Help center
Frequently asked questions
Detailed answers about Safe-Doc: shadow AI, pseudonymization, zero-storage architecture, GDPR compliance and limits. Updated with the product.
Because real usage already exists. In most organizations, employees use generative AI to move faster - often without a document security layer. That's Shadow AI: productive, but hard to audit.
Bans without alternatives create workarounds, lost productivity, and higher risk (people still send raw documents).
Safe-Doc doesn't replace internal policies. It enables controlled usage: pseudonymize before AI, stay in control, put the real values back on the way back when needed.
Shadow AI is the use of external AI tools outside an official framework: no dedicated DPIA, no logging, no document protection before sending.
It can be small (pasting a paragraph) or massive (contracts, audits, HR files). Safe-Doc targets the latter - where document exposure is structural.
No. Safe-Doc is not an analysis model. It's a document security layer before and after the AI you already use.
You keep your usual tools on a pseudonymized or anonymized version - Safe-Doc doesn't lock you into a proprietary chat or a workspace that stores your case files.
A summary tool processes content. Safe-Doc processes exposure risk: what goes to AI, what stays with you, and what can be re-identified later.
"Workspace + built-in chat" platforms often centralize files, history and mapping tables at the vendor. Safe-Doc inverts that: ephemeral processing, not a cloud archive of sensitive case files.
Safe-Doc follows four steps:
- Import a document (file, pasted text, or Data Room batch)
- Pseudonymize or anonymize depending on the chosen mode
- Review detected entities (uncheck, adjust)
- Export the AI-ready text, the mapping table (in pseudonymized mode) and the report
On the way back, the real values are put back on your side: by importing the mapping table with the AI answer, or automatically with the Chrome extension, the Claude skill or the MCP server.
See the visual guide →
Single document - one contract, report or memo: fast, ideal for daily use.
Data Room - a homogeneous batch (contracts + annexes + exhibits): consistent pseudonymization - the same entity gets the same alias across the batch. Essential for multi-document AI analysis.
- Anonymized - masking oriented toward irreversibility
- Pseudonymized - consistent aliases and an exportable mapping table
No. Safe-Doc is a browser-accessible SaaS. No workstation installation, no IT integration required. An account is enough to get started in minutes. The Chrome extension, to work right in the ChatGPT or Claude window, is optional.
For team deployments (Cabinet or Cabinet+ plan), users receive an email invitation. No IT department involvement required for standard plans.
90+ entity types detected - multinational coverage
- Personal data: names, emails, phones, addresses, national IDs…
- Confidential data: company names, internal references, contracts, amounts, IBAN…
- Identifying context: exact dates, fine locations, rare combinations
- Metadata: author, revision history, comments (DOCX), PDF layers
Exact scope depends on the chosen mode.
Pseudonymize if you need to reuse AI analysis with real values (mapping table, then restoration).
Anonymize if you don't need re-identification and want fewer traces (no correspondence table to secure).
Safe-Doc offers both - not a single marketing choice.
Yes, in pseudonymized mode. In the app: import the mapping table (mapping_xxx.json) and paste the AI answer, which contains the aliases; Safe-Doc puts the real values back locally.
It is automatic with the Chrome extension, right in the ChatGPT or Claude page. With the Claude skill and the MCP server, the result is written with the real names to a file on your computer.
You remain responsible for protecting the mapping table: don't share it on unsecured channels.
Yes by design - that enables "analyze then re-identify". Reversibility relies on the mapping table; if it leaks, re-identification risk rises sharply.
Irreversible anonymization aims not to leave an exploitable link - but strict GDPR "anonymity" remains hard in practice on rich documents.
Pseudonymization (GDPR Art. 4(5)) replaces identifying data with aliases while keeping a mapping table that allows re-identification. The data remains personal data under GDPR : processing stays subject to its obligations.
Anonymization aims for irreversibility: if no cross-matching can re-identify, the data falls outside GDPR scope. In practice on rich documents, that threshold is hard to reach.
Safe-Doc uses the right term : not approximate "anonymization" when re-injection is possible.
