The Invisible Risk

Imagine this: your line-of-business application sends citizen data to an AI service via API to make official notices easier to understand. Your law firm software has draft contracts containing client names reviewed. Your hospital system asks the AI about treatment options and transmits the patient record in the process. Or an employee copies sensitive data straight into ChatGPT^®.

It happens. Every day. In thousands of German companies and public authorities. And in most cases, management has no idea.

Sources: Bitkom Research 2025, Cornerstone OnDemand 2025

Why Bans Don’t Work

Companies that simply ban AI lose twice. First, employees use the tools anyway—just without oversight and without any safeguards. Second, these companies give up the productivity edge that AI tools provide. Market research firm Gartner predicts that by 2030, 40 percent of all organizations worldwide will experience security incidents caused by uncontrolled AI use.

The CLOUD Act—the Problem No One Sees

OpenAI^® is based in San Francisco. So is Anthropic^®. Google^®, of course. What many decision-makers don’t know: the US CLOUD Act of 2018 gives American authorities the right to demand that US companies hand over stored data. No matter which server the data is on. No matter which country.

When your software or your employees send personal data to OpenAI^®, an OpenAI^®-compatible provider such as DeepSeek^® or Mistral^®, to Claude^® or Gemini^®—whether through an API integration or direct use—US authorities can, in theory, access that data. This is in direct conflict with the GDPR.

The Supposed Solution: An EU Data Center

Many companies believe they have solved the problem by not booking AI models directly from the provider, but running them through AWS^® Bedrock or Microsoft^® Azure^® in a European data center. The data never physically leaves the EU—problem solved?

No. The CLOUD Act is tied not to the location of the server but to where the company is headquartered. AWS^® is a subsidiary of Amazon^® (USA). Microsoft^® Azure^® is a product of Microsoft^® Corporation (USA). Both are subject to the CLOUD Act—regardless of whether the data is processed in Frankfurt, Dublin, or Stockholm.

With Anonix, the server location no longer matters: the AI provider—whether accessed directly or through an EU data center—sees only anonymized data with placeholders. Even in the event of a CLOUD Act request, there is nothing usable to hand over.

The Consequences Are Real

Sources: DLA Piper GDPR Fines Survey 2025, Bitkom Research 2025

What Really Happens When Your Software Uses AI APIs

Let’s look at a typical process. Your line-of-business application calls an AI API to make an official notice easier to understand:

Without Anonix, four categories of personal data leave your infrastructure: name, case reference, IBAN, and full address. They end up on servers in the US.

BEFORE: Without Protection

AFTER: With Anonix Privacy Proxy

The AI works just as well with the placeholders as with the real data. In the response, Anonix automatically reinserts the original data—in real time, even for streamed responses. Nothing changes in the workflow for your software or your employees.

The Solution—Anonix Privacy Proxy

Anonix is a transparent proxy that sits between your applications and the AI provider—whether OpenAI^®, DeepSeek^®, Mistral^®, Groq^®, Perplexity^®, xAI^® Grok^®, Claude^®, Gemini^®, or a self-hosted model. Every request is automatically anonymized before it leaves the EU infrastructure. Every response is automatically de-anonymized before it reaches the employee.

What Makes Anonix Unique

Supported AI Providers and Models

How the Protection Works

Anonix doesn’t detect personal data with a single method, but through four successive layers of protection. Each layer catches what the previous one missed.

Layer 1: Pattern Recognition

More than 60 patterns detect structured data such as email addresses, phone numbers, IBANs, tax IDs, monetary amounts, and dates. This works reliably and quickly because these data types follow a fixed format.

Layer 2: Anonix ConvNet NER

For unstructured data such as personal names, company names, and places, a neural network is used. It understands linguistic context and recognizes that “Munich” is a place and “Ms. Weber” is a name. In eight languages: German, English, French, Spanish, Italian, Dutch, Portuguese, and Polish.

Layer 3: Anonix Zero-Shot NER

Some industries have their own sensitive data types: diagnoses, case references, project numbers, cost centers. Anonix detects these without any programming. You describe the category in your own words, and the system learns it instantly.

