a model should handle and which ones stay with a person.

This article walks through the invoice process step by step, shows where AI earns its place, and lists the controls we put around it in our projects.

Start with the format your invoices arrive in

Not every invoice needs AI. Under EU rules, an electronic invoice is one that is issued, transmitted and received in a structured data format that allows automatic processing, based on the European standard EN 16931. A PDF attached to an email does not meet that definition. When an invoice arrives as structured data, for example through the Peppol network, the fields are already there, and a rule-based import is cheaper and more reliable than any model.

The share of structured invoices is growing. In Germany, every domestic business has had to be able to receive e-invoices since 1 January 2025, and businesses may keep sending other formats only until the end of 2026, or the end of 2027 if their prior-year turnover was €800,000 or less. At EU level, the VAT in the Digital Age package, adopted in March 2025, brings digital reporting for cross-border B2B transactions from 1 July 2030 and lets member states introduce mandatory domestic e-invoicing. AI is most useful for what remains: PDFs, scans, photos of receipts and the long tail of suppliers who have not switched yet.

The invoice process, step by step

What to automate first

The best first steps are high in volume and low in judgement. In our projects, that usually means three things.

• Intake and classification. A shared accounts payable inbox collects invoices, payment reminders, statements and supplier questions. A model can sort them and route each one to the right queue, which removes a surprising amount of manual clicking.
• Field extraction from unstructured invoices. Language models that read images can handle layouts they have never seen, which older template-based OCR could not. The output should always be the same structured record your accounting system expects, so a scanned invoice and a Peppol invoice look identical further down the line.
• Coding suggestions. Most companies book the same supplier to the same account and cost centre month after month. A model that proposes the coding from past bookings, and shows why, saves time while keeping the decision visible.

What to keep manual

Some steps stay with people. Not because AI cannot produce an answer, but because a wrong answer costs too much, or because the responsibility has to sit with a named person.

• Changes to supplier bank details. A request to update a bank account is a classic route for invoice fraud. Confirm it with a known contact through a separate channel, never by replying to the message that asked for the change.
• Releasing payments. AI can prepare the payment run. A person with the authority to spend the money approves it.
• New suppliers and unusual VAT cases. A first invoice from a new supplier, or a cross-border or reverse-charge transaction, deserves human review until you have evidence that the system handles it correctly.
• Disputes and larger exceptions. Differences that need a conversation with the supplier are relationship work, not data work.

The controls that make it trustworthy

An invoice system is only as good as the checks around it. These are the ones we build in from the start.

1. Confidence thresholds per field. Values the model is unsure about go to a review queue instead of being booked silently.
2. Arithmetic and completeness checks. Line items must add up to the totals. The EU VAT Directive lists the details an invoice must contain, such as the supplier's VAT number, a sequential number and the VAT rate applied, so missing details are flagged before booking.
3. Duplicate detection. Compare supplier, invoice number, amount and date. The same invoice often arrives twice: once by email and once through a portal or by post.
4. Matching tolerances set by finance. The accepted difference in price or quantity is a business rule. Write it down and enforce it in code rather than leaving it to the model.
5. An audit trail. The VAT Directive requires the authenticity of origin, integrity of content and legibility of an invoice to be ensured, for example through business controls that create a reliable audit trail between invoice and supply. Keep the original document, the extracted data, every change and who approved it.
6. A weekly sample review. Check a random sample of invoices that were processed automatically, not only the flagged ones. It is the only way to find errors the system was confident about.

How to start

Measure the current process for a few weeks first: how many invoices arrive, through which channels, how long each one takes and where they get stuck. Then pick the step from the table that matches your biggest bottleneck. For most teams that is intake and extraction, with a person reviewing every proposed booking during the first month. Widen automation only where the review shows the error rate is low enough to trust.

To estimate what this could save your team, describe the process in the ROI calculator. If you are still choosing where to begin, read five questions that pick the right first automation, or browse common workflows by business area.