If a lawyer has a chatbot summarize a case file, the client has to know. Since October 10, 2025, that's the law in Italy: article 13 of Law 132/2025, the Italian AI law published in the Gazzetta Ufficiale on September 25, 2025, and it covers every intellectual profession. In the year since, the professional bodies have built templates and ethics rules on top of it, and they don't all say the same thing.
We're not lawyers and this isn't legal advice. We did write a statement on how AI is used, though, the AI Transparency page of a site we look after, and when we reread it for this article we found a sentence in it that was wrong. It's fixed now.
What article 13 says
Two paragraphs. The first limits AI, in the intellectual professions, to "instrumental and support activities," with the professional's own intellectual work remaining "prevalent." The second holds the obligation: information on the AI systems "used by the professional" must be given to the client "in clear, simple and exhaustive language" (current text on Normattiva, in Italian).
The text doesn't say what an intellectual profession is, or what happens to someone who doesn't inform the client. The professional bodies are filling those gaps, each one for its own members.
Who is bound, and who can sanction
The CNDCEC amended the accountants' code of ethics on November 20, 2025: article 21, paragraphs 8 to 10, in force since November 21, 2025, repeats the obligation and asks for the AI used to be mentioned "also in any documentation produced" (updated code). Since that day an accountant who doesn't tell the client breaks the code, and answers for it before the disciplinary board.
Engineers have a new article 13 in their code, approved by the CNI on June 23, 2026, and sent to the local orders with circular 457 of July 31, 2026. The engineer has to inform the client "in advance and in full," stating the characteristics and purposes of the systems, and the code calls missing or incomplete information an ethics violation. Each provincial order still has to adopt the code with its own resolution, so if you're an engineer, check first whether yours has.
For lawyers, the Consiglio Nazionale Forense circulated a disclosure template on October 13, 2025, that ends with the client's signature.
That leaves the professions without a register, like ours. Article 13 doesn't distinguish between those with a professional body and those without, but without one there's no disciplinary board. We behave as if the obligation applied to us too.
What to write in the notice
The CNF template says the lawyer "might use" AI "where considered useful," gives a few examples, such as admin work and drafts, and promises that every output will be checked. The CNDCEC model clause goes into the engagement letter and gives the client the right to ask that AI not be used on their matter; which tools, though, it only discloses "where requested." The strictest is the engineers' code: characteristics and purposes, before the work starts.
We side with the strictest reading. The law speaks of systems "used," and a sentence like "I might use AI," without saying which one or on what, matches the template, but it tells the person signing it nothing. For each tool we'd write its name, what it does in the engagement, which of the client's documents go into it and who checks the result. If personal data goes into it, you need the privacy notice as well.
In writing, before you start
The law doesn't ask for it in writing. A notice given out loud can't be proved, though, and the professional bodies' templates all lean toward writing: the CNF one ends with the client's signature, and the CNDCEC clause sits inside the engagement letter. For existing clients, something sent that stays on file is enough.
A statement we wrote, and what didn't add up
That AI Transparency page doesn't answer any obligation: we advise clients to disclose their AI use, and we did it there too. It says where a model runs (only in comment moderation), what it's sent, and where one isn't used: it makes no decisions about clients or visitors and doesn't run in the free tools. The detail is in the site's privacy policy: the provider is OpenAI, and the model receives only the text of the comment, without name, email or the fingerprint of the IP address.
Twice the text promised more than what was happening. The first version said AI wasn't used "in the texts you read here," which wasn't true: we corrected it on July 30. We found the second one while writing this article, and getting it wrong on the transparency page of all places didn't feel great. The page said a model flags comments and that "a human decides whether to publish it," but the code did something else: a comment the model judges fine goes live at once, and only the flagged ones, or those the model couldn't assess, wait for a person. The privacy policy described this correctly, the short page didn't. We corrected it on November 8, by writing down what actually happens rather than changing the code to match the sentence.
In a practice the risk is the same. The CNF template promises a "careful and accurate check" of every output: if an associate pastes a chatbot summary into an opinion without reading it, the notice the client signed says something false.
That page covers a website, not the projects we build for clients, so we'll say it here. Our developers get help from Claude, from Codex and from models running on our own computers: they write and review code inside work a person has already designed. How a piece of software is built and what it has to do is our decision, and the code that goes into production is reviewed by whoever signs off on it. Clients' personal data doesn't go through the models: we strip it out first, or that part of the work is done without AI.
What to do this week
- List the AI tools actually used in the practice, and ask associates and trainees too. Features inside software you already have count as well, such as automatic meeting transcription or the assistant built into your email.
- For each one, note the four things from the section on the notice: name, use, which client documents go into it, who checks.
- Take your professional body's template and replace the "might" with the tools on your list.
- Put it in the engagement letter for new work and send it to existing clients.
- Decide when you'll reread it, and update it before anyone adopts the next tool. On the page we wrote, we promised the same for the site's models.
If you can't put together the list in step 1, because someone else installed the tools or they run inside software you don't control, write to us. We start from what's on the practice's computers and accounts; the wording of the notice stays with you and your professional body.