Which data should you remove before sending a document to AI?

Date: 8 November 2026. Author: Riducta Editorial Team. Publisher: X20 Digital Trust Services S.r.l.

Before sending a document to an artificial intelligence service, remove information unnecessary for the task and check details that could identify someone even without a name. Work on a copy, verify the text actually exported and keep any restoration map separately. Masking some names reduces exposure, but does not automatically authorise sending the document or guarantee anonymity.

The first decision concerns the task. Rewriting a clause may require only that clause, without the full contract. Improving a letter may require only an invented description of the situation. If the result does not need real data, a synthetic example avoids transferring unnecessary information. The following are general organisational criteria, rather than a legal assessment of particular processing.

Establish what the service needs to know

Describe your intended result in one sentence: correcting style, comparing two texts or extracting deadlines. Assess each document section against that result. A home address rarely helps improve a communication's grammar; a date may be necessary to compare time limits. Removing everything indiscriminately may make the answer unusable, so document substitutions relevant to the work.

Before uploading, check whether the organisation permits use of that service and specific account. Provider settings, contractual terms and project access form part of this decision. The fact that an interface accepts attachments does not mean the transfer has internal approval. If necessary information is missing, pause the sending of real material.

Look for direct identifiers and contextual details

Check names, contact details, tax identifiers, identity document details and bank information. Examine headers, signatures and footers: these often repeat information removed from the body. The filename may also contain the client's name. Do not focus only on items highlighted by an automatic detector; abbreviations, spelling mistakes and photographs may escape detection.

Next examine context. A unique role in a small business, a public incident or an unusual combination of place and date may identify someone. Generalise or remove unnecessary details. Preserve relationships essential to the task, however: if two parties have different obligations, replacing both with “person” can change the document's meaning.

Use consistent replacements

Assign stable labels such as CLIENT_A and SUPPLIER_B. Check that one person does not receive two labels and that different people are not accidentally merged. If the work requires restoring data in the answer, keep the correspondence separately with limited access. Do not attach the map to the service alongside the masked text.

A reversible replacement should not automatically be called anonymisation. Article 4(5) of Regulation (EU) 2016/679 of 27 April 2016 defines pseudonymisation and requires additional information to be kept separately and protected. Recital 26 requires assessment of identifiability in the context of reasonably likely means. Source: GDPR, Article 4 and Recital 26.

Check the copy that will be sent

Export the text and read it in the exact form the recipient will receive. If sending a PDF, also check images and automatically recognised text. If pasting text, check what appears in the input field before sending. A preview may not represent comments, attachments or metadata in the original file. The check therefore needs to reflect the chosen format rather than only the visible page.

Complex documents may require additional technical examination. The ICO describes hidden information risks in its guide to disclosing documents securely. This British source is used here for technical guidance, rather than interpretation of Italian law. No single text search test establishes by itself that a file contains no identifying information.

A fictional example

A procurement manager wants to improve a communication about delayed delivery. She copies the letter and replaces the contact's name with CONTACT_A. She removes the telephone number, signature and personal address, then notices that the text identifies the only technical director at a small office. Because the specific role is unnecessary for rewriting, she uses “supplier contact”. She keeps the number of delayed days because it is essential to the message.

After receiving the answer, she checks that the AI has not changed facts or added accusations. She restores data only in the copy intended for the authorised recipient. The map remains separate from the AI service. The example demonstrates a cautious procedure, without claiming that the text is legally anonymous or that sending it is lawful in every situation.

Assess the response and retention too

The check does not end with upload. The answer may repeat details, invent information or reconstruct associations you tried to avoid. Verify the result before use and limit access to working copies. Define how histories, exports and maps should be managed according to actual needs and organisational rules.

The European Commission explains that pseudonymised information still attributable to someone remains personal data in its guide to the application of the GDPR. Masking is therefore one measure within the process, rather than a compliance certificate.

To see how to mask text in your browser and restore data in the response, read about Riducta.