Supplier invoices with AI: what it really does and what to check

In many small businesses there is someone who, every week, opens the inbox, downloads the suppliers' PDFs, rescues the receipts and paper invoices from the drawer and types them one by one into the accounting software. Supplier, tax ID, number, date, base, VAT, total. A wrong digit turns up months later, when the quarter's VAT does not add up. This post explains what AI does when it reads those invoices, where it gets things wrong, what it needs from your business and what it has to do with Verifactu.

What it does, step by step

"Digitising invoices with AI" sounds like scanning, but scanning only produces an image. What matters is the data. The process has four steps:

1. Collect the document wherever it arrives: an email inbox, a folder, a photo taken with a phone.
2. Read it and extract the data. An AI model reads the PDF or the photo and pulls out the fields that article 6 of Spain's Invoicing Regulation requires on every full invoice: number and series, issue date, issuer's name and tax ID, address, description of the goods or services, taxable base, VAT rate and VAT amount. If needed, the lines one by one too, with quantities and prices.
3. Match it to what you already have. The tax ID it read is looked up among your suppliers. If it exists, the invoice is attached to that record and inherits its usual account and VAT treatment. If not, the system proposes creating the supplier.
4. Send it to the accounting software and flag whatever is unclear. With Holded, for example, all of this goes through its API: the purchase is created, the original PDF is attached and the purchase is approved, with nobody typing.

The key point is the last one: the outcome is not "the AI does the bookkeeping on its own" but "the AI prepares and a person reviews the doubtful ones". Accounting stops transcribing and starts checking.

Where it goes wrong and how you notice

A modern model reads a clean PDF invoice well. The problems are rarely in reading; they are in interpreting. These are the ones we see most:

• The tax ID, with or without prefix. A supplier from another EU country shows its VAT number on the invoice, starting with the country code. If the system stores it without a rule, the same supplier ends up duplicated, or the return for intra-EU transactions comes out wrong. Pick a rule and apply it every time.
• The VAT that is not there. An invoice for a service from a supplier in another EU country usually arrives with zero VAT and a "reverse charge" note. Read literally, it looks like a purchase without VAT; in the books, you have to self-assess the VAT. The AI can spot the note, but someone has to decide the tax rule.
• Duplicates. Some platforms send two PDFs for every charge, the invoice and a receipt, with the same amounts. And some invoices arrive twice because someone forwards them again. If number and supplier are not checked before creating the purchase, it is booked twice.
• Bad photos and complex invoices. A crumpled receipt or one with the total covered by a thumb; an invoice with discounts, several VAT rates or income tax withholding. The AI will read something; the risk is that it reads an eight where there was a three.

The good news is that almost all of these errors can be caught with simple rules, without relying on the model being right. Base plus VAT must equal the total. The tax ID must have a valid format. The invoice number cannot repeat for the same supplier. The date cannot be in the future. An invoice that fails these checks does not go through on its own: it goes to a review queue, with the image next to the data read, so the person can fix it in seconds. That is why our page on invoice automation with AI says it plainly: we do not promise a hundred per cent accuracy.

What it needs from your business

Before talking about models, you need to know how invoices come in today. It is the most underestimated part and the one that shapes the result the most:

• An invoice inbox: an email address only for invoices, given to every supplier. It is the cleanest channel: they arrive as PDFs and are read with no one touching them.
• Phone photos, for the paper that arrives with the delivery driver or one-off receipts: sent by email or by WhatsApp to a company number, they come in the same way.
• The office scanner, if it saves to a shared folder or sends by email.
• Supplier portals. Electricity, phone or online platforms: those invoices do not arrive, they have to be downloaded, and they are the ones that are always missing.

On the destination side, the accounting software needs a way in. Holded has a public API and almost everything can be automated with it; other programs only import a file, which also works, with a manual step.

A word of warning: if you receive few invoices a month and almost all of them as PDFs, you may not need anything custom. Holded, for example, already includes a scanner that creates the purchase from the PDF. Custom work makes sense with high volume, documents in odd formats or data that has to reach your system in a specific way.

What to check before you trust it

Before letting the system work on its own, test it with your own invoices, not with demo samples:

• Take the invoices from a real month, including the ugly ones: crooked photos, foreign suppliers, invoices with several lines and several VAT rates.
• Compare field by field what it read with what the paper says. Do not look only at the total.
• Check that new suppliers are created with the right tax ID and country, and that existing ones are not duplicated.
• Look at what it sent for review and why. If it sent nothing, be suspicious: either your invoices are perfect or the rules are not checking.
• Make sure the original PDF is attached to every purchase in the accounting software. It is what any tax inspection will ask for.

Paper: reading it does not mean you can throw it away

The fact that AI has read a paper invoice does not allow you to destroy it. The Invoicing Regulation requires received invoices to be kept for the statutory limitation period, with their original content. To do without the paper, Order EHA/962/2007 requires certified digitisation: a faithful copy, electronically signed and made with software approved by the Spanish Tax Agency. A phone photo read by AI is not that. If you want to stop keeping paper, that is a separate project; if not, the paper is filed as usual and the AI just saves you the typing.

And Verifactu, where does it fit?

Less than it seems. Verifactu regulates the software you use to issue invoices, not how you record the ones you receive. Reading your suppliers' invoices with AI neither creates an obligation nor helps you comply. Under the text currently in the BOE, invoicing software must be adapted before 1 January 2027 for companies and before 1 July 2027 for everyone else. The Ministry of Finance has announced that it plans to postpone this, but until it is published in the BOE those are the dates; we cover it in Verifactu: 2027 or October 2028?. The only thing that touches you as a recipient is that invoices made with adapted software carry a QR code, and the regulation lets the recipient use its data to check them with the Tax Agency.

What will change how invoices are received is mandatory B2B e-invoicing. According to the Spanish Tax Agency's timetable, which applies Royal Decree 238/2026, it becomes mandatory on 6 October 2027 for businesses that invoiced more than 8 million euros the previous year and on 6 October 2028 for everyone else, both for issuing and for receiving. Those invoices arrive with structured data and do not need AI to be read. That is why it makes sense to build the invoice intake once: today it reads paper and PDFs, and it is ready to receive e-invoices through the same channel.

Where to start

List how your invoices arrive today, how many there are a month and which program they go into. That tells you whether what your software already includes is enough, whether an off-the-shelf product will do or whether custom work is worth it. It is what the free invoice automation assessment asks, in about two minutes. And if you work with Holded and want purchases to come in on their own, we explain it in Holded development and integrations.