AI invoice processing for a small business.

What it does, step by step

AI invoice processing takes over the typing and leaves the decisions with people. A supplier invoice goes through these steps:

• It arrives in one place: an inbox such as bills@ that suppliers send to, a shared folder, or a photo taken on a phone.
• The AI step reads it, even when every supplier uses a different layout, and pulls out the supplier, invoice number, invoice date, due date, each line, the tax, the total and the bank details.
• It matches the supplier to the one in your accounting software, by name, tax number or bank details.
• It runs checks: has this invoice number been seen before, do the lines add up to the total, are the bank details the ones on file, does it match an order or a delivery note.
• It creates a draft bill with the lines coded to your accounts, and attaches the original.
• A person reviews the draft, corrects anything wrong and approves it.
• Payment happens later, in a payment run that a person approves, ideally not the same person.

The typing disappears in the middle steps. The safety sits in the last two, and they stay with people.

Try what your accounting software already does

Before you buy or build anything, look at the bill capture you may already pay for:

• Xero lets you upload bills or forward them by email, and its AI pulls out the key details to create draft bills ready for approval. Xero also offers Hubdoc to photograph and upload paper bills.
• QuickBooks lets you email or upload supplier invoices, and its AI fills in a bill with the dates, terms, category and line items for you to review and save. Higher Bill Pay plans add approval rules before a payment goes out.
• Dext, a separate capture tool, reads the supplier, date, tax, totals and line items, applies rules you set for each supplier, flags duplicates and publishes nothing to your accounts until someone has reviewed it.

For a business with a steady set of suppliers and one accounting system, start with one of these, and look further only if it falls short. Ask your bookkeeper first. They may already use one of these tools for other clients and know where it trips up.

When a custom assistant is worth it

The built-in tools read an invoice and make a bill. A custom assistant earns its place when the work around the invoice is the hard part:

• Each invoice has to be matched to a job number, a site or an order that lives in another system.
• Prices have to be checked, line by line, against what you agreed with the supplier.
• Delivery notes and credit notes have to be paired with the invoices they belong to.
• Invoices arrive in several languages, or mixed in with other mail in a shared inbox.
• Who approves depends on who ordered it, the site or the amount.
• Your accounting or job software has no bill capture of its own.

Our AI assistants are built for one job like this. The assistant pulls the details out of emails and documents and prepares a draft, and a person checks it before anything is saved or sent. The prices page shows where an assistant of this kind starts.

One supplier invoice, from inbox to approval

A restaurant with two sites buys from a dozen food and drink suppliers, and invoices arrive every day. Here is one of them, handled by an assistant built for that job:

• At 7.40am the fish supplier emails a PDF to the bills inbox.
• The assistant reads it: supplier, invoice 88231, today's date, payment due in 30 days, six lines, the tax and the total. It finds the supplier by its tax number.
• It checks the invoice number against the past year of bills from that supplier. There is no match, so it is not a duplicate.
• It checks the lines against the delivery note the kitchen photographed yesterday. One line, a case of salmon, was sent back at the door but is still on the invoice.
• It compares the unit prices with last month's. Salmon went up, and it notes that for the manager.
• It checks the bank details against those on file. They match.
• It creates a draft bill for the second site with each line coded to food costs, and marks it amber for the returned case and the price change.

At 10am the manager opens the review list. The green drafts need a glance each. For the amber one, she asks the supplier for a credit note for the salmon, approves the bill with a note that the credit is due and passes the price change to the chef. On Friday the owner approves the week's payment run in the accounting software. Nobody typed an invoice, and nobody paid one without looking at it.

The checks a person keeps

Decide these before you switch anything on, and write them down:

• Every draft bill is reviewed before it is approved. Once the readings have earned your trust, ordinary invoices from known suppliers can move faster, but a person still approves them.
• Any change of bank details is confirmed with the supplier by phone, on a number you already have, before the bill is approved.
• New suppliers are set up by a person, never created from an invoice alone.
• Totals that do not add up, missing tax numbers and unreadable pages go to a person instead of into a guess.
• Credit notes are matched to the invoice they correct.
• Payments are approved by a person, ideally not the one who approved the bills.

The AI step never pays anything. It reads, flags and prepares. Paying stays in your accounting software or bank, behind an approval that a person gives.

Bank detail changes and fake invoices

Invoice fraud aims at the moment a business pays. Watch for four forms:

• An email that looks like it comes from a real supplier, saying their bank details have changed.
• An invoice from a lookalike address, one letter different from the real one.
• A genuine invoice sent a second time with a new number.
• An invoice for work or goods nobody ordered.

