Prepare a document for AI without leaving your data in it
To prepare a document for AI is to make it useful to a machine while keeping your data at home. A contract, a payroll spreadsheet, a client file, a letter: all of them contain names, identifiers and amounts that belong on your computer, and only there. And these are precisely the documents on which ChatGPT, Claude or Gemini are most helpful.
We wrote this page as a method that works whatever the tool: five steps, a list of what to remove, a list of what you can keep, and how to restore the original values once you have the answer. Biffer automates most of this work, offline, and the method applies to all your documents.
Why secure your documents before handing them to an AI
Key points
- Before sending a document to an online AI, you have full control over what it contains: that is the moment to act.
- Preparing data for AI means removing what identifies people and what is confidential, while keeping a rich, readable document.
- Stable aliases (PERSONNE_001, SIRET_002) keep the text consistent and let you restore the real values afterwards.
- Biffer does this work offline, with a preview you review before anything is sent.
The first reason is simple: once a file is dropped into an online AI, it is stored, processed and sometimes retained under the provider's terms. The settings on offer frame that use; preparing the document, on the other hand, is entirely up to you and happens before sending. It is the stage where you decide everything.
The second reason is legal. As soon as a document contains a person's name, email address, social security number or IBAN, it contains personal data within the meaning of Article 4 of the GDPR, and passing it to a provider engages your responsibility. Removing personal data before using AI, or pseudonymizing it, reduces that risk at the source. We detail this framework on the page personal data and AI.
The third reason concerns confidential information that is not personal data: a negotiated price, an exit clause, a project's name, the margin on a quote. You are free to remove it, and that is often the safest choice. Protecting your data before using an AI means treating both families with the same care.
The method in five steps
We apply this method to every document, whether it goes to ChatGPT, Claude, Gemini or to an outside firm. It comes down to five actions, in this order.
1. Define what you will ask the AI
Summarizing a contract, rephrasing a letter, spotting inconsistencies in a spreadsheet, translating meeting minutes: the request determines what the document needs to keep. For a contract summary, dates and amounts matter; for a rephrasing, they are superfluous. Write the question before preparing the file, not after.
2. Find what identifies and what is confidential
Go through the document looking for two things: what points to a person or a company (names, contact details, official identifiers, professional identifiers) and what falls under trade secrets. Biffer does this automatically, checking the check digits of French company numbers (SIREN), French social security numbers and IBANs, then shows a highlighted preview.
3. Replace with aliases rather than delete
A stable alias (PERSONNE_001 wherever that name appears) keeps the document's logic: the AI always knows who does what. That is what makes a document cleaned before AI still fully usable. For each data type, choose to pseudonymize, mask or keep.
4. Review the preview and correct
Untick false positives and add whatever the analysis missed by selecting the text: all occurrences are handled at once. Also check the file's metadata (author, company, comments), which Biffer clears on export. This review takes a few minutes, and it is what makes the result good.
5. Send, then restore
Send the prepared document to the AI and ask your question. When the answer comes back, replace each alias with its original value using the mapping table, encrypted or in plain text as you chose. You get a text you can use directly, and the real data stayed on your computer from start to finish.
Each screen, from analysis to the license key, is described on the page how Biffer works.
What to remove, what you can keep
The question comes up in every training session: should everything be anonymized? Our answer: remove what points to someone, keep what sheds light. A document that keeps its structure and useful figures remains fully usable. The table below sums up the trade-offs we make most often, category by category.
| Category | Examples | Our advice |
|---|---|---|
| People | Last name, first name, title followed by a name | Always pseudonymize: PERSONNE_001, PERSONNE_002 |
| Contact details | Email, phone, postal address | Pseudonymize or mask: the AI works just as well without them |
| Official identifiers | French company number (SIREN, SIRET), social security number, IBAN, license plate | Pseudonymize, or mask if the document is meant to be shared widely |
| Professional identifiers | Court case number (RG), RPPS, CRPCEN, training provider number (NDA), claim number, employee ID | Pseudonymize with the matching profession profile |
| Dates | Birth, signature, deadlines | Keep if the question is about a schedule; pseudonymize a date of birth |
| Amounts | Prices, salaries, balances, margins | Keep for a numerical analysis; mask for a rephrasing |
| Professional vocabulary | Standard clauses, technical terms, structure | Keep: this is what lets the AI understand the document |
Amounts deserve separate thought. They are not personal data in themselves, but a salary tied to an employee ID, or a balance tied to an IBAN, becomes personal data by association. Our rule: if the person is pseudonymized and the amount is needed for the question, keep it; if the amount is a trade secret and the question does not depend on it, mask it. The EDPB, in its Opinion 28/2024 on AI models, recalls that data minimization also applies to AI uses.
