What the rules of the profession say
In Romania, three texts set the frame:
- Law no. 51/1995, Article 11: the lawyer must keep professional secrecy regarding any aspect of the case entrusted to them, except in the cases expressly provided by law.
- Statute of the profession, Article 8: professional secrecy is a matter of public policy, and not even the client can release the lawyer from it (paragraph 3).
- Statute of the profession, Article 9: the duty is absolute and unlimited in time. It also binds the people the lawyer works with, and the lawyer must make them aware of it.
None of these texts mentions artificial intelligence. The duty does cover every aspect of the case, whatever the medium: the lawyer also answers for what they send to an AI tool.
Where the data goes when you use AI
An AI model usually runs on its provider's servers. Everything you give the assistant goes there:
- the question you type;
- the documents you upload, or that the assistant opens by itself;
- the conversation history, sent again with every new question.
What the provider does with this data depends on the contract: how long it is kept, in which country, and whether it is used to train models. The terms differ even between the free and the commercial version of the same product.
The exception is a local model, which runs on a server the firm controls. There the text goes nowhere.
What the CCBE guide recommends
On 2 October 2025 the Council of Bars and Law Societies of Europe (CCBE) published a guide on the use of generative AI by lawyers. The guide ties AI to the principles of the profession, above all confidentiality and professional competence. The guide is not binding, but it shows how the European bars, including the Romanian bar association (UNBR) as a CCBE member, look at the use of AI. Its basic recommendations:
- do not enter personal, confidential or client-related data without appropriate safeguards;
- safeguards can be: the provider's contractual obligation to keep the data confidential or not to retain it, a data processing agreement, technical measures, or running the model locally or in an environment the firm controls;
- read the provider's terms, to know how the data you enter is used;
- verify the output before using it in your work.
The guide also covers the lawyer's independence and informing the client about the use of AI.
Seven questions for any AI assistant
| No. | Question | Why it matters |
|---|---|---|
| 1 | Where does the model run: locally or at a provider? | This decides whether the text leaves the firm. |
| 2 | How long does the provider keep the data, and is it used for training? | Retention and training are the main risks. |
| 3 | In which country is the data processed and stored? | Transfers outside the European Union have rules of their own (Chapter V of the GDPR). |
| 4 | Is there a data processing agreement? | Article 28 of the GDPR requires one for every processor. |
| 5 | Is personal data removed from the text before it reaches the model? Where? | Anonymisation has to happen first, on a system of the firm. |
| 6 | Can the assistant send anything outside the firm without the lawyer's approval? | An e-mail that has been sent cannot be taken back. |
| 7 | Is what the assistant did kept in a log? | Without a log you cannot check anything afterwards. |
Anonymisation: useful, with limits
Taking names, personal identification numbers and addresses out of a document before it reaches the model lowers the risk a great deal. It does not remove it. A legal document can also show who it is about through context: facts, positions, places, case numbers.
In legal terms, a document from which only the names and personal identification numbers have been removed is in most cases pseudonymised, not anonymous: if the person can still be identified, the data remains personal data and the GDPR still applies (Recital 26).
That is why anonymisation is combined with the safeguards in the contract with the provider. For sensitive cases, such as those with health data, criminal cases or cases involving minors, the cautious option is a model that runs locally.
Verification stays with the lawyer
An AI model can write wrong things convincingly: a judgment that does not exist, an article of law with different content, a quotation attributed to someone else. The CCBE guide recommends verifying the output before using it, and the document remains that of the lawyer who signs it.
How the assistants in AVODesk work
- The lawyer approves. The assistant asks for the lawyer's agreement before any e-mail, any sharing and any document that is not anonymised.
- Anonymisation in the platform. Documents are anonymised on the firm's server before the assistants work with them. More: Anonymisation.
- Approved destinations. The assistants send data only where the firm has approved.
- No deletions. The assistants cannot delete clients, cases or documents.
- Everything is logged. Every action taken by an assistant is recorded.
- The model, where the firm chooses. The AI model runs locally or with a provider, depending on preference and budget.
The rules are described on the AI assistants and Security pages.