How AI maps invoices to the right Belgian VAT grids — accurately

Belgian VAT is not a single rate applied to a line. It is a set of grids — codes that determine how each amount is reported: domestic sales, intra-community supplies, reverse charge, deductible vs. non-deductible, and so on. Code a bill to the wrong grid and the periodic return is wrong, even if every euro is correct.

This is exactly the kind of judgement that has kept invoice processing manual. So how do you automate it without getting it subtly, dangerously wrong?

Why a raw language model isn't enough

Ask a general AI model "what VAT grid does this invoice belong to?" and it will give you a plausible answer — based on the average of the internet, not on how this firm treats this vendor. For a fiduciaire, plausible is not good enough. Two print suppliers might be coded differently because of how each client's books are structured. The right answer is specific, not generic.

Retrieval over the firm's history

Counzel's Invoicing Assist works in three moves:

• Read. It extracts the structured fields from the document — vendor, amounts, VAT, dates.
• Retrieve. It searches the firm's own past invoices for the closest matches: same vendor, similar lines, similar treatment.
• Propose. It drafts the coding — accounts and VAT grids — anchored to how the firm has handled comparable bills before, not to a generic guess.

Because the suggestion is grounded in the firm's history, it gets more accurate per client over time. The more invoices flow through, the better the matches.

The accountant stays in control

None of this replaces review. Each drafted bill lands in the Review Inbox. Clean, routine ones can post automatically once you trust them; anything unusual is flagged as an exception with a plain reason. When you open a draft, you can see exactly which field came from which part of the document — there is no black box, and no confidence percentage pretending to be certainty.

Why "Belgian-deep" matters here

The grids, the reverse-charge rules, the disallowed-expense logic — these are encoded specifically for Belgium. A horizontal AI bookkeeping tool would have to rebuild a country's rulebook to match it. That depth is the difference between a demo that looks impressive and a tool an accountant can actually trust with a return.