Supplier invoices. Checked, then booked.
To automate invoice processing, AI reads each supplier invoice, plain rules match it against the purchase order and the goods receipt, and only clean invoices are booked without a person. Here is how the flow works, what to prepare, which invoices to automate and which to keep reviewing.
For teams that key insupplier invoices.
Read, match, then book.
What it means to automate invoice processing
In most finance teams a supplier invoice arrives as a PDF in a shared inbox. Someone opens it, types the supplier, the invoice number, the amounts and the VAT into the accounting system or ERP, looks up the purchase order, checks that the goods arrived, and sends it to a budget holder for approval. Each step is simple. Together they fill a large part of the week.
Automating that work means building the same sequence as a small, dependable flow. AI does the one step that used to need a person: reading an invoice in any layout and turning it into structured fields. Everything after that is ordinary software applying rules you already have. The model reads. It does not approve, and it never decides to pay.
The flow has six steps. The reading gets the attention, but the value sits in the matching and the checks, because that is where a wrong invoice is stopped before it becomes a wrong payment.
- Collect: pick up every invoice from the invoice inbox, the supplier portal and the e-invoicing channel.
- Read: AI extracts supplier, invoice number, dates, lines, totals, VAT and the bank account.
- Identify: link the invoice to a supplier in your master data and to a purchase order number.
- Match: compare invoice lines with the purchase order and the goods receipt, within your tolerances.
- Check: test for duplicates, VAT that does not add up and bank details that differ from the supplier record.
- Book or route: post clean invoices for payment and send the rest to a named reviewer with the reason.
Keying in invoices hides its cost.
Why manual invoice entry costs more than the typing
The typing is the visible part. The hidden part is waiting. An invoice sits in an inbox until someone has time, then sits with a budget holder who is travelling, then misses the payment run. Suppliers call to ask where their money is, and someone spends twenty minutes finding out.
Then there are the errors that only show up later. The same invoice sent twice, once by email and once by post, and paid twice. A transposed digit in an amount. An invoice for a delivery that was only partly received. Each one is rare, and each one takes far longer to undo than it took to cause.
Measure it with your own numbers. Count the supplier invoices you receive in a normal month. Time ten of them from arrival to booked, including the lookups. Note how many days pass between arrival and approval, and count the duplicate payments, credit notes and supplier reminders of the last quarter. That is your business case, and nobody else’s average can replace it.
Gather the inputs before the build.
What to prepare before you automate supplier invoices
Start with real invoices. Pull a few months from the archive and make sure the awkward ones are in: the scan of a paper invoice, the invoice that covers three purchase orders, the credit note, the one in another language, the supplier who never quotes your order number. This set shows what the flow must cope with, and it becomes the test every later change has to pass.
Then write down the rules your team applies from memory. How much may a price differ from the order before someone asks questions? Which suppliers send invoices without a purchase order, and who approves those? What happens when ten units were ordered and eight were delivered? These answers exist in people’s heads. The automation needs them on paper.
- A sample of real invoices from many suppliers, including scans, credit notes and odd layouts.
- Supplier master data: names, VAT numbers, bank accounts and payment terms as your system holds them.
- Open purchase orders and goods receipts, and how to read them from your ERP.
- Your tolerances: the price and quantity differences you accept without a second look.
- The approval matrix: who signs off which cost centre, and up to what amount.
- How your accounting system accepts invoices: an API, an import file or an existing connector.
Automate the match. Review the rest.
Which invoices to automate and which to keep reviewing
Not every invoice deserves the same treatment, and the split is not about how well the AI read it. It is about what you can check the invoice against. An invoice with a purchase order and a goods receipt can be verified by software line by line. An invoice with nothing behind it can only be verified by the person who ordered the service.
This is the three-way match: invoice, purchase order and goods receipt agree on what was ordered, what arrived and what is being charged. Where all three agree within your tolerance, a person adds nothing by looking. Where one is missing, a person has to supply the missing evidence. The rule set below is the one we would start from, and you tighten or loosen it with your own data.
Two more points. If a supplier can send a structured e-invoice, for example a UBL file over the Peppol network, take it: the fields arrive as data and there is nothing for AI to read. And if most of your invoices have no purchase order, the first fix is your purchasing process, not an AI model.
- Book automatically: known supplier, purchase order found, goods received, all lines within tolerance.
- Automate the reading, keep the approval: invoices without a purchase order go to the budget holder, coded and filled in.
- Always review: a bank account that differs from the supplier record, however good the invoice looks.
- Always review: a new supplier, a possible duplicate, or totals and VAT that do not add up.
- Always review: any invoice above the amount your approval matrix sets, matched or not.
- Skip AI entirely: structured e-invoices, which software can validate and match as they are.
Go live one supplier group.
How to roll out invoice automation safely
Run it in the background first. For one full month, including the month-end close, the automation reads and matches every invoice while your team keeps working as usual. Nothing it produces is booked. At the end you compare field by field: where it agreed with your team, where it was wrong, and where your team was wrong.
Then switch on automatic booking for the safest group only: a handful of regular suppliers whose invoices always carry a purchase order. Everything else reaches a person as a filled-in proposal with the source invoice beside it, so the work changes from typing to confirming. Widen the group as the comparison stays clean, one type of invoice at a time.
Keep three figures in view after launch: the share of invoices booked without anyone touching them, the fields people correct most often, and the days from arrival to approved. Keep every decision in a log as well, with the invoice, the extracted fields, the checks that ran and who approved. Your accountant will ask for it.
How we approach it.
How Walhallah builds document automations
We start with one invoice stream and your own examples. A focused prototype that reads real invoices and shows what it would book, and why it would stop, often takes days. That is enough to judge the results on your own documents before committing to more. A production version connected to your accounting system or ERP usually follows in weeks rather than months.
The closest work in our portfolio is on the order side. For Wegter we built an automation that turns email-based orders into structured SAP workflows, removing a manual bottleneck from the order desk. Supplier invoices are a different document moving in the other direction, but the pattern is the same: AI does the reading, engineering does the matching, the logging and the connection, and people keep the final word on anything unusual.
Questions. Answered.
Does invoice automation work with my accounting system or ERP?
In most cases, yes. The automation posts invoices the way other software does: through the system’s API, an import file or an existing connector. What matters is whether it can also read purchase orders, goods receipts and supplier records from the same system, because the checks depend on them. That is one of the first things to verify.
What happens when the AI misreads an invoice?
A misread field should fail a check before it reaches the ledger. Totals have to add up, the VAT number has to belong to a known supplier, and the lines have to match the purchase order. When a check fails, the invoice goes to a reviewer with the doubtful field marked and the original document next to it. Nothing is booked on the model’s word alone.
Can invoice automation stop duplicate payments and invoice fraud?
It catches the common cases more consistently than a busy person does. Software can compare every invoice with everything already booked, on supplier, number, date and amount, and flag likely duplicates. It can also hold any invoice whose bank account differs from the supplier record until someone confirms the change with the supplier through a known contact. It does not replace that phone call.
Is it worth it to automate invoice processing in a small company?
It depends on volume and on how many invoices have a purchase order behind them. If one person spends days each month keying in and chasing invoices, it tends to pay off. If you receive a few dozen invoices a month, the scanning and e-invoicing features already in your accounting software are probably enough, and custom work would be the wrong spend.