Tell us the job

Blog / Invoice automation

How to automate supplier invoice entry without changing your accounting software

Let an AI model read the PDFs, let plain rules check the numbers, and keep a person for the invoices that need one. Here is the method, step by step.

EnglishPortuguêsEspañol

It is Monday morning. Someone on your team opens the shared inbox, finds fourteen PDFs from suppliers, and starts typing. Supplier name, invoice number, date, net, VAT, total. One PDF is a scan at an angle. One is in German. One is a reminder for an invoice that was already entered last week. By lunch the spreadsheet is up to date, mostly, and nobody is quite sure about the German one.

This is ordinary work in a team of 5 to 50 people. It is also slow and easy to get wrong. In a 2026 benchmark sponsored by Medius, an accounts payable software vendor, Ardent Partners puts the average cost of processing one invoice at $9.90, against $2.67 for its best-in-class group, and an average exception rate of 19.9% (Ardent Partners, via Medius). An older APQC benchmark of 1,485 organizations found a median of $5.83 per invoice, with the bottom quarter at $10 or more (APQC, via CFO.com, 2018). Neither survey is about teams your size, so treat them as a rough range, not your number. Our reading for small teams: the cost sits in people's hands and in the exceptions.

Below is a method you can build yourself, with any tool, without changing your accounting software.

The idea in one paragraph

Use an AI model for the one thing it does well here: reading a messy PDF and copying out the fields. Use plain fixed rules, not AI, for everything that can be checked with arithmetic or a lookup. Anything that fails a rule goes to a person. The person stops typing and starts reviewing.

Step 1: One inbox for invoices

Create one address that receives supplier invoices and nothing else, for example invoices@yourcompany. Ask suppliers to send there, and forward stragglers to it. It sounds minor, but it lets your flow watch one place instead of guessing which emails carry invoices.

Step 2: Decide the fields you need

Write the list down before you build anything. A typical set:

  • Supplier name and supplier VAT number
  • Invoice number
  • Invoice date and due date
  • Currency
  • Line items (description, quantity, unit price, line total)
  • Net amount, VAT rate, VAT amount, total
  • Document type (invoice or credit note)

Your tax authority already tells you most of this list. HMRC's guidance on a full VAT invoice includes a unique sequential number, the issue date, the supplier's name, address and VAT number, a description, quantity and unit price, and the VAT rate and amounts (HMRC). In the EU, the European Commission describes basic EU-wide rules plus some national ones. If a field is legally required on the invoice, it is worth capturing.

Step 3: Let the model copy, never fix

Send each PDF to an AI model with a strict instruction: return these fields exactly as printed. If a field is missing or unreadable, return it empty. Do not calculate, round, translate or correct anything.

That last rule matters most. If the model quietly "fixes" a total that does not add up, you lose the one signal that something is wrong. You want the raw reading, so the next step can catch the problem.

Step 4: Plain rules check every invoice

A spreadsheet formula or a few lines of code can do these. No AI involved, so they behave the same way every time:

  1. Lines add up. The sum of line totals equals the net amount.
  2. Net plus VAT equals total. Allow a few cents of rounding when a supplier rounds line by line, no more.
  3. VAT rate matches. For each line, VAT divided by net gives a rate on your list of expected rates for that supplier or country. Check per line, so an invoice with two rates does not fail by default.
  4. VAT number looks right. The format matches the supplier's country (the right prefix and length). For new suppliers, confirm it once in the free official lookups: VIES for EU numbers, Check a UK VAT number for UK ones.
  5. Not a duplicate. The same supplier and invoice number has not been entered before. This catches reminders and resends.
  6. Dates are sane. The invoice date is not in the future and not more than a few months old. The due date comes after the invoice date.
  7. Currency is expected. It is on your list of currencies you actually pay in.
  8. It is an invoice. Not a credit note, a quote, a statement or a delivery note.

