Everyone talks about sending e-invoices. The real work is receiving them.

In Italy every supplier invoice has arrived as structured XML through the SDI exchange system since 2019. In Germany, every company has had to be able to receive e-invoices since 2025, and issuing them becomes mandatory in phases from 2027. The format question is largely settled: FatturaPA, XRechnung, ZUGFeRD.

And yet, in most accounts payable teams I talk to, someone still sits in front of each incoming invoice and decides, line by line:

  • Which G/L account? Or which item?
  • Which cost centre, which project, which dimension?
  • Which VAT treatment? In Italy: which natura IVA, which non-deductible share?
  • Which accrual period, if the service spans months?

The XML made the invoice machine-readable. It did not make the coding automatic.

Why this is a good job for AI

Coding an incoming invoice is a matching problem. The answer is usually already in your history. If you posted the last twenty invoices from this electricity supplier to the same account with the same cost centre, the twenty-first is not a hard decision. If a software vendor sends a new line text that looks like last year’s licence renewal, a person would recognise it in a second.

That is exactly what language models are good at: reading short, messy line texts and recognising what they resemble. And it is safe ground, because the XML already carries the hard facts. Vendor, VAT number, amounts, tax rates and line texts are structured. The AI doesn’t have to guess them from a scan.

What I’m building

This is what I’m building next for Business Central: a coding assistant for incoming FatturaPA, XRechnung and ZUGFeRD invoices.

I want to be clear: it is in development. There are no results yet, and I won’t pretend otherwise. What follows is the design.

How it is meant to work

Same principle as the rest of my “Hack Business Central” series: the ERP keeps the rules, the AI explains its suggestion.

  1. Read the structure. The incoming e-document is parsed in Business Central: vendor, document type, lines, tax data. Nothing is interpreted that the XML already states.
  2. Apply the firm rules first. Where a vendor always goes to one account, or a purchase order exists, deterministic rules decide. No AI needed.
  3. Learn the rest from history. For lines without a firm rule, the assistant looks at how similar lines from this vendor, and from similar vendors, were posted before. It proposes account or item, dimensions and VAT treatment.
  4. Say how sure it is. Every suggestion comes with a confidence level and a short reason: “posted 14 times to 6200 with cost centre ADMIN in the last twelve months”.
  5. Send the uncertain lines to a person. High-confidence lines are pre-filled. Low-confidence lines are flagged for review.
  6. Post nothing on its own. The result is a purchase invoice draft. A person reviews and posts.

For Italy the VAT side matters most. Natura IVA codes N1 to N7, partially deductible costs and the TD16 to TD19 integration documents for reverse charge are where manual coding is slow and errors are expensive. For Germany, the same applies to VAT posting groups and to splitting a single invoice across cost centres.

What I expect to be hard

  • Thin history. A new vendor has no past. Firm rules and a person will carry more of the load there.
  • Inconsistent history. If the same kind of cost was posted to three different accounts, the assistant will learn three answers. Low confidence is the honest output. It is also a useful finding for the controller.
  • Free-text lines. Some suppliers write “as agreed” and nothing else. No model can code that reliably.

Limits, honestly

  • Nothing is posted automatically. That is a design decision, not a temporary restriction.
  • Confidence is not correctness. A confident wrong suggestion is possible. That is why every suggestion shows its reason, so the reviewer can see what it is based on.
  • Tax judgement stays with people. The assistant proposes a VAT treatment from history. Whether it is right for this transaction is decided by the accountant or the tax adviser.
  • Transport is not part of it. Receiving the invoice from SDI or a German channel is handled by whatever service you already use. The assistant starts once the e-document is in Business Central.

Read more

The pattern behind it, and the three use cases that already work, are in the white paper: Ask Business Central: three AI use cases that work today.

Is your team facing a similar problem?

If your accounts payable team still codes e-invoices by hand, I’d like to hear how you do it today. Real workflows will shape what I build. Discuss your challenge. I usually reply within two working days.