Privacy proxy for AI applications
Use AI.Protect data.
Anonix replaces personal data with placeholders before your request reaches the AI provider and automatically restores it in the response. Nothing changes for your software or your staff.
Your infrastructure
Public administration
Dear Ms Ingrid Bergmann, pursuant to §§ 1, 3 WoGG your application (ref. WG-2026-04817) is granted. The benefit will be paid to account DE89 3704 0044 0532 0130 00.
3personal data items detected
Hospitals & clinics
Patient Thomas Weber, born 14.03.1968. Diagnosis: type 2 diabetes. HbA1c: 8.2%. Medication: metformin 1000 mg.
5personal data items detected
Banking & insurance
Policyholder Sabine Hartmann, policy no. KFZ-2026-881234. Accident on 03.03.2026 in Munich, Leopoldstr. 42. Damage: EUR 12,400.
6personal data items detected
At the AI provider
Rewrite this official notice in simple, citizen-friendly language.
Dear [PERSON_1], pursuant to §§ 1, 3 WoGG your application (ref. [CASE_NO_1]) is granted. The benefit will be paid to account [IBAN_1].
0reach the AI
Turn these notes into a structured discharge letter to the GP.
Patient [PERSON_1], born [DATE_1]. Diagnosis: [DIAGNOSIS_1]. HbA1c: [VALUE_1]. Medication: [MEDICATION_1].
0reach the AI
Write a letter to the policyholder about settling this claim.
Policyholder [PERSON_1], policy no. [POLICY_NO_1]. Accident on [DATE_1] in [CITY_1], [ADDRESS_1]. Damage: [AMOUNT_1].
0reach the AI
For OpenAI® and all compatible APIs, Anthropic® Claude®, Google® Gemini®, Mistral®, DeepSeek®, xAI® Grok®, Groq®, Perplexity® and self-hosted models.
01The problem
Every AI promptis an uncontrolleddata transfer.
54%
of knowledge workers use unofficial AI tools at work.
49%
would keep using AI tools even if they were explicitly banned.
42%
of companies assume their employees use AI privately for work.
Sources: Bitkom Research 2025, Cornerstone OnDemand 2025
Business applications send citizen data via API, legal software checks contracts with client names, employees paste patient data into ChatGPT®. Every day. Usually without management knowing.
Bans don’t work.
Banning AI means losing twice: the tools get used anyway, just without protection. And the productivity gain is lost. The question is not whether your staff use AI, but which data leaves your organisation when they do.
According to Gartner, by 2030 40% of all organisations will experience security incidents caused by uncontrolled AI use.
02The EU data centre fallacy
A server in Frankfurtdoes not protect you.
The US CLOUD Act applies based on where a company is headquartered, not where its servers are. OpenAI®, Anthropic®, Google®, Microsoft® and AWS® are subject to it, even when data is processed in Frankfurt, Dublin or Stockholm.
With Anonix, server location no longer matters. The AI provider only receives placeholders and never the mapping back to the original data.
US group · subject to the CLOUD Act
Data centre Frankfurt · EU
- Ingrid Bergmann · housing benefit[PERSON_1] · housing benefit
- DE89 3704 0044 0532 0130 00[IBAN_1]
- Thomas Weber · type 2 diabetes[PERSON_2] · [DIAGNOSIS_1]
- TechVision GmbH · EUR 4.2 m[ORG_1] · [AMOUNT_1]
US authority
Disclosure order
1.2bn €
highest single GDPR fine, for data transfers to the US (Meta®, 2023)
5.9bn €
GDPR fines in Europe since 2018
62%
of companies in Germany transfer data to non-EU countries
Sources: DLA Piper GDPR Fines Survey 2025, Bitkom Research 2025
03The mechanism
There and back.In real time.
Anonix sits as a transparent proxy between your applications and the AI provider. Every request is pseudonymised before it leaves your EU infrastructure. Every response is restored before it reaches your staff.
- 01
Detect and replace
Names, addresses, account details, case numbers and other sensitive information are replaced with unique placeholders. Same value, same placeholder.
- 02
AI works with placeholders
The AI fully understands the context: who represents whom, what it is about, when the hearing is. It just doesn’t know who is involved.
