AI in law firms: secrecy, rules and the uses that hold up

01 Does professional secrecy apply to AI too?

Yes. Secrecy has no exception for tools. The law governing the legal profession binds lawyers to the “rigorous observance of professional secrecy” and to the utmost confidentiality about what they learn while assisting clients, in and out of court, and makes a breach a disciplinary offense (Law 247/2012, art. 6). The code of ethics repeats it in article 13 and spells it out in article 28: secrecy covers all the information received from the client, lasts after the engagement ends, and the lawyer must make sure employees, trainees, consultants and collaborators respect it too. A breach of professional secrecy carries a suspension of one to three years.

The code, even in its 2026 text, has no article dedicated to AI. It doesn’t need one for the most important question: when a lawyer pastes a client’s document into a chatbot, that information leaves the firm and goes to a provider, on the provider’s terms.

The guide of the Council of Bars and Law Societies of Europe (CCBE), published on 2 October 2025, says it plainly: lawyers should refrain from entering personal, confidential or other client-related data into a generative AI tool “unless there are appropriate safeguards in place”. And they should understand how the provider handles the data: whether it uses prompts to train the model, whether it shares data with third parties.

In practice, three questions to the provider make the difference. Is there a data processing agreement (the DPA required by article 28 of the GDPR)? Is the firm’s content used to train the models? Where is it stored? If the answers aren’t in writing, client documents stay out.

02 What Law 132/2025 requires of lawyers

Italy’s AI law, in force since 10 October 2025, has one article for all intellectual professions, lawyers included (art. 13). It says two things.

1. AI only supports. It is used “solely for ancillary and support activities”, and the professional’s own intellectual work must prevail. Research, summaries, first drafts: yes. The opinion, the strategy, the choice of what goes into a filing: no.
2. The client must be told. Information about the AI systems used must be given to the client “in clear, simple and comprehensive language”, to protect the relationship of trust. The law sets no form: what matters is that the client understands which tools you use and for what. The natural place is the engagement letter.

The same law also deals with judges (art. 15). When AI is used in judicial activity, every decision on the interpretation and application of the law, the assessment of facts and evidence and the adoption of measures always stays with the judge; until the AI Act is fully implemented, AI in the ordinary courts is authorized by the Ministry of Justice. It is the same idea from the other side of the bench: the machine helps, a person decides.

And there is the European AI Act: since 2 February 2025, an organization that uses AI systems in its work must take measures to ensure a sufficient level of AI literacy among its staff (art. 4). The lawyers’ code already has the duties of competence (art. 14) and of keeping up to date (art. 15), and the CCBE guide reads competence this way: it includes knowing how the tool you use works.

03 Invented citations: what happened in court

The most concrete risk is not science fiction. It is a ruling that doesn’t exist, written so well that it looks real. It’s called a “hallucination”: the model produces a plausible but false answer, such as a case number, a holding, a quote attributed to a judge.

The case that set the standard is American. In June 2023, in New York, Judge P. Kevin Castel sanctioned two lawyers and their firm, who had filed six non-existent opinions generated by ChatGPT and kept standing by them after the court questioned their existence: a $5,000 penalty, jointly and severally, and a letter to each judge falsely named as the author of those opinions (Mata v. Avianca). The judge wrote that there is nothing inherently improper about using a reliable AI tool, but that existing rules “impose a gatekeeping role on attorneys to ensure the accuracy of their filings”.

In Italy, in March 2025, it happened before the Florence court’s business section. In a dispute over T-shirts printed with protected cartoons, a defense brief cited Court of Cassation rulings that did not exist, or that dealt with something else. The defense lawyer explained that a collaborator had done the research with ChatGPT, without his knowledge. The court described the phenomenon of AI hallucinations and rejected the request to sanction the party for vexatious litigation (bad faith was not shown and no damage was claimed), but it put on record “the negative value of failing to check that the rulings actually exist”.

These are not isolated cases. The public database kept by Damien Charlotin, which collects decisions in which a court found that a party relied on material invented by AI, counted 2,184 cases worldwide on 10 October 2026, 15 of them in Italy.

And it is not only about general-purpose chatbots. A study by researchers at Stanford and Yale, published in the Journal of Legal Analysis, found that when asked “specific, verifiable questions about random federal court cases”, general models got it wrong between 58% (ChatGPT 4) and 88% (Llama 2) of the time. A second study by the same group found errors of this kind between 17% and 33% of the time even in the AI legal research tools of LexisNexis and Thomson Reuters, which answer from databases of case law. Fewer, but not zero.

04 Why answers with the source next to them matter

The lesson of these cases is not “don’t use AI”. It is that an answer without a source can’t be checked in time. An invented case number is only discovered by going to look for it, and once the filing is in, it’s too late.