Layer 4: AI Quality Check

A dedicated small language model checks the already anonymized text once more for overlooked fragments. It acts as a second opinion that judges independently of all other layers.

Before and After: An Example

The AI still understands the context perfectly. It knows that this is about a court hearing, who represents whom, and when the hearing is scheduled. It just no longer knows who exactly is involved.

Government, Public Administration & Public Sector

Ministries, state agencies, municipalities, and subordinate authorities process citizen data at the highest protection level. At the same time, Germany’s Online Access Act (OZG) mandates digitalization. Several German states have already adopted AI strategies, yet data protection officers block every direct API connection to US AI services. A political dilemma that can be solved technically.

The Challenge

• Citizen data is subject to BSI IT baseline protection and VS-NfD (restricted classified information) requirements
• Registration data, social benefit applications, building permits, and tax data must not reach third parties
• Police and security authorities process investigation data and witness protection information
• Educational institutions process student data and exam results
• State data protection laws prohibit cloud AI involving personal data

Anonix in Practice

Hospitals and Clinics

Under Art. 9 GDPR, health data enjoys the highest level of protection. Findings, diagnoses, medication plans, and treatment histories are special categories of personal data. A violation isn’t just a fine—it destroys the trust between doctor and patient. At the same time, AI-powered systems promise enormous advances in diagnostics, documentation, and treatment planning.

The Challenge

• Art. 9 GDPR: health data is a special category whose processing is prohibited in principle
• German state hospital laws impose additional requirements
• Hospital information systems (HIS) contain complete patient records
• Medical confidentiality under Section 203 of the German Criminal Code (StGB): disclosure to third parties is a criminal offense
• Enormous documentation burden: physicians spend up to 40% of their working time on documentation

Anonix in Practice

Care Facilities

Germany has more than 15,000 nursing homes and home care services. The skilled labor shortage is acute: over 35,000 positions are unfilled. Caregivers spend up to 30% of their working time on documentation instead of with residents. AI can reduce this burden—but care documentation contains health data under Art. 9 GDPR that must not reach external services.

The Challenge

• Care documentation, medication plans, and care logs contain Art. 9 data
• Resident data includes diagnoses, cognitive impairments, and behavioral issues
• Inspections by the Medical Service (MDK) require complete, audit-proof documentation
• Care home supervisory authorities and data protection authorities regularly review data processing

Anonix in Practice

Courts and Justice

The German justice system handles millions of proceedings every year. Case files, judgments, indictments, and witness statements contain the most sensitive personal data there is: victim data, witness protection information, statements by minors. Judicial confidentiality prohibits any uncontrolled disclosure. At the same time, the pressure is growing to shorten proceedings through AI-assisted research.

The Challenge

• Case files contain the names of defendants, victims, witnesses, and expert witnesses
• Juvenile criminal law: special protection for underage defendants and victims
• Witness protection programs: any disclosure of data can endanger lives
• Judicial independence requires that no external third parties gain access to case data

Anonix in Practice

Social Services and Youth Welfare Offices

Youth welfare offices and social service agencies process the most sensitive data in our society: child endangerment, taking children into care, addiction issues, domestic violence. Case reports and assessments concern people in crisis. A data breach here is not an abstract compliance violation—it can destroy families and put children at risk.

The Challenge

• Case reports contain highly sensitive data about children, parents, and family situations
• Section 65 of Book VIII of the German Social Code (SGB VIII): social data is subject to heightened confidentiality protection
• Assessments contain psychological evaluations and psychiatric diagnoses
• Caseworkers are chronically overloaded; AI could ease the documentation burden

Anonix in Practice

Law Firms

Attorney-client confidentiality is not only protected by professional rules—a breach is a criminal offense under Section 203 of the German Criminal Code (StGB). Draft contracts, briefs, due diligence reports, and client data must never be disclosed to third parties under any circumstances. At the same time, AI offers enormous potential: contract analysis, research, brief review, and legal tech are fundamentally changing the industry. Law firms that don’t use AI lose clients.