- 03
Original data restored
Anonix puts the original data back into the response, including streamed responses. The mapping only exists in memory.
04Detection stages
Four stages. Each catches what the previous one missed.
Anonix does not rely on a single method. Rule-based pattern recognition and three independent neural methods complement each other. A typo like “Bergman” instead of “Bergmann” is still recognised.
Eight languages
- DE
- EN
- FR
- ES
- IT
- NL
- PT
- PL
- 01
Pattern recognition
Structured data with a fixed format: email, phone, IBAN, tax ID, amounts, dates, addresses.
- t.richter@kanzlei-richter.de → [EMAIL_1]
- 22.04.2026 → [DATE_1]
- 02
Language understanding
Names, places and organisations in context. In eight languages, even with typos.
- Dr. Thomas Richter → [PERSON_1]
- Ingrid Bergman → [PERSON_2]
- 03
Industry terms
Your own categories, described in your own words and recognised without training: case numbers, diagnoses, cost centres.
- Ref. 7 O 234/26 → [CASE_NO_1]
- 04
Independent review
A dedicated AI re-reads the already pseudonymised text and finds what all other stages missed.
- LG Frankfurt → [ORG_1]
05Industries
For everyone who cannotcompromise on confidentiality.
Choose your industry and see what the AI provider actually gets to see.
All industries →Rewrite official notices in plain language without disclosing citizen data.
Instruction to the AIRewrite this official notice in simple, citizen-friendly language.
3personal data items replaced
01 · Your original
Dear Ms Ingrid Bergmann, pursuant to §§ 1, 3 WoGG your application (ref. WG-2026-04817) is granted. The benefit will be paid to account DE89 3704 0044 0532 0130 00.
02 · Sent to the AI
Dear [PERSON_1], pursuant to §§ 1, 3 WoGG your application (ref. [CASE_NO_1]) is granted. The benefit will be paid to account [IBAN_1].
03 · Your result
Dear Ms Ingrid Bergmann,
good news: your housing benefit application (WG-2026-04817) has been approved.
We will transfer the money directly to your account DE89 3704 0044 0532 0130 00. You do not need to do anything else.
Speed up research and drafting. Case data stays inside the court.
Instruction to the AIDraft the operative part of a court order for these facts.
4personal data items replaced
01 · Your original
Defendant Marcus Klein, ref. 7 O 234/26. Charge: fraud under § 263 StGB. Injured party: Müller GmbH, Frankfurt.
02 · Sent to the AI
Defendant [PERSON_1], ref. [CASE_NO_1]. Charge: fraud under § 263 StGB. Injured party: [ORG_1], [CITY_1].
03 · Your result
Court order
In the criminal case against Marcus Klein, ref. 7 O 234/26, for fraud under § 263 StGB to the detriment of Müller GmbH, a date for the main hearing at the Regional Court of Frankfurt is set.
Lighten the load of case reports without exposing families in crisis.
Instruction to the AICreate a structured case overview with recommended measures under SGB VIII from these notes.
4personal data items replaced
01 · Your original
Family Yilmaz, 3 children (aged 4, 7, 12). Report by school, suspected child endangerment. Home visit on 12.02.2026.
02 · Sent to the AI
Family [PERSON_1], [NUMBER_1] children (aged [AGE_1]). Report by school, suspected child endangerment. Home visit on [DATE_1].
03 · Your result
Case overview
Family: Yilmaz · Children: 3 (aged 4, 7, 12) · Reported by: school
Suspicion: § 8a SGB VIII · Home visit: 12.02.2026
Recommendation: initiate parenting support under § 28 SGB VIII.
AI in teaching and administration without disclosing student data.
Instruction to the AITurn this course evaluation into a structured improvement report for the study committee.
4personal data items replaced
01 · Your original
Course BWL-301, Prof. Dr. Andreas Keller. Grade: 2.3. Comment: “The lecturer explains well, but Ms Schneider in the tutorial was unprepared.”
02 · Sent to the AI
Course [COURSE_1], [PERSON_1]. Grade: [VALUE_1]. Comment: “The lecturer explains well, but [PERSON_2] in the tutorial was unprepared.”
03 · Your result
Evaluation report · Course BWL-301 · Prof. Dr. Andreas Keller · Overall grade 2.3
Strengths: clear explanations in the lecture.
