AI for Italian lawyers. What you can do without putting professional secrecy at risk.
One in three lawyers already uses AI, and they all share the same doubt: does uploading a client’s pleading to ChatGPT breach professional secrecy? Here is the answer, with the rules, and the client notice template the law asks for.
Yes, for ancillary and support tasks, as long as the lawyer’s own intellectual work remains predominant. Article 13 of Italian Law 132/2025 says so, in force since 10 October 2025, and it also requires the lawyer to tell the client which systems are used.
The National Bar Council (CNF) issued a model notice on 13 October 2025, for use where helpful. The Code of Conduct for Lawyers does not mention artificial intelligence: the secrecy and confidentiality duties in Article 28 apply, and Law 247/2012 makes them a legal obligation.
Secrecy comes down to the tool, and the defence to checking. On ChatGPT’s personal plans, content is used to train the models unless you opt out, and any judgment cited must be opened at the source before it goes into a pleading. Below is the client notice template in Word and PDF, in Italian and English, to download without leaving an email address.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
The full picture of Italy’s AI law is on the page about Law 132/2025, and the rules for the other intellectual profession that uses it most are on the page about AI for accountants. This page covers the part that affects a law firm.
The client notice template, to download
A two-page sheet to hand over with the letter of engagement: which systems the firm uses, what for, on which data, and what the client can ask for. The CNF model states that the lawyer might use AI for support tasks and will check the results. The template keeps those statements and adds the table of systems, one row per tool.

| File | Contents | Link |
|---|---|---|
| ItalianWord, to fill in | The notice with the table of systems, what AI does not do, the lawyer’s checks, confidentiality and the client’s rights. | |
| EnglishWord, to fill in | The same sheet in English, for a client of the firm who doesn’t read Italian. | |
| ItalianPDF | The copy to print. | |
| EnglishPDF | The copy to print, in English. | |
| ExamplePDF, filled in, in Italian | A fictitious law firm handling a debt recovery with two systems: an external service on a business plan and a model run on its own computer. |
The template also adds a line the model doesn’t have: every judgment or provision cited in a pleading is opened in the official source before signing. The client can check that promise, and the Florence case below shows why it matters.
What AI can do in an Italian law firm
One lawyer in three already uses it for work: 36% of respondents, according to the Ipsos survey of 2,532 lawyers for the National Bar Council, published on 15 October 2025. The share rises to 47% among those aged 35 to 44 and to 76% in firms with more than twenty staff, against 35% in smaller ones.
The same survey shows where trust runs out. The tasks seen as least risky are case law research and document analysis; drafting pleadings is judged unreliable by 72%, and 52% do not consider AI reliable for legal work. On 9 October 2025 the CNF launched a market consultation on AI services for lawyers.
The tasks a system can take on are the ones that save hours without touching judgement. In each one the machine prepares and the lawyer decides, and the right-hand column is the part no rule lets you delegate.
| Task | What the system does | What stays with the lawyer |
|---|---|---|
| Case law research | Finds and summarises rulings on a question. | Opens every judgment in the official database before citing it. |
| Case file reviewcontracts, correspondence, invoices | Reads the documents and builds an index of dates, parties and amounts. | Decides what is relevant and what needs proving. |
| Chronology of events | Puts the events shown in the documents in order. | Checks every date against the original document. |
| Letters and formal demands | Prepares the first draft. | Rewrites it, signs it and decides whether to send it. |
| Summaries for the client | Summarises a court order in plain language. | Checks that the summary says what the order says. |
| Deadlines | Reads time limits from notices and orders and puts them in the diary. | Calculates the time limits and answers for them. |
| Comparing contract versions | Highlights the clauses that changed between two drafts. | Assesses what the change means for the client. |
| Foreign-language documents | Translates correspondence from a foreign counterparty. | Decides where a professional translation is needed. |
| Front office | Sorts incoming requests and suggests appointment dates. | Chooses whom to see and when. |
Several steps chained together, from the incoming document to the case file index, are the work of an AI agent: a system that carries out a sequence of steps and stops where a person is needed. How it differs from an assistant you ask one question at a time is explained on the page about AI agents and chatbots.
