Using Token Mode Before You Send a Document to an AI Assistant
How to tokenize a contract or report with ONYRI Sanitize before you upload it to an AI assistant, and what token mode still misses.
Token mode replaces names, emails and other identifiers with a label like [NAME1] before you export a document. You can then upload that flattened file to an AI assistant. Its answer still makes sense, because every mention of the same person keeps the same label. This guide walks through the steps, and what token mode still misses.
What changes when you tokenize a document first
A raw contract or report can hand an AI assistant real names, emails, phone numbers and account details. Some assistants keep a copy of what you upload for weeks, months or longer, depending on your settings. Token mode swaps each of those values for a consistent label before you upload anything. The assistant reads [NAME1], [MAIL1] or [IBAN1] instead of the real value. Its answer still refers to the right person or number throughout the document.
How to tokenize a document, step by step
- 1Add the PDF, Word file or scan to ONYRI Sanitize. Nothing is analyzed until you confirm.
- 2Pick a detection profile, then choose Token mode instead of Black marker.
- 3Review the detected list. Untick anything you want to keep visible, and add missed terms under Also mask.
- 4Download the flattened, anonymized PDF. Word files come back as a PDF too.
- 5Upload that file to the AI assistant instead of your original, and keep the original on your side.
The assistant's answer will use the same labels you saw in the preview. Say it tells Jane Example that [NAME1] signed the agreement on [DATE1]: she opens her original file and matches the labels back herself. ONYRI Sanitize keeps no key and cannot restore the file for you. Your original is the only way back, so keep it somewhere safe.
What token mode does not catch
Detection is pattern based, so it is not exhaustive. Review the preview before you upload anything, and add missed terms under Also mask. Some categories are not detected at all, and stay visible in the tokenized file unless you mask them yourself.
- Medical terms, religion and union membership are not detected.
- Salaries, written as plain numbers with no currency or IBAN pattern, are not detected.
- Company names are detected only with a legal form attached (SAS, SARL, Ltd, Inc.…); logos are not detected.
- Faces, signatures, charts and images inside the document are not detected or masked.
- Use Also mask or a custom rule to cover a name, project or figure the detectors miss.
Can an AI assistant still read a flattened export
ONYRI Sanitize always exports a flattened file: a new PDF made of page images, with no hidden text layer under the masks. Anthropic's documentation for Claude says the system converts each PDF page into an image and extracts the text from it, then reads both together. That is one reason a flattened export still works well with an assistant built this way. The labels come through as plain, readable text, even though the page itself is an image.
| Question | Black marker | Token mode |
|---|---|---|
| Can the assistant follow who is who in the document? | No, names are gone | Yes, the same label repeats |
| Reversible with your original file? | No, redaction is permanent | Yes, if you keep the original and labels stay consistent |
| Included in | Free and Pro | Pro only |
| Best for | Sharing outside your organization | Going back and forth with an AI assistant |
What AI assistants do with what you upload
Vendors differ, and their policies change, so check the assistant's own settings before every sensitive upload. As of September 2026, Anthropic says it only uses your conversations to train Claude if you turn that setting on. Incognito chats are never used, even then. Google's Gemini Apps Privacy Hub says a subset of chats, including files you share, may be reviewed by human reviewers to improve its services. Those reviewed chats can be kept for up to three years, even after you delete your activity.
- Check whether the assistant's default setting shares your uploads for training, and turn it off if you would rather not.
- Look for a temporary or incognito mode that skips training and human review entirely.
- Remember a privacy setting does not replace a masked document. Token mode removes the value before upload, so there is nothing sensitive left to review.
- Re-read your own confidentiality or client agreements: some forbid sharing certain documents with any third-party tool, tokenized or not.
Is token mode included in every plan
Token mode is a Pro feature. The Free plan covers 3 documents a day, up to 10 pages each, in Black marker mode only. Pro removes those limits and adds token mode, technical-secret detectors, and unlimited custom rules and profiles. Pro costs €9.90 a month or €99 a year, billed excluding VAT, with a 7-day trial that asks for a card but charges nothing today.
Frequently asked questions
- Does ONYRI Sanitize keep a key to restore the original values later?
- No. Token mode keeps no key and no mapping file. Each label like [NAME1] is only meaningful next to the original document you already have. If you lose that original, the tokenized copy cannot be reversed, by ONYRI or by anyone else.
- Do I need the Pro plan to use Token mode?
- Yes. The Free plan offers Black marker mode only, for 3 documents a day up to 10 pages each. Token mode, unlimited documents and custom rules come with Pro. It costs €9.90 a month or €99 a year, with a 7-day trial that only asks for a card.
- Will the AI assistant still understand a tokenized document?
- In most cases, yes. Because the same value always gets the same label, an assistant can still follow who did what, such as [NAME1] signing an agreement with [NAME2]. It just cannot know who [NAME1] really is, which is the point.
- What should I still check before I upload a tokenized file?
- Read the preview and the detected list first, since detection is pattern based and not exhaustive. Add anything missed under Also mask, then check the AI assistant's own privacy settings. A document with fewer values still deserves the same care as any other file you upload.
Sources & references
- Article 4(5) GDPR — definition of pseudonymisation — Intersoft Consulting (GDPR full text)
- How do you use personal data in model training? — Anthropic (Claude Privacy Center)
- Gemini Apps Privacy Hub — Google
- PDF support — how Claude processes PDF pages — Anthropic (Claude Developer Platform docs)
Mask a document without uploading it
ONYRI Sanitize finds names, identifiers, bank details and secrets in a PDF, a Word file or a scan, and masks them in your browser. You check the preview, then download a flattened copy.