Action needed: involve and prepare the tutorial lead (Ms Schneider) more closely.
Discharge letters and documentation with AI. Patient confidentiality stays in-house.
Instruction to the AITurn these notes into a structured discharge letter to the GP.
5personal data items replaced
01 · Your original
Patient Thomas Weber, born 14.03.1968. Diagnosis: type 2 diabetes. HbA1c: 8.2%. Medication: metformin 1000 mg.
02 · Sent to the AI
Patient [PERSON_1], born [DATE_1]. Diagnosis: [DIAGNOSIS_1]. HbA1c: [VALUE_1]. Medication: [MEDICATION_1].
03 · Your result
Dear colleague,
we are writing regarding our patient Thomas Weber, born 14.03.1968.
Diagnosis: type 2 diabetes · Lab: HbA1c 8.2% · Medication: metformin 1000 mg
Next steps: we recommend adjusting the therapy and a follow-up visit in 3 months.
Shift handover reports in seconds. Resident data stays protected.
Instruction to the AITurn these notes into a structured handover report for the night shift.
4personal data items replaced
01 · Your original
Resident Ms Helga Schneider, room 214, increased risk of falls. Medication Marcumar adjusted to 2.5 mg.
02 · Sent to the AI
Resident [PERSON_1], room [ROOM_1], increased risk of falls. Medication [MEDICATION_1] adjusted to [DOSE_1].
03 · Your result
Night shift handover · Ms Helga Schneider (room 214)
Increased risk of falls, raise the bed rails.
Marcumar adjusted to 2.5 mg, close monitoring required.
Schedule the next INR check.
Analyse trial data with AI without exposing participants.
Instruction to the AICreate a structured adverse event report in CIOMS format from this data.
4personal data items replaced
01 · Your original
Participant ID-4821, investigator: Dr. Maria Engel. Adverse event: nausea grade 2 after Compound XR-7.
02 · Sent to the AI
Participant [ID_1], investigator: [PERSON_1]. Adverse event: [SYMPTOM_1] after [COMPOUND_1].
03 · Your result
CIOMS adverse event report
Participant: ID-4821 · Investigator: Dr. Maria Engel
Event: nausea grade 2 · Suspected substance: Compound XR-7 · Causality: possible · SAE: no
Action: symptomatic treatment, continued observation.
Contracts and briefs with AI without breaching client confidentiality.
Instruction to the AIDraft a liability clause based on these contract details.
4personal data items replaced
01 · Your original
Seller: TechVision GmbH, represented by Dr. Stefan Hofmann. Purchase price: EUR 4.2 m. Liability cap: EUR 840,000.
02 · Sent to the AI
Seller: [ORG_1], represented by [PERSON_1]. Purchase price: [AMOUNT_1]. Liability cap: [AMOUNT_2].
03 · Your result
§ 7 Liability
(1) The liability of TechVision GmbH, represented by Dr. Stefan Hofmann, is limited to EUR 840,000.
(2) The buyer’s claims become time-barred 24 months after handover.
(3) The limitation does not apply in cases of intent or gross negligence.
Objections and analyses with AI. Tax confidentiality is preserved.
Instruction to the AIDraft a letter of objection to the tax office for this client.
2personal data items replaced
01 · Your original
Client Bäckerei Kraus GmbH, tax no. 013/123/45678. Objection to the 2025 VAT assessment. Vehicle: 60% business, 40% private use.
02 · Sent to the AI
Client [ORG_1], tax no. [TAX_ID_1]. Objection to the 2025 VAT assessment. Vehicle: 60% business, 40% private use.
03 · Your result
Objection · Tax no. 013/123/45678
On behalf of our client Bäckerei Kraus GmbH, we hereby object to the 2025 VAT assessment.
Grounds: with 60% business use, a proportional input tax deduction under § 15 UStG must be recognised. We request a suspension of enforcement.
Claims and loan applications with AI. Customer data stays protected.
Instruction to the AIWrite a letter to the policyholder about settling this claim.
6personal data items replaced
01 · Your original
Policyholder Sabine Hartmann, policy no. KFZ-2026-881234. Accident on 03.03.2026 in Munich, Leopoldstr. 42. Damage: EUR 12,400.