What is never delegated: interpretation and strategy
Interpreting the law, assessing the facts and deciding the litigation strategy: Law 132/2025 allows AI only for ancillary and support tasks, with the lawyer’s intellectual work predominant. Article 13 sets this for every intellectual profession, and for lawyers it sits alongside the duty of competence in Article 14 of the Code of Conduct.
On the judge’s side the same law is even clearer. Article 15 provides that when artificial intelligence systems are used in judicial work, every decision on interpreting and applying the law, assessing facts and evidence and issuing orders is always reserved to the judge.
Lawyers agree. For seven in ten respondents to the Ipsos survey for the CNF, AI cannot yet interpret laws and precedents correctly or produce reliable legal work. Whether AI will replace lawyers has an answer in the rules: the work remains the lawyer’s, and the way of working that guarantees it is called human in the loop.
Invented case law: the Florence case
On 14 March 2025 the Florence Business Court ruled on an appeal in which a defence brief cited Court of Cassation judgments that do not exist. A junior at the firm had found them with ChatGPT, and the lawyer did not know. The court refused to award costs for vexatious litigation, because the defence had been the same since first instance, but it criticised the failure to check.
This is the first risk named in the guide published by the Council of Bars and Law Societies of Europe (CCBE) on 2 October 2025: a generative AI system can produce entirely fictitious case law, create decisions that do not exist or misattribute quotations. The phenomenon has a name, AI hallucination, and it does not depend on how carefully the question is written.
Where there is bad faith or gross negligence, the bill is set by Article 96 of the Italian Code of Civil Procedure: on top of damages, the court orders the party to pay between €500 and €5,000 to the Fines Fund, in proceedings started from 30 June 2023. Opening every citation at the source takes a few minutes.
ChatGPT and professional secrecy: where the data goes
A lawyer’s secrecy is a legal obligation: Article 6 of Law 247/2012 imposes it towards third parties, in the represented party’s interest, and extends it to employees, collaborators and trainees. Breaching it is a disciplinary offence, and Article 622 of the Italian Criminal Code punishes anyone who reveals, without just cause, a secret learned through their profession.
Article 28 of the Code of Conduct makes it a primary and fundamental duty, one that lasts after the engagement ends. Breaching confidentiality carries a reprimand; when it concerns professional secrecy, suspension from practice for one to three years.
The practical question is where a document ends up once it is uploaded to an AI service. ChatGPT’s plans page states that on the Free, Go, Plus and Pro plans content is used to train the models, with an option to opt out; on the Business plan business data is not used for training by default.
The CCBE guide adds two warnings. First, lawyers are responsible for what they enter, and can breach their obligations at the very moment they type the request. Second, if the same system works for several of the firm’s clients, one client’s information can end up in work for another, creating a conflict of interest. A model run on the firm’s own computers avoids both for documents that must not leave the firm.
What to ask an AI provider before uploading a pleading
Eight questions, with a written answer before the tool is used on client documents. They come from Articles 22, 28 and 32 of the GDPR and from the CCBE guide’s warnings: a provider who can’t answer on one point leaves exactly that point uncovered, and secrecy remains the lawyer’s duty.
| Question | The answer you need |
|---|---|
| Is the content used to train the models? | No, by default, and in writing in the contract. |
| Is there an Article 28 GDPR agreement? | Yes, with the list of sub-processors and a commitment to ask permission before adding any. |
| In which country is the data processed? | The country, and the legal basis for any transfer outside the European Union under Articles 44 to 49 of the GDPR. |
| How long is it kept? | A specific period, and a way to delete it sooner. |
| Is each client’s data kept separate from the others’? | Yes: no reuse across the firm’s clients, to avoid the conflict of interest the CCBE describes. |
| Can the cited sources be opened? | Yes: every judgment or provision comes with a reference you can check in the official database. |
| Do the accounts belong to the firm? | Yes: one account per person, opened and closed by the firm, not trainees’ or collaborators’ personal accounts. |
| What happens to the data if we stop? | Deletion or return, as the firm chooses, under Article 28(3)(g) of the GDPR. |
The same questions, seen from the side of a business buying a custom AI system rather than a subscription, are in the guide on how to choose an artificial intelligence agency. How the two European regulations fit together is on the page about the AI Act and the GDPR.
