Product 01
Your staff already use AI — often with texts they should not really let out of the building. VERLON privacy sits in between: names, customers, prices, contact details and bank details are replaced with placeholders before sending and restored in the response. The user hardly notices a thing.
What needs protecting is checked while you type. Several stages work together — fixed rules, your own master data and a language model that also recognises names without a form of address.
Every match is given a placeholder. The same customer keeps the same placeholder throughout the conversation, so the AI can follow the context.
Before sending, you see both versions side by side: your original and exactly what goes out. If something has been missed, you mark it and it is replaced as well.
The response comes back with placeholders and is turned back into plain text on your side. You read a finished text, not a puzzle.
A general model does not know your customers. You enter customer names, project numbers, machine designations or case references once — after that they are reliably detected, even when they look like an ordinary word.
You decide what is replaced, what is blocked and what is deliberately let through — differently for accounting than for sales. Amounts, for instance, can pass through when the AI is meant to calculate with them.
An English customer enquiry is answered in English, a German one in German — regardless of the language your people use for the interface.
Quotations, contracts, spreadsheets and emails can be attached. They are checked in exactly the same way as a typed sentence before anything leaves the company.
Every replacement is recorded and chained continuously, so it cannot be altered afterwards without it being noticed. In an audit, that is the difference between a claim and evidence.
For each request you see which model was used and what it cost. Limits per employee prevent surprises at the end of the month.
The complete system runs on your own hardware. Nothing of ours sits in between. For organisations bound by professional secrecy, often the only viable option.
A machine for you alone, set up and maintained by us. Separate from all other customers — no shared operation.
The quickest start, with no technology of your own. Separated by client, with the same protection rules as in the other models.
On this subject we stick to precise wording, because imprecise wording can become expensive.
Pseudonymisation is a protective measure under Art. 32 GDPR and Art. 8 of the Swiss Data Protection Act. It is not anonymisation: the personal reference does not disappear, because the mapping still exists on your side — which is precisely what makes it possible to restore the response.
The accurate statement is therefore: sensitive original data does not leave your environment, and the AI provider receives only a pseudonymised version. Anyone who promises you more is promising too much.
Nor is any software “data protection compliant”. Compliance always applies to a processing activity, never to a tool. We supply the measure and the evidence for it — responsibility for the processing remains with you, and we tell you what you need for it.
No Docker, no account, no administrator. A single command fetches the program, its own runtime and the detection model into one folder and creates an icon. From then on the protection layer runs entirely on your machine: detection, mapping table and keys stay there. We see nothing of your texts — not even the pseudonymised ones.
Applications → Utilities → Terminal. Paste the line and press Return.
curl -fsSL https://verlon.ch/install.sh | bashAfterwards VERLON is in the Applications folder and starts with a double-click — without a security prompt.
Start menu → type “PowerShell” → open. Paste the line and press Enter.
irm https://verlon.ch/install.ps1 | iexAfterwards VERLON is on the desktop and in the Start menu.
curl -fsSL https://verlon.ch/install.sh | bashEntry in the applications menu, controlled via ~/.local/share/verlon/bin/verlon.
Program, Node and Python runtime, detection model. Everything in one folder, nothing system-wide. You should have around 3 GB of free space and 8 GB of memory.
You start with a small starting credit. Top-ups are made in the console with one click (10, 25, 50 or 100 CHF); billing is per request in centimes, and you see every amount. If you have your own Anthropic or OpenAI key, you enter it and need no credit.
Detection needs no connection. Only the actual question goes — in pseudonymised form — to the AI provider.
Prefer a file to click? macOS · Windows — contains the same setup as a double-click file, with instructions. On the Mac, a downloaded file has to be allowed once under System Settings → Privacy & Security; the Terminal line above does not need this.
Bring a text you would never give to an AI as it stands. In the demo we show you what of it would go out — and what would not.