A reading assistant can flag all four: bank details that differ from the record, a sender it has not seen before, the same amount and lines as an earlier invoice, and no matching order or delivery. It cannot decide which flag is fraud. The person who reviews does that, and the rule for bank details has no exceptions: call the supplier on a number you already had, never one from the email.

Where AI reading goes wrong

Reading invoices is good enough to save the typing, and it still makes mistakes you should plan for:

• A date like 03/04 read as March or April, depending on the country the supplier is in.
• A statement that lists several invoices read as one invoice.
• A credit note read as an invoice, so you would pay money you should be receiving.
• Lines run together on a long invoice, or a total taken from the wrong box.
• Poor phone photos, handwriting and faded receipts.
• A trading name that differs from the name on your supplier record.

Two checks guard against these: a test that the lines add up to the total, and a person reviewing every draft before approval. Before you rely on any tool, run fifty of last year's invoices through it, the awkward ones included, and compare every field with what your bookkeeper entered.

Check supplier statements once a month

Many suppliers send a monthly statement listing every invoice they think you owe. Comparing it with the bills in your accounts catches what reading one invoice at a time cannot: an invoice that never reached the bills inbox, one entered twice, or one with a different amount in each place.

Dext offers this check, and connected to Xero or QuickBooks Online it also finds invoices that are in your accounts but never reached Dext. A custom assistant can do the same against the statement PDF. Either way, a person decides what to do about each difference, usually by asking the supplier for a copy or a credit note. Do it before the month's payment run, so the run pays what you really owe.

In parts of Europe, invoices are becoming data

Some countries now require invoices between businesses to be structured e-invoices: files that software reads directly, with no page to look at. A PDF does not count.

• In Germany, every business has had to be able to receive e-invoices from other German businesses since 1 January 2025, and a PDF no longer counts as one. Formats include XRechnung and ZUGFeRD.
• In Belgium, structured e-invoices between companies have been compulsory since 1 January 2026, sent over the Peppol network.

A structured e-invoice goes straight into accounting software that supports it, with no reading step at all. AI reading is for everything else: PDFs, paper, photos and invoices from suppliers in countries without such a rule. If you trade in the EU, ask your accountant which rules apply to you and whether your software can receive e-invoices, and choose a tool that handles both kinds.

How to set it up

• Give invoices one address, such as bills@, and ask suppliers to use it. Forward anything that arrives elsewhere.
• Collect fifty real invoices from last year, including credit notes, statements and the worst photos you have.
• Agree the coding rules with your bookkeeper: which supplier goes to which account, and how tax is treated. The tool follows these rules. It does not decide them.
• Write down the approval rules: who approves bills, above what amount a second person looks, and who approves payments.
• Test on the fifty invoices, and adjust the rules until the drafts match what your bookkeeper would have entered.
• Run the new way next to the old one for one month-end, then switch.
• Each month, look at the flags and corrections, and add a rule or a supplier setting for any that repeat.

Most of the setup time goes on the rules. The approval steps can run as an automation around the reading step, so a draft left waiting reminds its approver and a failed connection alerts someone.

What to look at after the first month

• How many drafts were approved without a change. That tells you how far to trust the readings.
• Which fields people corrected most, such as dates, tax or coding. Each one points at a rule or a supplier setting to fix.
• How many flags were real problems. Flags that are nearly always false teach people to click past them, so tune those.
• Invoices that arrived somewhere other than the bills inbox, and which suppliers sent them.
• How long a bill waited between arriving and being approved.

What it reads and keeps

Supplier invoices carry more personal data than people expect: sole traders' names and home addresses, bank details, sometimes staff names. Agree before you start:

• Where the invoices and the extracted details are stored, and in which country.
• Who can see the review list and the bank details.
• How long the reading tool keeps its copies. Your accounts need the invoices for years, but the tool may not need to keep anything once the bill is in your software.
• Which services see the documents, so nothing goes to one nobody agreed to.

How long you must keep invoices depends on your country's tax rules, so ask your accountant. And give staff an agreed tool, so nobody pastes invoices into a chat app they found online: AI agents for a small business covers how far to let any AI tool into your systems.

When it is not worth it

• You get a handful of supplier invoices a month. Typing them takes less time than setting up and checking a tool.
• Most of your suppliers already send e-invoices your software reads.
• Your bookkeeper already captures your bills in their own tool. Ask before you add a second one.
• Your invoices are mostly paper, and nobody will photograph them on the day they arrive. Fix that habit first.

If the real problem is bills approved late or paid twice, the fix may be a clear approval rule and one person who owns supplier payments, with no AI at all. And if the money is stuck on the other side, with clients who pay late, see how to automate invoice reminders.