With Safe-Doc, the process takes 4 steps:
- Import your PDF into Safe-Doc (drag-and-drop or upload)
- Review detected entities and export the pseudonymized text
- Paste the pseudonymized text into ChatGPT : no sensitive data goes to OpenAI
On the way back, import the mapping table to put the real values back into ChatGPT's answer. With the Chrome extension, you attach the PDF right in ChatGPT with the Safe-Doc paperclip, and the answer comes back with the real names.
An NDA typically contains party names, amounts, validity dates, confidentiality clauses and internal references : all elements that must not transit in plain text to a third-party AI model.
Safe-Doc automatically pseudonymizes all these entities before sending. You get a "neutralized" contract you can analyze freely with ChatGPT, Claude or Gemini : then re-inject the real values into the result.
Safe-Doc is built on no-retention:
- No document content kept in the database or application logs
- temporary technical processing (upload, pipeline), then purge
- the processed file and the mapping table (if exported) are downloaded to you
Job metadata (status, counters, technical keys) may be kept for tracking - never the full document text.
Many LegalTech solutions offer a persistent project space: file versions, chat history, hosted mapping tables, long retention to re-identify later.
Convenient - but each upload increases risk surface: data lifetime at the vendor, support/backup access, DPA and sub-processor complexity.
Safe-Doc doesn't turn client files into a cloud vault. It's a technical pass-through: in → process → out → purge.
No. Safe-Doc doesn't replace ChatGPT: it doesn't store your prompts or third-party model answers. You keep exchanges in the AI tool you already use - Safe-Doc provides the upstream/downstream document layer.
You are. The mapping table is generated for local export, or kept in your browser with the Chrome extension, where you can clear it at any time. Its protection (encryption, retention, internal sharing) is your organization's duty. Safe-Doc doesn't keep it as a central re-identification archive.
Safe-Doc is designed privacy-by-design: minimization, pseudonymization (GDPR Art. 4), operation traceability, EU hosting (Hetzner, Germany), 100% self-hosted AI, DPA available.
It's a compliance aid - not a substitute for the controller's obligations (legal basis, retention, transparency, etc.).
The European AI Act imposes transparency, documentation and risk-management obligations for AI system use, on a phased timeline. Since the AI Omnibus, published in the EU Official Journal in July 2026, high-risk system obligations apply from 2 December 2027 (Annex III) and 2 August 2028 (Annex I products).
Safe-Doc helps AI Act compliance by:
- Documenting each pseudonymization operation (audit report)
- Reducing personal data exposure to third-party AI models
- Enabling workflow traceability for AI on sensitive documents
Safe-Doc is not a standalone AI Act compliance tool : consult your DPO for a full analysis.
Safe-Doc targets sensitive documents and controlled AI usage - including law firms. Pseudonymization before third-party AI reduces exposure, aligned with sector recommendations.
Lawyers and DPOs remain responsible for data choices, AI vendor selection and final review. Safe-Doc does not provide legal advice.
Data processing infrastructure: European Union (Hetzner, Germany). Scaleway is used for transactional email only - no document hosting. No third-party AI API receives your content; you then send the pseudonymized version to the AI tool of your choice from your workstation.
No. Safe-Doc doesn't use your files to train models. For external AI you use afterward, check your account terms (professional tiers with training opt-out).
Technical indicators: job id, duration, counters (pages, PII replaced, warnings) - not document bodies, no plaintext PII in application logs.
PDF, DOCX and spreadsheets (XLS, CSV); text pasted in the UI. Scanned PDFs: OCR extraction depending on scan quality.
No. Any automatic engine has false positives (mask too much) and false negatives (miss an entity). Hence human review and residual risk indicators in the report.
Safe-Doc removes common metadata (author, tool dates, Word comments, etc.). The stylistic fingerprint, meaning the way the text is written, is not processed. No guarantee of removing all forensic traces.
Only if your organization has appropriate legal basis and contractual framework. Safe-Doc is not HDS-certified or cleared for national defense secrets. When in doubt, consult your DPO or compliance officer.
Yes. Safe-Doc is AI-tool agnostic : it works with any model or interface: ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, Mistral and any other LLM accessible via web interface or API.
