Sending a confidential document to AI requires checking what you are allowed to share, limiting the content disclosed and assessing the receiving service. A pseudonymised copy can reduce exposure of identities, but it does not automatically make the file anonymous or authorise its disclosure. For consumer services, the CNIL recommends not submitting confidential information. Start with the task, then choose a suitable environment.
When can you share a confidential document with AI?
The answer depends on the content, recipient and your organisation's rules. Distinguish public documents, working copies approved for disclosure and confidential documents that must stay in a restricted environment. An assistant's upload button does not settle that question.
Does AI need the whole file to prepare a summary? Does it need signatures and contact details to compare two clauses? Write the request before selecting files. An excerpt or description will often be enough. Article 5 of the GDPR sets out data minimisation: process only what is necessary for the purpose.
| Your situation | Before sending | What to watch for |
|---|---|---|
| You want to improve some wording. | Copy the necessary passage or create a fictional example. | The sentence may still reveal a strategy or identity. |
| You want to summarise a matter. | Check the authorised environment and prepare a reviewed copy. | Links between documents must remain understandable. |
| The content cannot leave your organisation. | Work in the approved internal environment or without disclosing the document. | Replacing names does not lift a restriction on sharing. |
Why is removing names not enough?
A confidential document may contain two kinds of information: details that identify someone, and information your organisation wants to keep private. The first includes direct identifiers and context. The second may include a negotiated price, margin, internal method or position in a dispute. Removing a name does not address both needs.
The CNIL distinguishes anonymisation, which aims to prevent identification in practice, from pseudonymisation, which replaces identifiers while allowing attribution through additional information. A stable code such as [PERS1] helps track a person across several documents. It does not make the matter anonymous.
A fictional example, before and after
Suppose your confidential document states: “Camille Martin, the sole manager of the Saint-Malo site, disputes the change to their contract announced on 12 June.” Replacing the name with [PERS1] leaves a role and location that may identify the person.
If your question is only about the letter's clarity, a working copy could read: “An employee disputes a change to their contract. Rewrite this paragraph to distinguish the facts, the request and the points requiring clarification.” You preserve the drafting need while removing unnecessary details. If dates or the role determine the answer, keep them only within the authorised scope. Excessive generalisation could distort the analysis.
For a lawyer, the parties' names are therefore only one part of the review. Strategy, unusual facts and duties relating to the matter also count. The same reasoning applies to an internal report or client correspondence. The method starts with the intended use of the confidential document, rather than an automatic list of words to mask.
Which safeguards should you check with the AI service?
Check content reuse, retention and access separately. A promise not to use content for training does not, by itself, answer questions about storage, deletion or access. Look for the terms applying to your account and the particular feature, rather than a general claim about the brand.
- Has your organisation approved the account and intended use?
- What does the service do with files, conversations and responses?
- What access rights and retention periods apply?
- Do connectors or called tools introduce other recipients?
- Who approves disclosure when content contains personal data or confidential information?
Where a provider acts as a processor of personal data, Article 28 of the GDPR requires contractual safeguards. Buying a licence does not replace that check. The appropriate person must also examine sharing restrictions specific to the matter. Depending on the organisation, this may involve the DPO, security officer or project lead.
How do you prepare a confidential document in six steps?
1. Define the task before attaching the file
Specify the result: summarise, rephrase, identify differences or prepare questions. A focused request helps you select only the necessary pages. Avoid uploading an entire confidential document for a question that two paragraphs could answer.
2. Work on a clearly identified copy
Keep the original confidential document intact and create a copy for the task. Give it a filename that does not reintroduce an identity or restricted reference. Distinguish it clearly from the original to avoid sending the wrong file.
3. Remove or replace unnecessary information
Check names, contact details, identifiers, signatures and indirect references. For multi-document matters, use consistent labels to preserve useful relationships. Keep the mapping table separate from shared content. Also review internal information that is not personal data.
4. Check the exported file, not just the preview
Reopen the version you intend to share. Search its text for names and contact details, try copying treated passages and inspect annotations. For scans, check both the image and OCR text layer. Also review comments, metadata and attachments where the format supports them. These practical checks complement review without providing exhaustive proof of anonymity.
5. Review the request and check your selection
A properly prepared confidential document may be accompanied by a prompt that repeats a removed name. Check the question, selected files and items added by connectors. Do not send the mapping table to explain codes to the model.
