A set of accounts to comment on, a reclassification to check, a contract to compare with last year’s. These are the things a language model is genuinely useful for in an accounting firm, which is why the file ends up in the text box before anyone has decided anything.
Removing the name does not remove the identification
The first thing people do, when they do anything, is delete the company name. What remains is a set of accounts with a turnover, a sector, a headcount and a province. To anyone who knows that market those are four coordinates, and the companies sitting in all four often fit on one hand.
The regulation puts it exactly that way. Recital 26 says that to decide whether a person is identifiable you consider all the means reasonably likely to be used by the controller or by another person. The companies register is one of those means, and it is public. Data is anonymous when nobody can get back; while the route exists, the data is personal and the regulation still applies.
The same holds in reverse: an IBAN or a VAT number left in because “they are only numbers” is the most direct identifier there is, since it takes no reasoning at all to trace it back to someone.
“But the accounts are public”
This objection always comes up, and it is a fair one: limited companies file their accounts with the companies register, anybody can buy them for a few euro, so what exactly is being protected. The answer is that the document being worked on is almost never the filed one.
What passes through a firm is the version before: the draft with adjustments still under discussion, provisions to be quantified, the note on a dispute that will not survive into the filing, the valuation choices being decided that week. Beside it sit the documents that are never filed at all: the trial balance, the ledgers, the interim position, the business plan, the loan agreements with their covenants, the letter of representation. And there is the plain fact that partnerships and sole traders file no accounts whatsoever.
Then there is the difference that matters most, and it is about time. Filed accounts say how a closed year went; an April draft says how a business is doing now, and those two pieces of information are worth very different amounts to whoever holds them. “It is public” is true of the final document and false of all the work that leads to it.
A payslip is not a document like the others
A payslip usually carries something the GDPRopens in a new tab treats separately. Sickness absence is health data; union deductions reveal trade union membership. These are special categories under Article 9, and for them processing is prohibited unless one of the conditions in paragraph 2 applies.
That is not a difference of degree. It is a difference of regime: for ordinary data you argue about the legal basis, here you start from a prohibition and look for the exception. The question “do I have an Article 9(2) condition for sending this payslip to an American supplier” has one answer, and it is no.
Who answers, and for how much
The controller is the firm, not the individual who pasted the file and not the supplier that received it. Article 83 grades fines in two tiers: up to 10 million euro or 2% of worldwide annual turnover for breaches of organisational and security obligations, up to 20 million or 4% for breaches of the principles, the legal bases, the special categories and the transfer rules.
The percentages are written for multinationals, and for a professional firm the numbers that matter are different: the client who leaves, the report to the professional body, the damages claim, and the insurance cover, which is worth reading closely when the event is a voluntary upload to an outside service.
And there is a new duty, which applies to you too
Article 13 of Italian Law 132/2025opens in a new tab speaks of the intellectual professions, and the category includes chartered accountants and accounting experts: the duty to tell the client you use artificial intelligence applies to you too (the statute in full is here).
The practical effect is that using AI stops being an internal matter for the firm. Either you write to the client that you use it, in which case it is worth being able to answer the next question too, which is always “and where do my numbers end up”, or you do not use it.
The question is data too, not just the document
One part of the problem gets missed because it has no file attached. A model receives two things: the document and the question, and almost always it is the second that says what really matters.
“A client in the province of Bergamo, forty employees, metalworking, is considering a restructuring procedure: how should the report be framed” contains no personal data in the strict sense, has no attachment, and needs no anonymising at all. It is also, in one sentence, the most confidential piece of information that firm holds at that moment. And the four coordinates that left are the same four that lead to the name.
The consequence is that a rule about files covers half the problem. The version that holds is a habit rather than a prohibition: describe the situation, not the case. “A company with forty employees” becomes “a small limited company”; the province goes unless the reasoning needs it; the sector stays only where it changes the answer. This is not caution for its own sake: a question framed that way also gets a better answer, because it asks for the principle instead of asking for advice on a case the model does not know.
The values that verify themselves, and the ones that get proposed
The technical side of this splits cleanly. Tax codes, VAT numbers and IBANs carry a check character that can be computed: a program does not recognise them by eye, it verifies them, and they either add up or they do not. Company names, personal names, addresses, amounts and dates come from a model instead, and a proposal can be wrong.
That is why nothing in Omissia is substituted before you have seen it: the review belongs to the firm, line by line, at the point in the document where the value sits. And the work happens on the office computer, not on a server: the file does not move, and the mapping between placeholders and real values stays there, encrypted, and disappears with the case file.
What still identifies a document after the data has been taken out is in the piece about metadata.