Invoice Processing with AI: From PDF Inbox to Verified Booking Entry

When a mid-sized company looks for a first AI use case that pays off and hurts no one, invoice processing comes up surprisingly often. The reason is simple: the task is frequent, follows rules, is based on documents – and nobody enjoys doing it.

At the same time, it’s an area where mistakes cost money directly. That’s reason enough for a sober look at what automation can actually deliver here – and where its limits lie.

The workflow today

In many companies, invoice processing looks roughly like this: invoices arrive by email in a shared inbox, some still on paper. Someone opens them one by one, reads off the supplier, number, date, amounts and tax rates, and types them into the accounting system. Then it’s checked whether a purchase order exists and whether quantity and price match. The invoice then goes to the relevant department for approval – usually by email, occasionally as a printout in an internal mailbox. After that comes the account assignment and finally payment, often just before the early-payment discount deadline – or just after it.

The time involved goes less into typing than into searching, following up and reminding people. That’s exactly where the leverage lies.

What AI takes over in this workflow

Extraction – even for unstructured formats

The first step is recognizing the invoice data. Classic systems work with templates: for each supplier, it’s defined where each field sits. That works as long as the layout doesn’t change – and it changes constantly.

Modern models work differently. They read the document for its content and find the invoice number because they understand what an invoice number is, not because it sits in a fixed position. The same applies to line items, discount lines, different tax rates, and the special cases that come up in practice all the time, such as consolidated invoices or credit notes.

One important side effect: structured formats such as ZUGFeRD or XRechnung, which are already mandatory in business with public-sector clients, can be integrated into the same workflow. Where structured data is available, it’s used preferentially – the AI is then only responsible for the rest.

Matching and reconciling

The second step is reconciling against what the company expected: is there a purchase order? Do quantity, price and terms match? Was the service recorded as delivered? This is where the difference between pure text recognition and a connected system shows – without a link to inventory management or ERP, the matching stays manual work.

Checking and flagging anomalies

The third step is the most valuable and the most often overlooked: the plausibility check. A system that knows your invoice history can flag anomalies that a person under time pressure would miss. A bank account that differs from the one used before. A price well above the usual level. An invoice number that has already been used. A service that was already billed for the same period.

The flag for changed bank details deserves particular attention: fraud involving fake invoices or supposedly changed account details is one of the most common attacks on mid-sized companies. An automatic comparison against the master data is one of the most effective and simplest countermeasures.

Pre-assigning accounts

From the supplier, line item texts and history, a suggested account assignment can be derived that is correct in most cases. It isn’t posted automatically, only suggested. Accounting confirms or corrects it – and every correction improves the basis for the next suggestion.

Routing approvals

The path to the relevant department’s approval can be derived from cost center, amount and purchase order reference. Instead of an email that gets buried in an inbox, there’s a task with a deadline and a reminder before the early-payment discount period expires.

Where people still have to stay in the loop

This isn’t a matter of courtesy – it’s a requirement, both for proper bookkeeping and for the internal control system.

Approving the service itself. Only the person who placed the order can judge whether a service was actually delivered. No system can replace that.

The final posting. A suggestion is a suggestion. Responsibility for the posting stays with accounting.

Approving the payment. Especially for anomalies and changed bank details, a human confirmation is essential here, ideally following the two-person rule.

The special cases. Every company has them: the invoice that needs to be split across three projects, the partial invoice with retention, the credit note referencing a consolidated invoice. A good system recognizes that it’s dealing with a special case and presents it – instead of guessing.

Traceability and retention

Two requirements need to be considered early. First, the logging: it must remain traceable which value was recognized automatically and which was changed manually, by whom and when. Second, the audit-proof retention of the original documents for the legally required periods – a requirement that’s independent of the processing itself, but is easily overlooked once a new tool is inserted into the process.

If personal data is processed – which is the case with travel expense or entertainment receipts – data protection considerations come into play as well: legal basis, data processing agreement, deletion concept.

When the effort pays off

An honest assessment: at twenty invoices a month, a project doesn’t pay off. It gets interesting once several of the following points apply:

• Monthly incoming invoice volume is in the hundreds.
• There are many different suppliers with different layouts.
• Early-payment discounts regularly go unused because approvals take too long.
• Approval runs by email or paper and can’t be tracked.
• During vacations or sick leave in accounting, the inflow backs up.

The benefit comes less from faster typing than from fewer follow-up questions, shorter turnaround times and fewer missed discounts. To put a number on the effect, it’s best to measure three figures before the start and again three months later: average turnaround time from receipt to posting, the share of discounts actually used, and the number of follow-up questions per hundred invoices.

