How to automate invoice processing, and what a person should still check

Automate invoice processing: the short answer

To automate invoice processing well, let software do the reading, the sums, the matching and the routing, and let clean invoices arrive in your accounts as draft bills. Keep a person on the exceptions: invoices that don't match, first invoices from new suppliers, changed bank details, and the decision to release a payment. The queue where those exceptions wait decides whether the whole thing works.

This guide is for the owner or operations lead of a firm of 10 to 100 people. Right now someone in your firm, often the owner or the finance manager, types supplier invoices into Xero, QuickBooks or Sage and chases approvals by email.

Most guides on this topic come from software vendors, and they promise "touchless" processing, where no person handles an invoice at all. We build business automations, and we'd set a different goal: a person touches fewer invoices, and always the ones where a mistake costs money. And the part vendors sell hardest, reading the invoice, is the part e-invoicing is about to make routine.

Which steps to automate, and what a person checks

Invoice processing runs in eight steps, from the invoice arriving to the money leaving. Software can do most of the work in six of them.

Beyond the approval itself, three decisions need a person every time. They get their own section below, after the step most vendor guides get wrong for small firms: matching.

This is a workflow with one AI step, the reading, surrounded by rules and a review queue. Every invoice follows the same path, and the model never decides what happens next. Our comparison of AI agents vs workflows explains why that shape suits most document work.

Matching when you don't raise purchase orders

Matching means checking an invoice against a record of what you agreed to buy. Two-way matching compares the invoice with the purchase order: same supplier, items, quantities and prices. Three-way matching adds the goods received note, the record that the delivery arrived.

Vendor guides assume you have both. If your firm raises no purchase orders, as many small firms don't, a "three-way matching" feature has nothing to compare. You still have other records, and each one can do the job:

• The supplier's history: the same lines and a similar total as their last six invoices.
• A price list or contract, such as a subcontractor's agreed day rate.
• Job numbers that exist and are still open in your job system.
• Delivery notes or timesheets that someone on your side has already signed.
• For rent, leases and software, a schedule with the same amount on the same date.

Say you run a 25-person building services firm. Subcontractors invoice against jobs, and merchants invoice against deliveries to site. The automation reads the job number, checks that the job is open, and compares the hours billed with the timesheets your site managers already approve. An invoice for a closed job, or for hours nobody signed, goes to a person.

Set the tolerance for each supplier, and write it down. A fuel card bill changes every month; a rent invoice should never change.

If most invoices have nothing to match against, the better first project may be purchase orders for your five biggest suppliers. Map the process first, delete what nobody needs, and automate what survives.

The review queue decides whether it works

The review queue is where the automation hands an invoice to a person. Its design decides whether that person checks the invoice or just clicks approve.

Send an invoice to the queue when any of these is true:

• A field was read with low confidence. Reading tools report this for each field. Microsoft's invoice model, for example, returns a confidence score between 0 and 1 for the total, the due date and every other field. Set a stricter threshold for the total than for the delivery address.
• The lines don't add up to the total, or the VAT rate isn't the one you expect.
• That supplier has already sent an invoice with the same number.
• The match fails, or there is nothing to match against.
• It is the first invoice from a new supplier.
• The bank details differ from the ones on file.
• It is a credit note, which reverses something a person should recognise.

What the person sees matters as much. Put the original document beside the fields the automation read, with the reason it stopped.

Store that reason as one value from a short fixed list, such as "no open job" or "bank details changed", rather than as a sentence. A field can be counted. After a month you'll know whether most of the queue comes from one supplier's layout or one missing job number, and you can fix the cause.

Watch for two warning signs. If a person approves everything in the queue for weeks without changing a thing, the rules are too cautious or nobody is really looking. If the queue grows faster than anyone clears it, the automation has only moved the pile. Our guide to designing the approval step so it isn't a rubber stamp goes further.

What a person should always decide

Three decisions stay with a person, however good the automation gets: changing a supplier's bank details, accepting a new supplier, and releasing a payment run.

Changed bank details

Mandate fraud, also called payment diversion fraud, usually starts with an email that seems to come from a supplier you already pay. It says their bank details have changed. Report Fraud is the police reporting service for England, Wales and Northern Ireland. It advises checking any request to pay a new account with the supplier directly, on their official contact details, and letting only designated staff change payment arrangements.

So the automation never updates bank details, whatever the email says. It spots the change, holds the invoice and sends it to the queue. A person phones the supplier on the number already on file, not the one in the email.

Large companies fall for this too. From 2013 to 2015, a Lithuanian man used a company registered in Latvia under the same name as Quanta Computer, a Taiwanese hardware maker that supplied Google and Facebook. He sent both companies forged invoices, contracts and letters, and they paid into his accounts (NPR). He pleaded guilty to wire fraud in 2019 (US Department of Justice).

A check on the supplier's name would have passed every one of those invoices, because the name was right. Match on details a fraudster can't copy from a website: the bank account on file, a job number you issued, the supplier's company number.

New suppliers

A first invoice needs a person to confirm that the supplier is real and that someone ordered the work. For a UK company, that includes a look at the Companies House register. The same person sets up the supplier record with its bank details, and every later invoice is checked against that record.