What you can keep without worry: the document's structure, section headings, legal or technical vocabulary, quantities, durations, units. That is precisely what the AI needs to produce a faithful summary or a sound analysis. Preparing your data for AI means removing what points to someone and keeping everything that sheds light.
Restoring afterwards: the mapping table
Restoration is the step that turns a good practice into an everyday working tool. With it, a summary that talks about PERSONNE_001 and SIRET_002 becomes a complete document again in one action, with each name automatically back in place.
If you wish, Biffer saves a mapping table that links each alias to its original value. Two forms are possible, plus the option of having none:
- Encrypted, with a password you choose: the table is protected by Argon2id and AES-256-GCM. You alone hold the password that opens it, and that applies to us too. This is the form we recommend as soon as the document leaves your office.
- Plain text, readable without a password: handy for quick processing on a computer only you use, provided you delete it afterwards.
- None: you work without a table, and the processing is permanent. This is the right choice when the document must be anonymized for good, for example before it is shared widely.
A legal caveat: as long as a mapping table exists, the prepared document remains pseudonymized data under the GDPR, and therefore subject to the regulation. Without a table and with no reasonable possibility of re-identification, you approach anonymization. Biffer gives you the means to minimize; review and the decision are yours, and that is what your compliance rests on. The complete process is described on the page anonymize a document.
ChatGPT, Claude, Gemini: the same steps, three dedicated pages
The method is the same from one tool to the next, and each assistant has its own settings, common use cases and documentation. We wrote one page per tool, with concrete examples and the Biffer workflow applied to each.
Preparing documents for ChatGPT
Summarizing a contract, rephrasing a letter, analyzing an Excel spreadsheet: the three most frequent uses, and what OpenAI's settings let you control. Read anonymize before ChatGPT.
Preparing a document for Claude
Reviewing a contract clause by clause, synthesizing a long file, drafting a reply to a complaint, with Anthropic's documentation as a reference. Read anonymize before Claude.
Preparing a document for Gemini
Sorting job applications, translating a business letter, working with a CSV export, with Google's documentation for account settings. Read anonymize before Gemini.
Biffer is developed by Romain Rissoan, a trainer and consultant based in Lyon, France, and founder of the training company Optédif formation. The software runs on Mac and PC (Windows), and it is free for 5 files per month, with every feature included, after a simple email confirmation in the software: you can try the method on your own files from the download Biffer page.
Frequently asked questions
Should I anonymize a document even for a trivial question?
Yes, because the document always contains more than the question. Asking an AI to correct the spelling of a letter seems simple; yet the letter contains the recipient's address, their case number, sometimes their IBAN, and the whole file goes out. The habit we teach: prepare the document before every upload, whatever the request, letting Biffer do the detection and reviewing the preview. Once the habit is in place, it adds a minute to the job and gives you most of the protection.
Is a document prepared with Biffer anonymous under the GDPR?
It depends on the choice you make about the table, and we prefer to be precise on this point. If you keep a mapping table, the data is pseudonymized: re-identification remains possible for whoever holds the table, and the GDPR continues to apply. If you work without a table and people remain beyond the reach of re-identification by reasonable means, the document approaches anonymization. In both cases, having removed the data before sending greatly reduces exposure. Biffer gives you the means to decide; the legal classification and the final review are yours.
What about a scanned PDF or a photo of a document?
Drop them into Biffer like any other file. A scanned PDF or a photo is made of page images with no text layer; the software runs built-in character recognition on it, on your computer with no connection, then the usual detection. In an image, the identified areas are redacted by blacking out pixels: the redaction is permanent, which guarantees that masked data stays masked in the shared file; for these files, processing is done without a mapping table. The preview lets you check the recognition result and manually add a name or number that was misread. The other formats, Word, Excel, CSV and text PDFs, keep aliases and reversibility.
Your business documents anonymized, ready to be published or handed to an AI tool.
Biffer runs on your computer without sending a single file. Free for five files a month, unlimited for €10 excl. VAT per computer.
100% offline · Free: 5 files a month · No file sent