Five Questions From the Paris Bar's AI Charter

Use the Paris bar's AI charter to audit any AI tool before sending client data.

By RAGbase Published on Updated on 5 min read

The Paris bar published a model AI charter in July 2026. A lawyer who reads it lands quickly on article 4, case 2: an online tool is allowed after validation or sufficient guarantees, without confidential or sensitive data, with systematic human review. Article 10 adds security guarantees, including encryption and logging. Those two articles give a firm a real standard, not a slogan. The problem is that “sufficient guarantees” is not a checkbox. It has to be turned into questions a firm can actually ask about Claude, or any other AI tool, before a single client file goes near it.

This piece turns the charter into five such questions. We answer each one for RAGbase Shield, the local, reversible pseudonymisation tool made by RAGbase, because it is the tool we know in detail. The answers for another product will differ, and that is fine: the point of the charter is that every firm checks, every time.

Question one: has the tool been validated, or does it meet the charter’s own standard

Article 4, case 2 of the model charter gives two paths: formal validation inside the firm, or guarantees judged sufficient. Shield does not decide this for the firm. It is free and open source under the Apache-2.0 licence. What it can offer the firm’s AI lead is the source code itself, available to read and audit, and a compliance sheet that says plainly: “This sheet is not legal advice. It describes a tool. Compliance with professional secrecy and the GDPR remains the responsibility of the lawyer and the firm.” Validation, in the sense the charter means it, stays a decision made inside the firm, informed by that audit.

Question two: does confidential or sensitive data ever leave the computer

This is the heart of article 4. Shield’s architecture answers it directly: a local engine of rules, dictionaries and check digits runs on the lawyer’s own computer, in French and in English, before any text reaches an AI provider. Shield makes no network call and sends no usage statistics. Shield masks the identity, not the secret itself. That sentence matters here: a name replaced by a code is not the same as a fact becoming harmless. A fact covered by professional secrecy, under article 66-5 of law 71-1130 of 31 December 1971 and article 2 of the national internal rules, stays covered after the name is masked. Detection is never complete, so a firm cannot treat the tool as a final filter. It is one layer, run locally, before the text leaves the machine.

Question three: is there systematic human review of the output

Article 4 asks for review, not for trust in the machine. Shield cannot do this part, by design. The real names come back only on your computer, in the final document that Shield saves in the “Shield - sorties” folder, but Shield does not read the answer for accuracy or for secrecy. That review stays a human task. This is also where the limit about sent text bites: text already sent cannot be recalled. Text typed or pasted straight into the conversation is not masked, and no later step undoes that. Shield’s own rules say the same: choose the client folder in the Shield window, and refer to people by their code. The charter’s human review requirement points the same way: the lawyer stays the last filter, not the tool.

Question four: what does article 10 require for encryption and logging

Article 10 names encryption and logging as concrete guarantees, not abstractions. Shield keeps one encrypted vault per dossier, built on SQLite and Fernet encryption, with the key stored in the Mac keychain or in the Windows Credential Manager. The local journal records what happened, when, and how many times, but never a real name. A firm’s AI lead can read that journal to see activity without seeing client identities. None of this is a promise of compliance. What the architecture gives the firm is a concrete answer to a concrete question: where is the key, who can read the vault, what does the log actually record.

Question five: can the tool be audited, and by whom

A model AI charter written for a whole bar has to assume firms differ in technical capacity. Some have an AI lead who can read code; most do not. Shield’s answer is that the source is open, so a firm with the right skills, or an outside auditor it trusts, can check the claims instead of taking them on faith. That does not replace the four previous questions. It makes them checkable rather than asserted. For a firm without that capacity, the published facts still describe what happens. Shield runs on the lawyer’s computer, inside the Claude app. Its tools take only codes and dossier identifiers, never raw text. That holds even if nobody inside the firm reads the code line by line.

What this checklist does not settle

Five answered questions are not a cleared tool. For the lawyer, pseudonymised data remain personal data, even once names become codes like PERSONNE_A. A well-known case, a unique job title, or a precise amount can still identify a client with no name attached at all. Recommendation 2026-01 of 2 April 2026, from the ethics board of the Avocats au Conseil d’État et à la Cour de cassation, makes the underlying point without naming any product: outside a closed AI, do not transmit elements that identify the client, though an abstract legal question does not breach secrecy on its own. A checklist built from article 4 and article 10 helps a firm ask the right questions before it signs anything. It does not answer them. That stays the firm’s job, dossier by dossier.

The limits of Shield are on the Limits and compliance page.

Sources

https://www.avocatparis.org/sites/bdp/files/2026-07/Mod%C3%A8le%20charte%20IA.docx; https://ragbase.ai; docs/CONFORMITE.fr.md and README.md of RAGbase Shield.

Tags: Professional secrecyAI charterCompliancePseudonymisationAuditing

Frequently asked questions

What does the Paris bar's AI charter say about using AI tools with client data?
Article 4 requires either formal validation or sufficient guarantees, with no confidential or sensitive data sent, and systematic human review. Article 10 adds requirements for encryption and logging. These provide a real standard, not a slogan, but firms must turn abstract language into concrete questions.
What is pseudonymisation and why does it matter in Shield?
Shield replaces names with codes before text reaches an AI provider, keeping professional secrecy intact. However, the article emphasizes that pseudonymisation does not make data harmless; unique details like case facts or job titles can still identify clients even without names.
Can Shield prevent all breaches of professional secrecy?
No. Shield is one local layer that runs before text leaves the machine, but detection is never complete. The lawyer remains the final filter and must apply human judgment to avoid sending identifying details, even if names are masked.
RAGbase
RAGbase Shield team

RAGbase builds RAGbase Shield, a free, open-source tool (Apache-2.0) for local, reversible pseudonymisation, and private AI systems installed at law firms.

Back to the blog