The AI layer

AI that prepares the work. Not AI that decides the tax.

Almost every accounting product now puts “AI” on its home page. What is rarely explained: if the AI is wrong, what happens, and who carries it? This page answers that.

First rule

The AI may not compute tax, and the ban lives in the code

A rule that lives only in an internal document will be broken by someone who never read it. So we put the ban in the code structure: the part that computes tax may not touch the AI part at all. If anyone writes such a call, the build fails.

This is not a rule you can waive for a moment. It is part of the project structure, checked on every build, with no switch to turn it off.

What the AI does

Four jobs, all of them stopping before the decision

Reading documents

PDF invoices, bank statements, and vendor files are read and split into fields: date, number, tax base, VAT, counterparty. The result fills the form as a suggestion, and each field can be clicked to see where it came from.

Proposing accounts

The AI looks at your own recording habits, not another company’s, then suggests the account you normally use for a similar cost. A suggestion, not a decision.

Flagging anomalies

A cost that suddenly jumps, a new vendor appearing at month end, an invoice booked in a different month from its invoice date. All flagged for you to check, never quietly fixed.

Explaining rules

You can ask in plain language, and the answer comes with a rule reference. That reference is pulled from the rule list in the system, not composed by the AI. So a non-existent article cannot appear.

Two metaphors

Ink and pencil, used the same way across the app

You should not need the manual to know where a figure came from. Its appearance already tells you.

Ink

Solid line, blue, firm edge. This is a figure a person approved and that has entered the ledger. It cannot be changed. If it is wrong, it is reversed by an entry linked to the original.

Output VAT, July 2026 Rp168.300.000

Posted and locked

Pencil

Dashed line, grey, always with a reject button. This is anything from the AI. Reject it and nothing is left in the ledger, because it never got there.

Suggested account: 6-2200 Professional fees

Draft 12 similar transactions

Citations

Why the AI here cannot invent a statute

The most dangerous AI mistake in tax work is citing an article number that sounds convincing but does not exist. We close that path not with better prompting, but by never letting the AI write the reference itself.

The AI points at a rule, it does not write it

What comes out of the AI is not the words “article 6(1)(a)” but an index number in an internal list. Like picking from a list, not typing freely.

The system fetches the real rule

That index number is swapped for the real content: regulation number, effective date, and the text itself. If the number is not in the list, nothing appears.

What you read comes from the rule list

The text on screen comes from a list compiled and checked by our tax team. The AI never touches it, so it cannot invent it.

The effective-dated rule catalogue

Your data

What leaves the system, what does not, and what is logged

Personal data masked first

Tax IDs, account numbers, and personal data are replaced with random codes before a document is sent out. The AI reads the document’s shape, not a person’s identity.

An audit record per call

Every call to the AI is logged: when, for which document, and what came back. Anyone with the right permission can review it at any time.

Providers are replaceable

The AI sits behind a single door. Switching providers, or turning it off entirely, is a settings change. Nothing needs rewriting.

If your internal policy forbids data leaving at all, the AI can be switched off completely. You lose only convenience: automatic invoice reading and account suggestions. Everything else keeps running: every tax calculation, the monthly checks, and the working papers. None of them ever used the AI.

Next step

Test it with your messiest invoice

Bring the vendor invoice that is hardest to read — a crooked scan, a stamp over the figures, an unusual layout. We will run extraction in front of you and show what happens when the result is wrong.

Book a demo See pricing

No credit card. No commitment. We prepare the sample data.