The client notice: the law and the CNF model
Information on the artificial intelligence systems used must be given to the client in clear, simple and complete language. Article 13(2) of Law 132/2025 requires it, to safeguard the relationship of trust between professional and client, and on Normattiva the text is in force, unamended, as of 8 November 2026.
The CNF model, sent to the local Bar Councils on 13 October 2025 under reference 7816, lists the permitted support uses: organisational and secretarial work, research into legislation and case law, preliminary document analysis, drafts and summaries. It states compliance with the GDPR and with ethical duties, the checking of results and sources, and asks the client to sign for acknowledgement.
The CCBE guide says when transparency becomes a must even outside Italian law: when it is reasonable to assume that an informed client would object, set conditions or have reservations about using AI for a particular purpose. The template on this page answers that case before it arises, with the list of systems and the right to ask for their exclusion.
Decree 179/2026: training and fees
Since 23 October 2026, Legislative Decree 179/2026 has required the Bar Councils to include artificial intelligence in lawyers’ training, with a minimum number of hours in every continuing education period. Article 47 gives the national councils, the National Bar Council included, six months to update their rules: until 23 April 2027.
The content is set by law: how the systems used in the profession work, including techniques for querying and instructing them, with their potential and limits; the rules, starting with the European regulation; and the ethical side, meaning the professional’s responsibility, the duty to inform the client and the human-centred principle.
Article 49 deals with fees: a professional’s use of AI systems may lead to an adjustment of the fair pay set by Law 49/2023, with parameters tied to the risk class of the system used. The decrees setting the parameters, including the one on lawyers’ fee parameters under Article 13(6) of Law 247/2012, must be updated within twelve months: by 23 October 2027.
How to start without taking risks
With a list of the tools the firm already uses, including free accounts opened by trainees and collaborators. Law 247/2012 requires lawyers to make sure these people respect secrecy too, and the Florence case shows that research done by a collaborator ends up in the pleading the lawyer signs.
One, choose the tool for client documents: a business plan with the Article 28 agreement, or a model run in the firm. Two, remove the data you don’t need: party names, tax codes and amounts replaced before use, where the task allows. Health data and criminal records never go in.
Three, a rule on citations: no judgment goes into a pleading unless someone has opened it in the official database. Four, hand over the notice with the letter of engagement and update it whenever a system changes.
Five, train the people who use it. Article 4 of the AI Act requires deployers of artificial intelligence systems to take measures for their staff’s AI literacy, as the page on the AI literacy obligation explains, and lawyers’ continuing education will include it under Article 47 of Decree 179/2026. Six, start with a single task, with the lawyer checking every output, and measure the time before and after.
What it costs, and what mistakes cost
A business plan for an AI assistant is paid per user, and the price changes often: check the provider’s price list on the day you decide. A local model shifts the cost to hardware and to the time of the person who sets it up. For a service aimed at its members, the market consultation the CNF launched on 9 October 2025 asks providers for solutions, costs and pricing models.
The cost of a mistake, on the other hand, is written into the rules. An invented citation, with bad faith or gross negligence, leads to the €500 to €5,000 payment under Article 96 of the Code of Civil Procedure, on top of damages. Breaching secrecy leads to suspension for one to three years under Article 28 of the Code of Conduct.
On the data side, the GDPR fines breaches of the obligations of controller and processor, including Articles 28 and 32, up to €10 million or 2% of worldwide annual turnover, and breaches of the principles and of the rules on transfers up to €20 million or 4%. These are the maximums in Article 83.
If a system talks directly to clients, for example an assistant on the firm’s website, the transparency obligations in Article 50 of the AI Act apply. Article 23 of Decree 179/2026 punishes breaches with fines of up to €15 million or 3% of worldwide annual turnover, and for small and medium-sized businesses the lower of the two applies.
Questions and answers
Can an Italian lawyer use ChatGPT?
Yes, for ancillary and support tasks, as long as the lawyer’s own intellectual work remains predominant, as Article 13 of Law 132/2025 says. The same provision requires the lawyer to tell the client which systems are used.
For client documents, the plan matters. On ChatGPT’s personal plans, content is used to train the models unless you opt out; on the Business plan it is not used for training by default.