02 · Sent to the AI
Policyholder [PERSON_1], policy no. [POLICY_NO_1]. Accident on [DATE_1] in [CITY_1], [ADDRESS_1]. Damage: [AMOUNT_1].
03 · Your result
Dear Ms Hartmann,
regarding your claim of 03.03.2026 (policy no. KFZ-2026-881234), we are pleased to inform you that settlement has been initiated.
Location: Munich, Leopoldstr. 42 · Reported damage: EUR 12,400
An assessor will contact you within 5 working days.
Job ads and HR texts with AI without disclosing employee data.
Instruction to the AITurn this internal job description into an external job advertisement.
4personal data items replaced
01 · Your original
Successor for Martin Schäfer, controlling department. Salary: EUR 78,000. Manager: Dr. Lisa Berger. Notice effective 30.06.2026.
02 · Sent to the AI
Successor for [PERSON_1], controlling department. Salary: [AMOUNT_1]. Manager: [PERSON_2]. Notice effective [DATE_1].
03 · Your result
Controller (m/f/d)
Your tasks: budget planning and forecasting, monthly and annual closing, reporting to Dr. Lisa Berger.
We offer: salary EUR 78,000, starting as soon as possible.
AI features in your own product, with privacy as a built-in component.
Instruction to the AIWrite an acknowledgement of receipt to the applicant based on this application data.
5personal data items replaced
01 · Your original
Application from Anna Fischer, anna.fischer@email.de, for Senior Controller at Autohaus Schmidt GmbH. Expected salary: EUR 85,000.
02 · Sent to the AI
Application from [PERSON_1], [EMAIL_1], for [POSITION_1] at [ORG_1]. Expected salary: [AMOUNT_1].
03 · Your result
Dear Ms Fischer,
thank you for your application as Senior Controller at Autohaus Schmidt GmbH. We have received your documents and are reviewing them carefully.
You will hear from us within 10 working days.
06Integration
Change two lines.Done.
In your existing software you only swap the API address and the API key. No rebuild, no new interface, no server of your own. Anonix works with any software that uses OpenAI®-compatible, Claude® or Gemini® APIs.
- Streaming responses in real time
- Admin interface for providers, rules and costs
- Complete audit log with CSV export
- Multi-tenant with separate databases
client = OpenAI( base_url="https://api.openai.com/v1", base_url="https://proxy.anonix.ai/v1", api_key="sk-…", api_key="anx-…",)07Law & security
Every sentence verifiable by your data protection officer.
Legally precise: technically, this is pseudonymisation. For the AI provider, the data is effectively anonymous because it never receives the mapping to the original data.
- Art. 5 GDPR
- Data minimisation. Only placeholders reach the AI provider.
- Art. 25 GDPR
- Data protection by design, before every AI call.
- Art. 28 GDPR
- The AI provider receives no personal data in plain text.
- Art. 30 GDPR
- Complete audit log of all requests.
- Art. 32 GDPR
- AES-256-GCM, Argon2id, two-factor sign-in, separate tenants.
- § 203 StGB · DORA · SGB
- Professional, banking and social secrecy are preserved.
Whitepaper
US AI and data protection.All in 26 pages.
How your organisation can use ChatGPT®, Claude® and Gemini® without giving away personal data. Prepared for management, data protection officers and IT.
- CLOUD Act and GDPR: why an EU data centre alone is not enough
- Practical examples from 12 industries, from public administration to software
- Architecture, pricing and sample calculations
- Compliance check under the GDPR, the EU AI Act and industry law
PDF, 26 pages, 1.3 MBDeutsche Version (opens in a new tab)
08Your path to Anonix
Productive within days.
- 1
Consultation
Which AI providers do you use, which data must be protected, which rules apply in your industry?
- 2
Pilot
Access to your Anonix account, configuration of the rules, trial run with one team.
- 3
Production
Rollout to all workplaces. Swap the API keys, done. Support, updates and maintenance included.
You pay for usage, not for headcount.
Billing is based on the number of requests. Unlimited users in every package. Fully managed on EU infrastructure, on-premise on request.
Let’s talk abouthow you useAI.
In a short conversation we show you how Anonix fits your organisation: which providers, which data, which rules. No obligation.