Releasing the payment run

Payment is the step that is hardest to undo. The automation can prepare the run: which bills are approved, which are due, and what is on hold and why. A person with authority reads that list and releases it. It's the same rule as in our checklist of what not to automate with AI: let the machine read and prepare, and keep the decisions you can't take back.

Disputes stay with people too. The automation can draft the query, but the conversation with a supplier you've used for ten years belongs to someone who knows them.

Scans and photos: make every invoice readable first

Invoices arrive as PDFs from accounting software, as scans with no text layer, as phone photos, and as text typed into an email. A reading step that can't see the text will skip fields or guess them, and a guessed total looks exactly like a read one.

So convert every invoice to text when it arrives, before any other step sees it. Split PDFs that hold more than one invoice, too. Microsoft's documentation for its invoice model warns that reading pages from two invoices at once returns partial data from both.

Keep the original attached to the bill in your accounts. The fields the automation read are a copy; the document is the record. Our guide to reading documents with AI covers this step in depth, including the part most projects skip.

E-invoicing shrinks the reading step, not the checking

An e-invoice is invoice data in a structured format that the buyer's software reads directly, with no scanning or extraction. A PDF attached to an email doesn't count. Germany's finance ministry says so in its e-invoicing FAQ: a plain PDF has no structured format.

Several European countries have made e-invoices compulsory, and the UK is next:

Sources: Germany's finance ministry, the European Commission on Belgium, France's Service Public, HMRC's consultation response and the Council of the EU.

For suppliers who send e-invoices, the reading step disappears, because the fields arrive as data. Everything after it stays. The sums still need checking, the match can still fail, and a fake supplier can send a perfectly structured e-invoice with a new bank account on it.

So if you are choosing a tool or a build in 2026, don't pay mostly for clever reading. Check that it accepts e-invoices as well as PDFs, and put the effort into matching and the queue, the parts that will still matter in 2030. If you sell to businesses in these countries, your own invoicing may have to change too; ask your accountant.

Rolling it out without trusting it blind

Count invoices per supplier over the last three months. If a handful of suppliers send most of them, start there, because each supplier you add brings its own layout and its own exceptions.

1. Run the automation in parallel at first. It proposes each bill, the person enters it as before, and you compare the two.
2. Once the comparison holds up, let clean invoices through as drafts and keep the queue for everything else.
3. Measure before and after: hours a week spent on invoices, days from arrival to approval, the share of invoices that reach the queue, and errors found after posting. We take that measurement before we start and again afterwards, because a feeling of "faster" isn't evidence.
4. Every month, read the queue's reasons and fix the most common one.

Each month, also pull a handful of invoices that passed every check and compare them with the originals. The dangerous failure is the quiet one: a supplier changes their layout, and the reader starts putting the wrong number in the total with full confidence.

An invoice sent to the queue costs someone a few minutes. A wrong total that passes every check becomes a payment.

When you don't need a custom build

Skip the build, and possibly us, in these cases:

• You get a few dozen invoices a month from the same suppliers. Your accounting package may already cover it. In Xero, you can forward a bill to your own bills address, and Xero creates a draft bill with the key details and the original attached. Check what yours offers before paying anyone.
• Most of your suppliers already send e-invoices, or soon will.
• Nobody has agreed who approves what. Write the approval rules down first, because software won't settle that argument.
• Your process is standard: purchase orders, one accounting system, one company. A packaged accounts payable tool fits that shape, and buying one usually costs less than building.

A build pays when invoices have to be checked against a system the packaged tools can't see. That might be your job management system, your stock records or a booking system. It also pays when you run several companies from one office, or have approval rules specific to your trade. Our guide to custom software vs off-the-shelf has a test for that decision.

Sort last month's invoices into three piles

Take last month's supplier invoices and sort them into three piles. Pile one went straight through with no questions. The second pile needed a question to someone, such as a missing job number or a price that looked wrong. Those in pile three needed a decision: a new supplier, changed bank details, a dispute.

Count each pile, and write down why each invoice in piles two and three stopped. Pile one is what to automate. The reasons behind pile two are your queue rules. Whatever sits in pile three stays with a person, and now you know how big it is.

Questions

Does this cover the invoices we send to customers?

No. This guide is about supplier invoices, the ones you pay. Your accounting software probably sends your own invoices already. Chasing customers who pay late is a separate job, where the tone of each reminder matters and a person should decide who gets a firmer one.

How accurate is AI at reading invoices?

It depends on your documents more than on the tool. Clean PDFs from accounting software read well; phone photos of creased paper don't. Ignore headline accuracy figures and test on your own last few hundred invoices. Judge the result by what reached your accounts wrong, because one wrong total that passes every check costs more than ten invoices sent to the queue.

Will our accountant be happy with this?

The automation creates draft bills in the accounting software you already use, with the original invoice attached, so the records your accountant works from stay the same. Ask them to agree the coding rules and approval limits before you switch it on. They often know VAT rules your team doesn't.

How we wrote this: drafted with AI, then checked and edited by the people who build our projects. Figures come from our own work or from the linked sources.