Answers grounded in documents work the other way round. The AI first searches the documents you gave it (the firm’s opinions, filings, contracts, the rulings you downloaded and read) and answers only with what it finds there, pointing to the document and page for each statement. If something isn’t there, it says so. Checking takes one click: you open the cited page and read it.

It doesn’t eliminate errors, and the Stanford study shows it well: even tools built this way get things wrong, less often than general chatbots. But it moves the lawyer’s work from “find out whether it exists” to “read the cited page”, which is the check that falls to the lawyer anyway. We go deeper in why answers with sources are more reliable.

05 The uses that hold up, and the ones to avoid

Law firms that use AI already use it mainly to search and summarize. According to the Politecnico di Milano observatory on professional firms, in 2026 36% of Italian law firms use AI tools they chose themselves; of these, 68% use them to summarize and rework documents, 48% to analyze documents and prepare reports for clients, 34% to draft or review opinions, contracts and deeds, and 34% to translate. Only 4% use them to assist clients through a chatbot. Searching legislation and case law is the main use in every profession, at more than 80% of the firms that use AI.

To avoid: pasting client documents into personal accounts, filing citations nobody has opened, sending an opinion no lawyer has reread, letting a chatbot answer a client on the merits.

06 How to start in a law firm

1. Map what is used today. Who uses which tool, with which account, for what. No consequences: the point is to know, not to punish.
2. Choose tools that come with a contract. A signed DPA, content that doesn’t train the models, a stated storage location. Personal accounts stop being used on client files.
3. Update the engagement letter. A few clear lines on which systems you use and for what, and on who checks the result. It is the notice Law 132/2025 requires.
4. Start from the firm’s own documents. Templates, opinions, filings already made: they hold your know-how and the scope is easy to decide. Client files come later, with clear permissions on who sees what.
5. Write down the citation rule. No ruling goes into a filing unless someone at the firm has opened and read it. It applies to trainees and partners alike.
6. Train on your own cases. Twenty real questions from your work, each with a good and a bad answer: it is the literacy the AI Act asks for, and it teaches people to spot a suspicious citation.

This is how ekory works: Ekos, ekory’s AI agents, read only the documents the firm chooses, answer with the document and page, and say so when something isn’t there. The archive is stored in Europe and your content never trains the models. Concrete examples are on the page for law firms; the answers for your data protection officer are on the security page.

07 Note on sources

This article is not legal advice. It reports what the rules, the professional bodies’ documents and the decisions cited say as of 10 October 2026. Your own situation needs a professional who knows your firm.

• Laws: the text in force on Normattiva (Law 132/2025, Law 247/2012) and EUR-Lex (AI Act, GDPR). Translations into English are ours.
• Lawyers’ code of ethics: the 2026 text published by the National Bar Council (CNF), with its table of sanctions.
• Decisions: the Florence order in its full (anonymized) text published by Diritto.it; Mata v. Avianca in the document filed with the New York federal court. Other Italian rulings circulate in the legal press: we cite only those whose text we have read.
• Case count: Damien Charlotin’s database on 10 October 2026; it grows every week and includes some cases in which AI use was only alleged.
• Data on firms: Politecnico di Milano, 2 July 2026; shares by use refer to firms that use AI. The summary gives 37% to 47% across professions, the detail 36% for law firms: we report the detail.

Frequently asked questions

Sources

1. Legge 23 settembre 2025, n. 132, art. 13 (intellectual professions), Normattiva, 2025.
2. Legge 23 settembre 2025, n. 132, art. 15 (AI in judicial activity), Normattiva, 2025.
3. Legge 31 dicembre 2012, n. 247, art. 6 (professional secrecy), Normattiva.
4. Codice deontologico forense, 2026 text with sanctions, Consiglio Nazionale Forense, 2026.
5. CCBE guide on the use of generative AI by lawyers, Council of Bars and Law Societies of Europe, 2 October 2025.
6. Tribunale di Firenze, Sezione Imprese, order in appeal proceedings no. 11053/2024, March 2025 (anonymized text published by Diritto.it).
7. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions, U.S. District Court for the Southern District of New York, 22 June 2023.
8. AI Hallucination Cases database, Damien Charlotin, consulted on 10 October 2026.
9. Dahl, Magesh, Suzgun, Ho, Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, Journal of Legal Analysis 16(1), 2024.
10. Magesh, Surani, Dahl, Suzgun, Manning, Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 2024.
11. La spesa ICT degli studi professionali in Italia supera i 2 miliardi, Osservatorio Professionisti e Innovazione Digitale, Politecnico di Milano, 2 July 2026.
12. Regulation (EU) 2024/1689 (AI Act), articles 4 and 113, EUR-Lex, 2024.