Does using artificial intelligence breach a lawyer’s professional secrecy?
It depends on where the data goes. The CCBE guide on generative AI warns that lawyers can breach their obligations at the very moment they enter information into the system, and that the provider may reuse the data to train the model unless agreed otherwise.
Secrecy is a legal obligation, under Article 6 of Law 247/2012, and an ethical duty, under Article 28 of the Code of Conduct: breaching it leads to suspension for one to three years.
What must an Italian lawyer tell clients about using AI?
Which artificial intelligence systems they use, in clear, simple and complete language, as Article 13(2) of Law 132/2025 requires. The notice covers both the client and the represented party.
The National Bar Council issued a model notice on 13 October 2025, for use where helpful. A list of the systems, with what they do and on which data, completes it.
What does a lawyer risk by citing case law invented by AI?
Aggravated liability under Article 96 of the Italian Code of Civil Procedure, where there is bad faith or gross negligence: on top of damages, the court orders a payment of between €500 and €5,000 to the Fines Fund.
In the case decided by the Florence court on 14 March 2025, the order was refused because the invented citations did not change the defence; the court still criticised the failure to check.
What is the best AI for a lawyer?
The one that meets three conditions before any feature: it does not use client documents for training, it comes with a data processing agreement under Article 28 of the GDPR, and it shows its sources so the lawyer can open them.
Among the tools that meet them, you choose by task: a writing assistant for drafts and summaries, a legal database with assisted search for case law, a local model for documents that must not leave the firm.
Notes on sources
- Italian Law no. 132 of 23 September 2025, Article 13 and Article 15, on Normattiva: text in force as of 8 November 2026, in force since 10 October 2025, unamended. The sentence on the relationship of trust is a faithful paraphrase.
- Italian Law no. 247 of 31 December 2012, Article 6, on professional secrecy; Italian Criminal Code, Article 622; Italian Code of Civil Procedure, Article 96, with the fourth paragraph added by Legislative Decree 149/2022. Texts in force as of 8 November 2026, on Normattiva.
- National Bar Council, Code of Conduct for Lawyers, 2026 update with sanctions: Articles 13, 14 and 28. The text was read in full: the words “intelligenza artificiale” do not appear. The phrase from Article 28 is a faithful paraphrase.
- National Bar Council, letter to the local Bar Councils of 13 October 2025, reference 7816, with the model notice republished by the Milan Bar; resolution of 9 October 2025 on the market consultation for AI services.
- National Bar Council, Ipsos survey “Avvocati e attualità: Intelligenza artificiale”, 2,532 lawyers, published on 15 October 2025.
- CCBE, guide on the use of generative artificial intelligence by lawyers, 2 October 2025, in the informal Italian translation by the CNF office in Brussels: sections 3.3 (risks), 4.1 (confidentiality), 4.4 (transparency), 4.5 (conflicts of interest).
- Florence Business Court, order of 14 March 2025, appeal no. 11053/2024: the facts come from the commentary in Agenda Digitale, because the text of the order is not published on an open source.
- Italian Legislative Decree no. 179 of 7 October 2026, on Normattiva, in force since 23 October 2026: Article 47 (training for the professions), Article 49 (fair pay), Article 23 (penalties). The deadlines are counted from the date of entry into force.
- Regulation (EU) 2016/679, the GDPR, Articles 22, 28, 32, 44 to 49 and 83(4) and (5), on EUR-Lex; OpenAI, ChatGPT plans page, read on 8 November 2026.
- The template and the example are mock-ups built by Itria on the texts above; the example uses fictitious names and two systems chosen as examples, not as purchasing advice. All sources were opened and read on 8 November 2026.
The machine prepares the case file, the lawyer decides. And every citation gets opened before signing.
An AI agent can read the documents that reach the firm, recognise their type, pull out dates, parties and amounts and build the case file index; the lawyer checks and approves before anything goes out. The piece that reads documents recognised the right type thirty times out of thirty and got 98.7% of fields right, across invoices, delivery notes, spec sheets and orders, in 6.3 seconds each: measured on a test bench on 4 September 2026, on a test system. Where documents must not leave the firm, the model can run on local or dedicated infrastructure.