A sensible starting point

Don’t start with all suppliers. Take the twenty that make up the largest share of your invoice volume, and initially let the system only extract data and suggest values – without anything moving forward automatically. During this phase you see the accuracy rate under real conditions, and your accounting team builds trust because it checks every step.

Only once recognition has proven reliable over several weeks does the second step pay off: automatic routing to approval, deadline monitoring, pre-assigned accounts. This order is the difference between a system accounting trusts and one it works around.

We’re happy to take a look at your actual document flow in an initial consultation – including the special cases that never appear in any process description but shape everyday work.

What you should measure before you start

To judge later whether the effort paid off, you need three figures from the current state. Collecting them takes about half a day.

• Turnaround time from the invoice arriving to it being posted, averaged over a typical month.
• Discount capture rate: the share of invoices where an early-payment discount was available and actually used.
• Follow-up questions per hundred invoices – to suppliers as well as internal approvers.

The second figure is, by experience, the most convincing. Missed discounts are a real, easily calculated amount, and they almost always arise from waiting time in approval, not from slow data entry.

The special cases are the real test

Every system looks good on a clean standard invoice. Whether it’s actually useful shows up in the cases that really occupy your accounting team: consolidated invoices with many line items, partial invoices with retention, credit notes referencing several documents, invoices in foreign currency, line items with different tax rates, scanned paper receipts of moderate quality, supplier invoices that need to be split across several cost centers.

Take exactly these documents into the test – and judge not whether the system solves them correctly, but whether it recognizes that it’s dealing with a special case and presents it instead of guessing.

Frequently asked questions

Does this replace our existing invoice capture system?

Usually not. In most cases, the recognition step is placed upstream and hands the verified data to the existing system. That’s the gentler path, because posting logic, interfaces and archiving remain untouched.

What’s a realistic recognition accuracy rate?

That can only be reliably determined using your own documents. What matters more than the rate anyway is how the system handles uncertainty: a field that’s flagged as uncertain and put up for review is harmless. A wrong value that goes through without any flag is not.

What about audit-proof retention?

That remains independent of the processing. Clarify early which system is the authoritative storage location for the original documents and how the retention periods are met there.

And data protection?

For pure supplier invoices, the personal-data aspect is minor. As soon as travel expense, entertainment or outlay receipts are added, you’re processing employee data – then the usual rules apply: legal basis, data processing agreement and deletion concept.

E-invoicing and structured formats

One development has changed the starting point: structured invoice formats such as XRechnung have long been mandatory in dealings with public-sector clients, and structured formats are increasingly taking hold in business-to-business dealings as well.

That’s good news for processing, because a structured document needs no recognition – the data is already machine-readable within it. In practice, that means a clear order for the architecture: where a structured format is available, its data is used. Recognition is only needed for everything else – PDFs without a structured component, scans, paper receipts.

When choosing a system, make sure it actually follows this order. Some solutions read even structured documents via their image rendering – which creates errors where none need to exist.

A second point: for mixed formats such as ZUGFeRD, which carry structured data inside a PDF file, both parts should be compared against each other. If they diverge, that’s a case for human review, not for automatic selection.

Rollout in three stages

The path from test to live operation should proceed in stages, each tied to a condition.

Stage 1 – Recognize and display only. The system reads the data and displays the results next to the document. Nothing is taken over, nothing moves forward. Accounting keeps working as before and sees, along the way, how reliable the recognition is. Duration: two to four weeks. Condition for the next step: recognition is stable for the twenty most important suppliers.

Stage 2 – Take over with confirmation. The recognized values are suggested, and accounting confirms or corrects them. This is where the first real time savings appear, and every correction improves the basis. The plausibility checks also run in parallel – deviating bank details, duplicate invoice numbers, unusual prices. Duration: four to eight weeks.

Stage 3 – Automate the approval path. Only now is the path to the relevant approval actively managed, with deadline monitoring and a reminder before the early-payment discount period expires. This is the step that actually changes turnaround time – and it requires the data foundation to be solid.

This order feels slow, and it’s the reason accounting ends up trusting the system. Whoever starts with stage 3 gets a system that gets worked around.

What accounting should know beforehand

A common source of friction comes from a misunderstanding: the worry that postings happen automatically. So make clear from the start what doesn’t change – posting responsibility, approval of the service by the relevant department, the two-person rule for payment, and control over every individual case.

What does change is the typing and the chasing people up. That’s exactly how it should be announced, too.