Step 5: Failures go to a review queue

Any invoice that fails even one rule goes to a separate tab or list called "Needs a person", with the reason written next to it: "total does not match lines", "possible duplicate of row 212". A person opens the PDF, decides, and moves it across. Invoices that pass every rule go straight into the main sheet.

The goal is not zero human work. It is that a person only looks at the invoices that need one.

Step 6: A short morning report

Once a day, send one message to whoever owns the process: how many invoices came in, how many were entered, how many wait for review and why. If the message stops arriving, the flow has stopped too, which is a check in itself.

Step 7: Start in a spreadsheet you own

Write results to a spreadsheet first, not straight into your accounting system. It is easy to inspect, correct and undo. Once you trust the output for a few weeks, connect the accounting system or keep importing from the sheet.

Step 8: Test on a month of past invoices

Before you rely on it, take one full month of invoices that are already entered by hand. Run them through the flow and compare. Count how many passed, how many went to review, and, most important, whether any wrong value got through the rules. If wrong values pass, add a rule or tighten one before going live.

Where it goes wrong

  • Bad scans. Photos taken on a phone, faint faxes and skewed pages produce bad readings. The rules will catch many of them because the numbers stop adding up, but not all. A clear "unreadable" route to a person is safer than a guess.
  • Handwritten notes. A "paid by card" or a corrected amount in pen can be ignored or misread. Send those to review.
  • Several invoices in one PDF. Split them first, or flag any PDF with more than one invoice number.
  • Credit notes. They look almost exactly like invoices. If one is entered as an invoice, you pay instead of being credited. Detect them and send them to review until you have a rule for them.
  • Tax treatment. Reverse charge, exempt supplies and mixed rates are judgment calls. HMRC, for example, expects a reference on the invoice when the reverse charge applies (HMRC), and says customers need VAT invoices as evidence to reclaim VAT (HMRC VAT Notice 700). The flow can capture the fields. Deciding how to treat them stays with a person or your accountant.
  • Supplier data leaving your accounts. Invoices carry names, bank details and prices. Before you send them to an AI provider, read its data processing terms and check that it does not train on your files.

How to build it

There are three common routes, and none is right for everyone:

  • Built-in capture in your accounting software. Many packages read PDFs sent to a dedicated address. Easiest to start; check how much control you get over the checks.
  • A dedicated invoice capture tool. More features for approvals and matching. Usually a monthly fee per user or per document.
  • A custom flow in an automation tool such as Make or Zapier, with an AI model for reading and your own rules for checking. Most control, and it runs in your own accounts. It needs someone to set it up and look after it.

What we built to test the method

We built a demo of this flow to check the method end to end. To be clear: it is a demo build on invented invoices, not client work. The sample set includes invoices in Portuguese, Spanish, French, German, Dutch and English. In our own tests it passed 101 of 101 offline checks. It runs in the owner's own accounts, writes to a Google Sheet they own, and sends one line at 08:00 each morning with what it did and what waits for a person.

What it does not prove yet: it has not been tested live with a real model reading real documents, so we make no claim about reading accuracy. That is exactly why step 8 exists, and why the rules, not the model, decide what gets through.

If you want to try this yourself, start with steps 1 and 2 this week. An inbox and a written field list cost nothing, and they make every later option easier. More practical guides are on the blog.

Sources

  1. Ardent Partners' Accounts Payable Metrics that Matter in 2026, Ardent Partners report sponsored and hosted by Medius (accessed 8 October 2026)
  2. Metric of the Month: Accounts Payable Cost, CFO.com with APQC data, 5 February 2018 (accessed 8 October 2026)
  3. VATREC5010: VAT invoice: Details which must be shown on a full VAT invoice, HM Revenue and Customs (accessed 8 October 2026)
  4. VAT guide (VAT Notice 700), HM Revenue and Customs (accessed 8 October 2026)
  5. VAT Invoicing, European Commission, Taxation and Customs Union (accessed 8 October 2026)

More from the blog