How to automate invoice processing

Invoices arrive by email, by post and through supplier portals. Someone collects them, retypes the header and line details, finds the right cost centre, emails for approval and books the invoice. That work is predictable and it happens again every week. That's exactly why it's one of the first processes you can automate.

Where does the time go with purchase invoices?

Check it in your own finance team. The time is rarely in the booking itself. It's in the steps around it.

• Collecting: gathering invoices from several inboxes, portals and the post.
• Retyping: entering supplier, invoice number, date, amount, VAT and lines into the accounting system.
• Matching: comparing the invoice with the purchase order or delivery note and working out the differences.
• Chasing: calling the budget holder to get approval.
• Searching: later finding out which invoice got stuck where, and why.

All five of those steps can be supported by software. Not all to the same extent, and not all with AI.

What can software really take over today?

Recognising invoice data is a mature technology. A model extracts the supplier, invoice number, date, amounts and VAT from a PDF, even when every supplier uses a different layout. Recognising invoice lines is harder, but for suppliers you use often it can be made to work well.

A model learns to assign the right ledger account and cost centre from your own booking history. The longer you use it, the less you correct.

What software won't solve for you: whether you want to pay this invoice. That decision stays with a person, and that's how it should be.

Step 1: map your invoice flow

Write down how many purchase invoices you process each month, which channels they come in through and how many suppliers account for most of the volume. It often turns out that twenty suppliers make up the majority of invoices. That's your starting point, because that's where the gain is biggest and recognition is easiest.

Note straight away how much time each step takes and how often something goes wrong. Without that baseline, you can't show afterwards that it delivered anything. The page on what manual work costs shows how to set up that calculation.

Step 2: choose one place where invoices come in

The biggest loss of time usually isn't in the processing but in the scattering. Invoices in three inboxes, a few in someone's personal inbox, a pile from the post.

Set up one address for all purchase invoices and ask suppliers to send them there. Scan the post in one place. From then on, you have one queue to automate instead of five. An inbox that sorts and routes itself is worked out under Shared Inbox Triage.

Step 3: connect recognition to your accounting software

Recognition without an integration gives you a tidy overview and still leaves you retyping. The gain is in the connection to the package you book in.

With Exact Online, Twinfield, Moneybird, SnelStart and Business Central, accounting packages widely used in the Netherlands, that's possible through their API. You read suppliers, ledger accounts and open purchase orders, and you write the invoice back as a booking or as a draft. What's possible for each package is under Exact Online, Twinfield and Moneybird.

Watch out for the details that hurt later: VAT codes, currencies, invoice dates that cross a period boundary, and credit notes. You deal with those cases in the design, not afterwards.

Step 4: put approval on one route

Decide who signs off on what, and up to which amount. Record that in the system instead of in someone's head. An invoice waiting for approval should have an owner, a deadline and a reminder that goes out automatically.

Whoever approves should be able to see the invoice, the purchase order and the difference on one screen. Without those three side by side, you get approval without checking, and that's worse than manual work.

The check for discrepancies can be done in advance. A checker that compares amount, VAT and order number and only flags the discrepancies is described under Invoice Checker.

What should you look for in invoice recognition software?

Ask four questions before you choose. How good is line recognition on your type of invoices, not on the demo? Where are documents processed and stored, and does that stay within the EU? How is the invoice written back to your accounting software, and is that a real integration or an export file? And what happens when the recognition isn't sure?

That last one matters most. Software that simply fills something in when in doubt costs you more time than software that honestly says someone needs to take a look.

When is this not your first step?

If you process a few dozen invoices a month, a project won't pay for itself. Check first whether your current accounting package already has a scanning feature you're not using. That's almost always cheaper than having something built.

If your supplier records are full of duplicates and outdated details, clean that up first. Recognition let loose on a mess produces a tidy mess.

And if your real bottleneck is on the sales side, for example in outstanding receivables, start there. Automating what already runs well delivers less than fixing what gets stuck.

What does it deliver?

The gain is in three things: fewer hours spent on admin, fewer booking errors and invoices that are ready for payment sooner. That last one gives you more control over your cash flow and saves on payment reminders.

Work it out with your own numbers: invoices per month, minutes per invoice, hourly rate. Set the licence and build costs against that and you have a payback period. How to build up that calculation is in our article on the AI business case. The approach itself is under AI and automation.