The Challenge

• Section 203 StGB: disclosing data is a criminal breach of attorney-client confidentiality
• Contracts contain company names, trade secrets, purchase prices, and terms
• M&A transactions: due diligence documents of the highest confidentiality
• In the event of a data breach, opposing counsel could claim a conflict of interest

Anonix in Practice

Tax Advisors and Auditors

Tax confidentiality under Section 203 StGB protects client data under criminal law. Tax returns, balance sheets, salary information, and tax audit documents are among a company’s most confidential business data. A violation doesn’t just threaten the advisor’s license—it can trigger liability claims that threaten their very existence.

The Challenge

• Tax confidentiality under Section 203 StGB and Section 57 of the German Tax Advisory Act (StBerG)
• Client data includes revenue, salaries, assets, and tax IDs
• Tax audit documents contain detailed business figures
• The German Public Accountant Act (WPO) and ISA standards require strict information security for auditors

Anonix in Practice

Banks and Insurance Companies

BaFin-regulated companies are subject to banking secrecy and, since January 2025, under DORA (Digital Operational Resilience Act), to the strictest IT security requirements in the EU. Account data, loan applications, claims reports, and risk assessments require the highest level of protection. At the same time, banks and insurers are investing billions in AI-powered process automation.

The Challenge

• Banking secrecy under Section 383 of the German Code of Civil Procedure (ZPO) and the banks’ general terms and conditions
• DORA: strict requirements for IT risk management and third-party oversight
• MaRisk/BAIT: outsourcing rules for IT services
• Customer KYC data, account transactions, and credit scoring results are highly sensitive

Anonix in Practice

HR Departments

Every company—regardless of industry—processes highly sensitive employee data in its HR department. Personnel files, applicant data, salary information, sick notes, and employee evaluations are personal data whose disclosure to third parties has consequences under both employment and data protection law. AI-powered recruiting, performance reviews, and workforce planning are booming—often without the data protection officer’s knowledge.

The Challenge

• Section 26 of the German Federal Data Protection Act (BDSG): strict purpose limitation when processing employee data
• Applicant data: compliance with the German General Equal Treatment Act (AGG) and discrimination risks in AI-assisted pre-screening
• Works agreements often explicitly regulate the use of AI tools
• Salary data, sick notes, and formal warnings are particularly sensitive

Anonix in Practice

Educational Institutions

Universities, colleges, schools, and research institutions are under double pressure: on the one hand, they are expected to be AI pioneers and provide students with modern tools. On the other hand, they process student data, exam results, and personal research data covered by state data protection laws. Several state data protection commissioners have already published recommendations against using US cloud AI services.

The Challenge

• State data protection laws and higher education acts govern the handling of student data
• Exam results, student ID numbers, and reasons for withdrawal are specially protected
• Research data involving personal information is subject to ethics committees
• Lecturers are already using AI extensively—often without institutional approval

Anonix in Practice

Pharma and Life Sciences

The pharmaceutical and life sciences industries process clinical trial data, patient information, and proprietary research results. GxP regulations, FDA 21 CFR Part 11, and the EU Clinical Trials Regulation set the highest standards for data integrity and confidentiality. AI speeds up drug development by years—but trial data must not fall into the wrong hands.

The Challenge

• Clinical trial data contains patient names, diagnoses, and treatment histories
• Proprietary compound data and research results are trade secrets
• GxP compliance requires complete traceability of all data processing
• Adverse event reports contain personal reports of side effects

Anonix in Practice

Software Companies and IT Service Providers

Software companies and IT service providers integrate LLM APIs into their own products—for customers in every industry. They are responsible for ensuring that their end customers’ data remains protected. A data protection violation affects not only their own company but every customer whose data flows through the API. Here, Anonix becomes an infrastructure component built into the product itself.

The Challenge

• Data processing under Art. 28 GDPR: the software provider is liable for its customers’ data
• Customer data from regulated industries (healthcare, finance, legal) flows through the API
• SOC 2 and ISO 27001 certifications require demonstrable data protection measures
• Customers increasingly demand proof of GDPR-compliant AI use in tenders

Anonix in Practice

Why Existing Solutions Fall Short

There are certainly products that can detect personal data. But none of them solves the actual problem: using AI APIs securely without data leaving the organization.

Where Is the Difference?

