E-invoicing is already mandatory. Use it as your first automation.

Since January 2026 Belgian B2B invoices must be structured Peppol documents. That gives you machine-readable data you did not have before.

Most companies treated the Belgian e-invoicing mandate as a compliance chore: pick a provider, tick the box, move on. That is a missed opportunity, because compliance has just handed you the one thing every AI project asks for and most companies cannot easily produce — structured, machine-readable operational data.

Where things stand

Since 1 January 2026, VAT-registered businesses in Belgium must issue and receive structured electronic invoices for domestic B2B transactions, over the Peppol network, in EN 16931-compliant formats. A PDF by email no longer satisfies the requirement on its own.

The tolerance period ended 31 March 2026. Penalties now apply, escalating across repeat offences (€1,500 / €3,000 / €5,000).

Real-time reporting follows in January 2028, and intra-EU B2B in 2030. So whatever you build now sits on a foundation that is going to get richer, not be replaced.

Why this matters more than it sounds

Ask any AI vendor what they need and the answer is clean, structured, historical data about your operations. Ask most Belgian companies what they have and the answer is a folder of PDFs and an accounting package.

Size does not remove the problem, it relocates it. A two-hundred-person company usually has an ERP and plenty of data — and still cannot answer a question that crosses two systems, because the invoice line, the quote and the delivery note live in different places with different formats. Structured invoicing gives every company a common spine those systems can be joined on.

Structured invoicing changes that, for the transactions that describe most of what your business actually does:

what you sell, to whom, at what price, how often
how payment terms are honoured, by customer
which lines get disputed or credited
how quotes convert into invoices, and how long that takes

None of that required AI to be useful. All of it becomes usable because it is now structured.

Three uses that pay for themselves before anything clever happens

Payment behaviour, per customer. Not "average days to pay" — the distribution. Which customers slipped this quarter versus last. That is a phone call your finance lead can make on Monday, and it is arithmetic, not AI.

Quote-to-invoice leakage. Compare what was quoted with what was invoiced. Systematic gaps are either a pricing discipline problem or a scope problem, and both are expensive. Most companies have never measured this because the two documents lived in different systems and different formats.

Dispute and credit-note patterns. Credit notes concentrate. Once you can group them by customer, product and reason, the pattern is usually obvious and usually fixable.

Where AI actually adds something

Once the numbers are structured and computed by code, a language model is useful for the part that is genuinely linguistic: explaining a pattern, drafting the customer conversation, summarising a quarter for a management meeting.

Keep the split hard. The code computes the figures; the model writes about them. If a model is doing your arithmetic you have built something that will be confidently wrong at the worst possible moment — see the note on making skills deterministic.

Practical starting point

You already have a Peppol access point, because you had to. Ask your provider two questions:

Can we get structured exports of our own sent and received invoices, not just the compliance flow?
How far back does it go?

The answer to the second sets how soon this is useful. If you only have data from January 2026, you have a comparison in your hands by year end. Start extracting now rather than when you want the answer.