The model sees tokens.
You see everything.
Proxy AI One swaps names, amounts and IDs for tokens before a prompt leaves your company. When the answer comes back, it puts the real data back in.
Summarize the contract with Acme Corp from 12 Jan 2024. Contact: john.doe@acme.com
Summarize the contract with [ORG_001] from [DATE_001]. Contact: [EMAIL_001]
The contract with Acme Corp from 12 Jan 2024 runs for two years. Questions go to john.doe@acme.com.
Summarize the contract with [ORG_001] from [DATE_001]. Contact: [EMAIL_001]
Your teams want AI. Your legal team says no.
Every prompt is a possible leak
When people use ChatGPT, Gemini or Claude at work, customer records, financial data and internal messages leave the company. Most teams only notice later.
“We can't allow that” costs time
Legal and security block AI tools by default. The risk is real, and until now there was no good alternative. So people use the tools anyway, out of sight.
Your own model is not the answer
Private models are expensive, slow to set up and behind the frontier. Your teams need the best models, and a safe way to use them.
Four steps. Sensitive data stays home.
- 01 Inside your company
Someone sends a prompt
An employee asks an AI tool to summarize a contract, draft an email or read a report. Nothing changes for them.
- 02 Inside your company
Sensitive parts become tokens
Customer names, amounts and internal IDs are replaced with anonymous tokens before the prompt goes anywhere.
- 03 At the model
The model gets a safe prompt
The cleaned prompt goes to the model you chose. The model works without access to the detected real values.
- 04 Inside your company
Your team gets a full answer
Proxy AI One puts the original context back before the answer reaches the employee. Complete, readable, protected.
Same document. Two versions.
The model reads the version with tokens. Your team keeps the original.
This agreement is made on [DATE_001] between [ORG_001] and [ORG_002], represented by [PERSON_001] ([EMAIL_001]).
- [ORG_001]Company names
- [EMAIL_001]Email addresses
- [DATE_001]Dates
- [ID_001]Internal IDs and project codes
This agreement is made on 12 January 2024 between Acme Corp and Baltic Freight, represented by John Doe (john.doe@acme.com).
- The real names, amounts and addresses
- The map from each token to its real value
- Your rules for what counts as sensitive
- The record of what was processed, and when
What your legal team needs to say yes.
One place for all AI traffic
Policies, logs and data protection sit in one layer, not per tool, per team or per model.
Tokens instead of detected data
The model gets a prompt in which detected sensitive values are replaced with tokens.
An evidence trail
Built to support GDPR accountability and DPIAs.
Switch models, keep policies
Your rules and access controls stay the same whichever model you use. No lock-in, no rework when the market moves.
Designed around EU rules
Built around GDPR principles, with EU-hosted options for processing.
Security and legal get control. Teams get the tools. Nobody has to choose.
Made for organizations that take it seriously.
European law sets a high bar, and for good reason. Proxy AI One is designed around GDPR principles and EU data protection requirements.
We are working toward formal certifications. Documentation is available on request, and we are glad to talk through your specific requirements.
- Data minimization. Only what is needed leaves your perimeter.
- Less data at the AI provider. Detected sensitive values reach the model as tokens.
- Audit trail support. See what was processed, when and how.
- Retention you set. Configurable to match your own data governance.
- EU-first hosting options. Processing within European data boundaries.
- Your own classification rules. Fit them to your sector and your data.
Start your onboarding.
Tell us about your organization. We will set up your account and walk your team through the configuration.
Prefer to talk first? Book a call.
It sits between your systems and outside AI models. Before a prompt leaves, it finds sensitive information and replaces it with placeholders. After the model answers, it restores the original values. The model does not receive the detected values in readable form.
No. It exists to remove the blockers that make companies restrict AI in the first place. Your teams keep using the models they want.
It is model agnostic. It can route cleaned prompts to any LLM endpoint, such as OpenAI, Anthropic, Mistral or Azure OpenAI. You can add or switch providers without changing your tokenization setup.
Yes. You can define your own entity types, extend the default detection and add rules for your domain, for example project codes, product names or client identifiers that generic detectors miss.
It reduces how much sensitive data leaves your company in readable form, supports data minimization and keeps an audit trail of what was processed. Formal certifications are in progress. Ask us for the current status.
Detected sensitive values reach the provider as tokens, so a breach on their side would expose tokens instead of those values. The map from token to real value stays inside your company. The rest of the prompt is processed under the provider's terms.
It is live. Individuals can subscribe on the pricing page today. For teams and regulated organizations we run a guided onboarding with custom data handling policies.
Use AI and protect your data.
Start with the Individual plan today, or talk to us about rolling it out across your organization.