Guide · AI for lawyers
AI and professional secrecy: using AI without exposing your client's file
The problem isn't AI. It's who receives the case file, under what contract, and what's kept afterwards. This guide walks through a real-world case and shows, clause by clause, how Custodio Legal answers those three questions.
At six in the evening a lawyer receives the defence to an employment claim: forty pages with the worker's name, ID number and salary, the medical records behind the sick leave, and the strategy the client shared at their first meeting. The hearing is tomorrow morning. She opens a general-purpose chat in her browser, signed in with her personal account, pastes the whole document and asks for a summary of its weak points.
The summary comes back in seconds, and it's good. What doesn't come back is an answer to three questions her client is entitled to ask: who received that text, under what contract, and what's kept of it.
Why it happens
Professional secrecy binds a lawyer to keep what they learn through their client, and in much of the region that duty also reaches the people who help them work the matter. Data protection law adds a second layer: for the third-party data that lives in a case file — the client, the other side, the witnesses — the firm is the data controller, and whoever processes that data on its behalf is a processor who has to act on its instructions.
A consumer chat opened with a personal account doesn't fit that picture:
- The person signed up for it, not the firm. What governs that upload are the terms that person accepted when they opened the account. Whatever they say about the text, how long it's kept or whether it's used for training, they say it to an individual user, not to a firm answerable for its client's data.
- There's no written mandate. Nobody set what the text may be used for, who else receives it, or what happens to it when the relationship ends.
- The text isn't tied to the matter. It sits in a chat history, mixed in with everything else that account has ever asked.
Bar associations are already putting this in writing. The guide published in July 2025 by the bar that registers lawyers in the city of Buenos Aires (CPACF) is blunt about it: avoid uploading case files with sensitive data to a tool that offers no guarantees, and without clear privacy guarantees, upload nothing confidential. For tools that don't offer them, its answer is to anonymise by hand — masking names, ID documents and references before pasting anything. It works, and it costs exactly the time the AI was supposed to save.
Every country has its own rule on professional secrecy, and every bar its own view. This guide doesn't replace them: what follows is what Custodio Legal puts in writing on its side, so you can hold it up against the rule that binds you.
How Custodio handles it
The same three questions, each answered with the published clause behind it.
Who receives the text, and under what contract
- The mandate is in writing. Section 4 (b) of the Terms says that for the data in the matters you manage, the firm is the data controller and Custodio Legal the processor, handling it on your behalf and on your instructions. That section is the processing agreement; where a country's law calls for a standalone instrument, it's published separately, in full.
- Only to provide the service. Section 4 (d) says your matter data isn't used for Custodio's own purposes, isn't passed to third parties for theirs, and isn't used to train models.
- Treated as privileged. Section 4 (h) says Custodio Legal treats your matter data as confidential information covered by your firm's professional secrecy, that its staff are bound to confidentiality, and that the obligation survives the end of the contract.
- Nobody signs in as you. Section 9 says the service has no feature that lets Custodio staff sign in as you or see your screens as if they were you: support works from what you choose to share.
- The list of who else receives data is public. Subprocessors are named one by one in the list of subprocessors, with what each receives and under which safeguard. If the text-generation provider changes, Custodio asks for your fresh authorisation before carrying on; for any other new subprocessor it gives you thirty days' notice (Terms, Section 4 (g)).
- Only the matter you pick is sent. Section 21 (f) says data from matters other than the one selected is never sent to the AI provider.
What's kept
- The original isn't kept. The file you upload is processed transiently to extract its content and then deleted; what remains is the extracted text and which document it came from (Terms, Sections 4 (a) and 9).
- The text is stored encrypted. The text extracted from each document, the passages the AI looks up, questions and answers, and drafts are all encrypted at rest with AES-256-GCM, and access is separated by firm and by role (Terms, Section 4 (i)).
- Zero retention at the AI provider. Per what that provider publishes, with the zero data retention Custodio has had switched on since 4 September 2026, what's sent to it isn't stored after each request is processed and isn't used to train models (Terms, Section 21 (h)). That's the provider's own statement, not an audit by Custodio; the guide Legal AI that doesn't train on your documents explains when it applies from and what it doesn't cover.
What the firm controls
- Features that produce something to read run on your click. Drafting and cited questions only run when you ask for them on the matter you're working on (personal data processing policy, Section 14 (viii)).
- What runs on its own can be switched off. Indexing for questions, fact extraction and matter reassessment run without a click, always on your own firm's matters; the firm owner can switch each one off in the firm settings, and switching indexing off also deletes the passages already indexed (same Section 14 (viii)).
Try the AI on a real matter, with the processing terms in writing
Open a free account, upload a document from a matter and read the Terms that cover it first. No card required.
Create your free accountWhat stays with you
A well-written mandate splits the responsibilities; it doesn't erase them. Under the Terms, the firm declares that it has the data subjects' authorisation, or another legal basis, to process the data it brings into its matters (Section 4 (e)), and it reviews every AI output before using it in any filing or professional act (Section 21 (b)). If your rules or your bar require you to tell your client you use AI, or to ask for their consent, that remains your call, made with your own country's rule in hand.
Confidentiality is one of the four problems with generic AI covered on the AI for lawyers page. The other three are the citations AI makes up when it answers from memory, AI that doesn't know the case file you're working on, and AI that acts without asking you first.
How to check before you decide
- Read Section 4 of the Terms of Service: it's the processing agreement, and it's what you'd show a client who asks.
- Go through the list of subprocessors and what each one receives.
- Open a free account, no card needed, upload a document from a real matter and check that what's kept is the text with its source, not the file.
Frequently asked questions about AI and professional secrecy
Does using AI on a client's file breach professional secrecy?
That depends on your country's rules and on what your bar says, and this page doesn't decide it for you. What does change the answer is who receives the text, under what contract and what's kept. At Custodio Legal that's in writing: the firm is the data controller, Custodio Legal is the processor, and it treats matter data as information covered by your firm's professional secrecy (Terms, Section 4).
Does Custodio Legal train models on my case files?
No. Section 4 (d) of the Terms says your matter data isn't used for Custodio's own purposes or to train models. On top of that, the text-generation provider has had zero data retention switched on for Custodio's organization since 4 September 2026, per what that provider publishes; the guide on legal AI that doesn't train on your documents explains what it covers and what it doesn't.
What's kept from the document I upload?
The extracted text and which document it came from, not the file. The original is processed transiently and then deleted (Terms, Sections 4 (a) and 9), and the extracted text is stored encrypted with AES-256-GCM.
Who else receives my matter data?
The subprocessors named in the public list at custodio.legal/subprocessors, with what each one receives. When an AI feature runs, only the matter you selected is sent, never another one. Custodio staff don't sign in to your account as you: support works from what you choose to share.
Use AI on your case files with the processing terms in writing
Open your free account, upload a real document and see what's kept. No card required.