OpenRouter^® is an AI gateway with over 400 models and has recently added redaction of sensitive data. But the check only happens on the servers of the US company, so the data leaves the organization in plain text. The redaction is permanent: every name becomes [PERSON_NAME], the AI can no longer tell people apart, and the response contains no original data. OpenRouter^® detects names and addresses only as a beta feature. If this check times out, the request is forwarded unredacted.

Google^® Cloud DLP and AWS^® Comprehend do detect personal data, but they are themselves cloud services of US companies. Taking a detour through AWS^® Bedrock or Microsoft^® Azure^® in an EU data center doesn’t change anything either: the CLOUD Act applies regardless of server location, because the parent company is based in the US.

Microsoft^® Presidio^® is open source and can be run locally. But it has no proxy function, no de-anonymization, and no zero-shot detection. It is a toolkit, not a finished product.

The Architecture at a Glance

Anonix consists of five components that are operated as a fully managed SaaS service in the EU. Not a single byte of personal data leaves the European infrastructure.

The Five Building Blocks

Security at Every Level

Investment and Cost-Effectiveness

Data protection doesn’t have to be a cost driver. Anonix costs a fraction of what a single data protection violation would cost. And at the same time, it enables the productive use of AI that your employees are already asking for.

What a Violation Costs

What Anonix Costs

Anonix charges by usage—not by users. A base fee plus a price per anonymization request. Unlimited users in every plan.

Base fee: EUR 99 – 999 / month depending on plan and company size. All prices plus VAT. SLM premium module (AI quality check with dedicated GPU inference): +1.4 cents per request.

Basis: ~10 AI requests per employee per working day × 22 working days = approx. 220 requests per employee per month. The more your company uses AI, the cheaper each individual request becomes.

All prices are net prices plus VAT.

Compliance at the Push of a Button

Data protection is not a project with a beginning and an end. It is an ongoing process. Anonix automates the technical measures that regulators require.

GDPR Compliance

• Art. 5(1)(c) (data minimization)—Only placeholders reach the AI provider. Personal data is reduced to the technically necessary minimum.
• Art. 25 (data protection by design)—Anonymization takes place before the API call. Without the proxy configuration, unprotected use is not possible.
• Art. 32 (security of processing)—AES-256 encryption, Argon2id hashing, two-factor authentication, and isolated tenant databases.
• Art. 28 (processors)—The AI provider sees no personal data. The requirements placed on the processor are significantly reduced.
• Art. 30 (records of processing activities)—Complete audit log with full logging of all proxy requests.

CLOUD Act and Third-Country Transfers

EU AI Act (Regulation 2024/1689)

Anonix is a data minimization tool and does not fall into any high-risk category of the EU AI Act. It makes no automated decisions about individuals, creates no profiles, and does not affect anyone’s rights. Rather, it reduces the risks posed by the actual high-risk applications (the external AI models).

Industry-Specific Compliance

Voices from the Field

The challenge is the same across industries: how can organizations capture the productivity gains of AI without losing control over sensitive data? Three perspectives.

Your Path to Anonix

Anonix is designed to take you from first contact to productive use in just a few days. No lengthy implementation projects, no rebuilding of your existing systems.

Deployment

Anonix is provided as a fully managed SaaS service. Hosting, updates, maintenance, and support are included in the package. You can focus on your core business while we run the infrastructure—exclusively on EU servers, of course.

Legal Notice

Copyright, Trademark, and License Rights

All copyright, trademark, and license rights to Anonix Privacy Proxy and this document are held by:

Trademarks

Anonix, Anonix Privacy Proxy, Anonix ConvNet NER, and Anonix Zero-Shot NER are trademarks of Convecto GmbH. OpenAI^®, GPT^®, Anthropic^®, Claude^®, Google^®, Gemini^®, Microsoft^®, DeepSeek^®, Mistral^®, xAI^®, Grok^®, Groq^®, Perplexity^®, Together AI^®, OpenRouter^®, Ollama^®, LM Studio^®, Cursor^®, GitHub^® Copilot^®, LangChain^®, LlamaIndex^®, and AWS^® are trademarks of their respective owners.