Reading time: 7 min | Published: June 2026 | Category: Privacy, AI Security, GDPR
Every day, millions of professionals send sensitive documents to AI models. Contracts. Medical records. Financial statements. HR files. They paste in the content, hit send, and get a useful summary back in seconds.
What most of them don't think about is what happens in between.
Before the AI responds, it reads everything. Your client's full name. Their IBAN. Their date of birth. Their home address. The diagnosis on page 3. The salary figures in the appendix. None of it is filtered. None of it is necessary.
The AI doesn't need any of it to do its job.
The task and the identity are two different things
When you ask an AI to summarize a contract, it needs to understand the clauses, the obligations, the deadlines, and the risks. It does not need to know that the party signing is named Jean-Pierre Moreau, born March 15, 1972, residing at 14 rue des Acacias, Lyon.
When you ask an AI to flag red flags in a financial statement, it needs to analyze the numbers, the ratios, the anomalies. It does not need to know the account holder's IBAN, their bank's BIC code, or their tax identification number.
The content that makes the AI useful and the personal data that makes it a privacy risk are almost never the same thing.
This is the insight ArcanAI was built on.
What actually happens when you send a document to an AI
Most AI providers — OpenAI, Anthropic, Google — have privacy policies that describe how they handle your data. Some offer enterprise agreements. Some allow you to opt out of training. Some promise they don't store your conversations.
All of that is worth reading. None of it changes the fundamental fact: the model sees what you send it. The raw text, including every name, number, and identifier, passes through their infrastructure before any policy applies.
For a personal use case, this is a minor concern. For a professional context — a law firm, a hospital, a financial institution, an HR team — it raises real questions. GDPR questions. Professional secrecy questions. Client confidentiality questions.
The standard answer has been: don't send sensitive documents to AI. Or: use an enterprise agreement and hope for the best.
There is a better answer.
Strip the identity before the document leaves your device
The approach ArcanAI takes is simple in principle: detect every piece of personally identifiable information in your document, replace it with a neutral token, and only then send the anonymized version to the AI.
Your document goes from this:
"The patient, Marie Dupont, born 04/12/1985, residing at 8 allée des Pins, Bordeaux, was referred by Dr. Carrera on 15 March 2024..."
To this:
"The patient, [NAME_1], born [DOB_1], residing at [ADDRESS_1], was referred by [DOCTOR_1] on [DATE_1]..."
The AI receives the anonymized version. It analyzes the structure, the meaning, the risks — everything it needs to be useful. It returns its response using the same tokens.
Then, locally in your browser, the tokens are replaced back with the real values. You see the full result. The AI never saw the names.
The token map — the dictionary that links [NAME_1] to "Marie Dupont" — is stored only in your browser's memory. It is never transmitted. It disappears when your session ends.
Why client-side matters
The critical word in the previous section is locally. The anonymization doesn't happen on a server ArcanAI controls. It happens in your browser, on your device, before anything is sent anywhere.
This matters because it means the sensitive data never has to leave your device at all. ArcanAI can't read it. The AI provider can't read it. If servers were compromised, there would be nothing to find.
This is what "Upload your documents. Not your identity." actually means in practice.
The minimum necessary principle
Beyond anonymization, ArcanAI applies a second layer of control: before your document is processed, a local model analyzes the task you're asking for and determines the minimum information the AI actually needs to complete it.
Summarizing a lease agreement? The AI needs the terms, the duration, the obligations. It probably doesn't need the property address or the tenant's contact details.
Reviewing a medical report for treatment recommendations? The AI needs the clinical findings. It doesn't need the patient's full name, date of birth, or social security number.
This principle — only share what is strictly necessary — is already embedded in GDPR. It's called data minimization. ArcanAI builds it into the product architecture rather than leaving it as a policy aspiration.
Who this is for
ArcanAI was built for anyone who works with sensitive documents and uses AI to process them. In practice, that means:
Lawyers analyzing contracts, due diligence files, or litigation documents. Client confidentiality is not optional. The AI doesn't need to know who your client is.
Healthcare professionals reviewing patient files, lab results, or referral letters. Medical data is among the most protected in the world. The AI doesn't need the patient's identity to flag a clinical risk.
Finance and accounting teams processing statements, reports, or payroll files. Financial data is sensitive. The AI doesn't need the account numbers to identify an anomaly.
HR teams handling employee records, performance reviews, or recruitment files. Employee data is personal. The AI doesn't need the name to evaluate a profile.
Anyone who has ever hesitated before pasting a document into ChatGPT, Claude, or Gemini — because something felt wrong about sending all of it.
The tradeoff that isn't
Some people assume that anonymizing a document before analysis makes the results worse. If the AI doesn't know the names, does it produce a worse summary?
In almost every case: no.
The quality of an AI analysis depends on the structure and content of the document, not on the identity of the parties involved. A contract is a contract whether the signing party is called [NAME_1] or Jean-Pierre Moreau. A medical report is a medical report whether the patient is [NAME_1] or Marie Dupont.
The AI does its job. You get the result. The identity stays with you.
That is the whole point.
Before AI can think — ArcanAI decides what it is allowed to know. Try it free at arcanai.co.