6. Review the response and decide how to retain it
Compare the response with the supplied passages. A summary may omit a qualification or invent a connection. A response based on a confidential document may also reproduce restricted information: decide where to keep it and who can access it. Record which working copy was used if you need to retrace your process.
Should you send the PDF, Word file or just an excerpt?
The best format provides the necessary information with the least unnecessary material. A text excerpt often suffices to rephrase a paragraph. Comparing tables may require their structure. Layout analysis may require the file itself. Choose the format after defining the task.
Converting a confidential document to PDF does not sanitise it. Printing, taking a screenshot or placing a rectangle on a page does not guarantee that all restricted information has gone. Check the final copy with tools suited to its format. To understand the difference between covering a passage and removing its content, read our comparison of redaction and pseudonymisation.
Official guidance to remember in October 2026
This guide draws on official sources consulted on 13 October 2026. The CNIL's generative AI FAQ, published on 18 July 2024, distinguishes non-confidential uses from processing involving personal or strategic data. It recommends defining authorised tools and uses. Its advice against sharing confidential information with consumer services remains a useful reference.
Articles 5 and 6 of the GDPR address purpose, minimisation and lawfulness, among other requirements. Articles 28 and 32 cover processor arrangements and security measures. These obligations do not disappear after a format change or name replacement. A useful procedure therefore connects file preparation with approval of the recipient.
How can you prepare this copy with Lexanon?
Lexanon helps you prepare confidential documents directly on your computer. Banco detects personal data and proposes replacements for you to check and correct. Consistent pseudonyms preserve useful links across the matter. Context and restricted information still require review.
You then choose approved documents to use with the tools and models available in your environment. Local preparation avoids sending the original confidential document to a server for this step. If you then choose remote AI, the content you send does leave your computer: check that recipient's safeguards. Our guide to local pseudonymisation explains this workflow.
Start with a representative working copy and check the results before extending the method to other documents. Explore Lexanon's preparation and review features.
Frequently asked questions
Can I send a confidential document to free AI?
Price alone does not determine safeguards. Before sending anything, check that the service is authorised by your organisation and suited to the content. The CNIL recommends not sharing confidential information with consumer services. A paid subscription alone does not authorise disclosure.
Is removing names enough to protect a document?
No. An address, an unusual job title, a date or a combination of details can still identify someone. A document may also contain confidential business information without identifying anyone. Review the copy, remove unnecessary details and check the context. Replacing names with codes generally amounts to pseudonymisation.
Is a pseudonymised document still subject to the GDPR?
Yes, when people remain identifiable using additional information. Pseudonymisation reduces direct identification, but the data remains personal. Keep the mapping table separate from the shared file. Processing must still meet applicable requirements, including its purpose, legal basis and the recipient's safeguards.
Is a PDF with black rectangles ready to send?
Not necessarily. A rectangle drawn on a page may merely cover text that remains recoverable. Use a redaction function that actually removes the relevant content, then check the exported file. Also test search and copy-and-paste, and inspect annotations and metadata. Appearance alone is not enough.
Does locally installed AI eliminate every risk?
Truly local execution limits disclosure to a remote provider. It does not remove risks relating to computer access, backups or external connectors. Check the full data flow and users' permissions. Model responses must also be reviewed: local execution does not guarantee accuracy.
What if a confidential file has already been sent?
Stop sending files and report the situation to the responsible person in your organisation. Record the service, account, time and information disclosed. Ask about deletion options without assuming they remove every copy. The appropriate person then assesses the consequences and any obligations, particularly where personal data is involved.
Three decisions before sending
For each confidential document, decide what the task needs, what can be shared and with whom. Then prepare a copy, check the export and review your request. These decisions, together with the chosen environment's safeguards, give file protection its purpose.
Sources
- CNIL, questions and answers on using generative AI systems, 18 July 2024, particularly questions 5 to 7.
- CNIL, anonymisation of personal data, 19 May 2020.
- GDPR, Chapter II, Articles 5 and 6, text presented by the CNIL.
- GDPR, Chapter IV, Articles 28 and 32, text presented by the CNIL.
- Adobe, sanitise a PDF and remove hidden information, technical documentation on metadata, comments and hidden content.
This article provides a preparation method and general information. It is neither personalised legal advice nor permission to disclose a matter. The examples are fictional.