AI and legal documents: the GDPR risk starts before the prompt

AI and legal documents: the real risk starts before the prompt

Short answer: a law firm should not start with “which AI tool should we use?”. The first question is: which parts of the file may be exposed, which ones must be pseudonymized, and who validates the working copy before analysis?

The promise is obvious: summarize a matter, prepare a chronology, compare contracts, extract obligations, review a clause, draft a memo or save time on a first-pass analysis. But the problem starts before the drafting stage. It starts when the raw document is copied, uploaded, converted or pasted into a system.

Extractable block: AI becomes useful for legal work when documents are prepared before use: clean extraction, pseudonymization of identities, separate correspondence table, human validation, then analysis on a protected version. Without that chain, the firm may save time while increasing leakage, confusion or uncontrolled-use risk.

The real issue: compliance, GDPR and professional secrecy

Sharing a raw client file with Claude, ChatGPT or any AI assistant is not a neutral gesture. It may involve transferring personal data, confidential exhibits, defence strategy, correspondence, privileged information and sometimes sensitive data under the GDPR.

The painful question is simple: if the client, an opposing party, a regulator or a bar authority asked tomorrow “where was this file sent, why, on what basis, with which safeguards and with which trace?”, could the firm answer cleanly? If not, the problem is not AI. The problem is the lack of governance.

Risk point: the GDPR allows fines of up to €20 million or 4% of annual worldwide turnover in the most serious cases. For a firm, the real risk is not only the fine: it is loss of client trust, challenge to the method, breach of secrecy and inability to explain what was shared.

The false debate: “can law firms use AI?”

The real answer is not yes or no. The real answer is: under which framework, on which data, with which validation, for which deliverable, and with which responsibility limit?

A law firm can have strong AI use cases. But a legal file is not marketing copy. It contains names, addresses, exhibits, dates, amounts, strategy, sensitive exchanges and sometimes family, health, employment or litigation data. Not everything should enter a tool as-is.

The problem many firms underestimate

The risk is not only the client name on page one. Leakage often comes from secondary details: file names, footers, scanned signatures, exhibit references, Word comments, tracked changes, exact addresses, matter numbers or combinations that make the case identifiable.

Removing three names is therefore not enough. Serious protection must preserve legal reasoning while removing information that unnecessarily identifies people, companies or the dispute.

What must be preserved for AI to remain useful

Confidentiality should not destroy the file. For legal work, the chronology, roles, obligations, clauses, deadlines, useful references and relationships between parties must remain usable.

The right model is not “hide everything”. It is stable pseudonymization: CLIENT_001, OPPOSING_PARTY_001, CONTRACT_001, EXHIBIT_001. AI can reason about relationships without knowing real identities.

The layer the client does not see, but that changes everything

A serious setup separates three spaces: the raw file, the protected working copy and the correspondence table. The table that can restore real names does not travel with the document used for analysis. It stays in a vault or confidential area.

This separation is the difference between “we use AI on a sensitive file” and “we have a controlled method to prepare a usable version”.

Pseudonymization, anonymization, vault: the foundation of the Company Brain

The risk is not only a badly written AI prompt. The real risk appears when a sensitive file circulates too quickly, without sorting, separation or clear rules about what AI is allowed to see.

This is where the difference between pseudonymization, anonymization and a vault becomes strategic. The firm must be able to work on a usable version of the file while keeping the most sensitive elements in a controlled space.

When designed properly, this becomes the foundation of a Company Brain: a working and steering surface where teams can understand priorities, risks, questions to handle and next actions, without unnecessarily exposing the raw file.

The point is not to give more documents to AI. The point is to create a controlled chain, tailored to the firm, that makes AI useful without weakening professional secrecy. This is exactly the kind of system Say Digital helps frame, prototype and deploy.

What Say Digital builds for this type of firm

Say Digital does not sell a magic prompt to law firms. The subject is more operational: creating a document-preparation chain that saves time without turning every matter into an avoidable risk.

• conversion of exhibits into usable text;
• detection of sensitive zones;
• stable pseudonymization;
• protection manifest;
• firm validation;
• analysis or drafting on the protected copy;
• controlled restoration of the final deliverable when required.

We intentionally do not reveal the full mechanics here. What matters for a firm owner is the problem: AI becomes risky when documents enter tools too quickly, without preparation or control.

Warning signs in a law firm

• Team members paste file extracts into several different tools.
• Word documents still contain comments or tracked changes.
• Names are hidden, but addresses or references remain visible.
• The same matter is summarized multiple times without a reference version.
• No one knows where the mapping between pseudonyms and real names lives.
• The final deliverable is restored without a leakage check.

The right commercial angle for a firm

The firm does not need to buy “AI”. It needs a system that prepares documents, preserves useful roles, protects identities, produces drafts and leaves the lawyer in control.

This distinction matters: AI does not replace professional judgment. It prepares, classifies, summarizes, compares, flags and drafts a first version. The professional keeps validation, strategy, advice and client-facing delivery.

FAQ

Can a law firm use AI on sensitive files?

Yes, but not by pushing raw documents into tools without a framework. The safer path is to work on protected versions validated and limited to the actual task.

Is pseudonymization enough?

No. It must come with a separate correspondence table, human validation, metadata checks and a final review before sharing.

Why not simply anonymize all matters?

Because full anonymization can break legal reasoning. In many cases, stable pseudonymization preserves roles and chronology without exposing real identities.

Next step: Say Digital can help a firm set up a first protected workflow on a few matters: document preparation, pseudonymized version, test deliverables and internal validation method.

Sources and resources

• CNIL — personal-data anonymization
• CNIL — personal data security
• GDPR — official text
• French National Bar Council — professional secrecy

Version française : IA et documents d’avocat : le risque RGPD commence avant le prompt