Customer due diligence produces a file nobody enjoys reading: identity documents, company searches, beneficial ownership declarations, source of funds, internal notes on the risk assessment. It is exactly the sort of material an automatic summary looks useful for, and exactly the sort it is not.
That data has one purpose
Italian Legislative Decree 231/2007opens in a new tab does not collect data “for the office”: it collects it to prevent money laundering and terrorist financing, and that purpose carries everything else with it, including the ten-year retention period running from the operation or the end of the continuing relationship.
One GDPR principleopens in a new tab applies here in full: purpose limitation. Data collected for one purpose is not used for an incompatible one. Sending it to a third party supplier to produce a summary is another purpose, and compatibility has to be argued beforehand, not afterwards. “It was for the same matter anyway” does not carry much weight.
What is actually inside it
The file is denser than the phrase “due diligence” suggests, and it is worth having the list in front of you before deciding what may leave. Inside there are: the identity document, with photograph, number, issuing authority and date; the tax code and full personal details; the declaration of beneficial ownership, which carries names, dates of birth and shareholdings of people who are often not even in the room; the company search with the whole corporate chain; the documents on the source of funds, meaning bank statements, prior deeds, sometimes tax returns.
And then there are the two parts that do not come from the client but are produced by the firm, and they are the most sensitive: the screening against politically exposed persons and sanctions lists, and the risk assessment with the score assigned. That last one is the firm’s opinion about a person, written down and kept for ten years. It is not data the client handed over: it is a judgement about them, which is why that line leaving weighs more than an identity document leaving.
And then there is Article 39, whose exceptions do not help
The point where this stops being a risk assessment is the disclosure prohibition. Article 39 forbids anyone under a duty to report a suspicious transaction, and anyone else who knows of it, from telling the client or any third party that a report has been made, that the Financial Intelligence Unit has asked for more, or that investigations exist or may be coming.
The exceptions exist and are listed: they cover professionals practising in association, as employees or collaborators, including in third countries, provided they apply equivalent measures, and the information exchanged may be used only to prevent money laundering and terrorist financing.
An AI supplier is in none of those categories. If pasted text carries the assessment that led to a report, or even the internal note contemplating one, disclosure to a third party has happened. There is no exception for “I had a program read it”.
The penalty here is not an administrative fine
On almost every subject on this blog the consequence of getting it wrong is administrative proceedings, or a disciplinary problem, or litigation. Not here, and it is worth spelling out because it changes the weight of the decision.
Article 55(4) of Legislative Decree 231/2007 punishes anyone who, being under the duty, breaches the disclosure prohibition in Article 39(1), and the one in Article 41(3), with arrest from six months to one year and a fine of 5,000 to 30,000 euro, unless the act constitutes a more serious offence. For a notary, who on the same material is already exposed to Article 326 of the criminal codeopens in a new tab as a public official, the two frames add up rather than cancel out.
The practical consequence is not to be more careful: care, at seven in the evening, is the first thing to run out. It is not to end up needing any: if the anti-money-laundering file physically cannot leave, because the tool works where the file already is, the question “should I have” never comes up.
So what can be automated
A great deal, on condition that the file does not leave. Reading an identity document, comparing a company search against a declaration, preparing a summary sheet: all of it can be done with tools that run on the office computer, where the data is never disclosed to anyone because it never moves.
The useful distinction is not between using AI and not using it. It is between tools that take the document to the model and tools that take the model to the document. Only the first make a disclosure to a third party, and that is what Article 39 forbids.
If part of the text has to leave anyway
For questions that need a large model, the route is substitution before sending: names, tax codes, VAT numbers, IBANs, addresses, land registry references and amounts become stable placeholders, and the mapping between placeholders and real values stays encrypted on the computer.
One limit is worth writing down: pseudonymisation reduces the risk, it does not repeal Article 39. A text stripped of every identifier but describing a recognisable transaction is still a communication about that transaction. The first question is still whether the content should leave at all, and only the second is in what form.
What a draft deed contains, value by value, is in the other piece.