The principle is the same: pseudonymize with Safe-Doc, send the protected version to Copilot, then put the real values back locally. In the browser, the Chrome extension also works in Copilot.
All AI tools accessible via web interface or API:
- OpenAI: ChatGPT (all plans), OpenAI API
- Anthropic: Claude (claude.ai, Claude Cowork, Claude Code, API)
- Google: Gemini, NotebookLM
- Microsoft: Copilot, Azure OpenAI
- Mistral, Perplexity, LLaMA via third-party interfaces
Safe-Doc works upstream (pseudonymization) and downstream (restoring the real data), whatever the AI tool. There are four ways to use it: the web interface, the REST API, the MCP server and, if you like, the Chrome extension, which does the same thing directly in the ChatGPT, Claude, Gemini, Mistral, Perplexity or Copilot window. The extension is optional: you don't need to install anything to use Safe-Doc.
Both tools pseudonymize documents before AI. The main differences:
- Market: Anonym-IA mainly targets lawyers; Safe-Doc targets all professionals handling sensitive documents (M&A, consulting, HR, finance, audit)
- Architecture: Safe-Doc does not store your data or see your AI queries : unlike solutions with built-in AI chat where every query transits their servers
- PII coverage: 90+ entity types detected across 6 countries
- Formats: PDF, DOCX, XLS and CSV spreadsheets, OCR; multi-document Data Room
Yes, and it is optional. The Chrome extension (beta) pseudonymizes your message and attachments before sending, in ChatGPT, Claude, Gemini, Mistral, Perplexity and Copilot. The answer comes back with the real names, right in the page, with no copy and paste.
With the Safe-Doc skill for Claude. When you hand a file to Claude, it does not open it: the skill sends it to Safe-Doc, and only the pseudonymized version enters the conversation. The final result is restored on your computer.
Yes, with the REST API: a text in one call, documents (PDF, Word, spreadsheets) through asynchronous processing, consistent aliases across a whole batch. The API is included from the Pro plan.
Yes, with the MCP server: reminders, bank reconciliations, summaries, onboarding. The agent only works on aliases, never reads the original file, and the result with the real names is written to your computer.
Yes, with the x402 protocol: the agent sends a text, receives the price, pays for the call in USDC and gets the text back pseudonymized or anonymized. No sign-up, no API key.
Solo (€19/month · 1 seat · 1,200 pages), Pro (€49/month · 1 seat · API and integrations), Cabinet (€149/month · 3 seats · 6,400 pages), Cabinet+ (€399/month · 10 seats · 20,000 pages). Enterprise on request. Add-ons: S/M/L page packs · extra seat +€35 (Cabinet and Cabinet+ only).
Safe-Doc offers 5 plans:
- Solo : €19/month · 1 seat · 1,200 pages/month
- Pro : €49/month · 1 seat · everything in Solo + API and integrations (Claude Cowork…)
- Cabinet : €149/month · 3 seats · 6,400 pages/month (~€50/user)
- Cabinet+ : €399/month · 10 seats · 20,000 pages/month (~€40/user)
- Enterprise : on request · unlimited seats and pages · SSO · dedicated DPA
Add-ons: page packs S (500 / €15), M (2,000 / €45), L (5,000 / €99) · +€35 / extra seat (Cabinet and Cabinet+ only).
- Legal, compliance, audit teams
- Law firms and regulated professions
- Finance, HR, M&A, operations
- SMEs adopting AI without a dedicated Shadow AI function
No. Many customers start with a pilot team (legal, compliance, innovation) then extend usage rules: pseudonymize before any external AI on sensitive documents.
No. Safe-Doc significantly reduces exposure and re-identification risk - it doesn't remove all cross-matching possibilities (rare amounts, unique context, style).
The service is provided on a best-efforts basis. Human review and settings adapted to the case remain essential.
The user: mode choice, entity validation, decision to send to third-party AI. Safe-Doc provides indicators and an action report - not legal validation of the final document.
Technically yes - not recommended. Any automated output should be reviewed before external transmission, especially on high-stakes matters (litigation, M&A, disciplinary).
A question not covered?
Contact us or try Safe-Doc directly on one of your documents.