New: AI Workspace – Secure AI for Organizations With Sensitive Data. Learn more.
The Use of AI in Healthcare: What Matters Most
What matters, therefore, is how patient data reaches an AI model—and whether this can be proven later.
In short: For hospitals, the risk arises when patient data is sent to external services without a protection tier, without logging, and without a confidentiality obligation for those involved.
Myra AI Workspace addresses this as a privacy-first AI platform: employees see how their input is protected before sending each message, and administrators define which data may reach which model.
Learn more about Myra AI WorkspacePricing: from €13 per user per month, excl. VAT
Features and monthly usage allowance vary by plan; Agents and Workflows are available as premium features.
Patient Data in Uncontrolled AI Tools
Physicians, nursing staff, medical controlling, and administration want to use AI to reduce documentation workload. If no approved solution is available, many turn to freely available services. This creates a blind spot for data protection, information security, and compliance.
The central question is therefore: How can hospitals use AI productively without handing patient data to parties that do not meet these requirements?
Patient Data in Prompts
In everyday hospital work, prompts and uploaded files quickly contain names, dates of birth, health insurance numbers, diagnoses, or findings. When a discharge letter is pasted into a freely available chatbot, health data goes to a third party—usually without a data processing agreement and without a confidentiality obligation.
The Myra AI Adoption Report 2026 shows that this is not a theoretical risk: across industries, 25% of surveyed AI users have already processed personal data with AI, and 15% have processed explicitly confidential data or data not approved for AI use.
Lack of Evidence
Without central logging, hospitals can hardly prove later which tools were used, which data went to which model, and which safeguards were in place—whether to data protection authorities or in information security audits.
Clinical Responsibility
Generative AI can produce statements that sound plausible but are wrong. Physician review remains essential for diagnosis, therapy, and medication.
Strict Legal Requirements
Several sets of rules apply to patient data at the same time. Rather than regulating AI as a technology, they govern the processing of health data, the disclosure of professional secrets, and the institution’s information security:
Article 9 GDPR: Health data is a special category of personal data. It may be processed only under strict conditions.
Section 203 StGB (German Criminal Code): Professionals bound by confidentiality may disclose patient secrets to external service providers only to the extent their involvement requires it. They must ensure that these providers are bound to confidentiality.
Section 393 SGB V (German Social Code, Book V): Healthcare providers, statutory health and long-term care insurance funds, and their processors may process health data via cloud services only under certain conditions. These include: Processing takes place in Germany, in the EU, or in an equivalent state.; the processing entity has a branch in Germany; a current C5 attestation is in place. Since July 1, 2025, it must generally be a Type 2 attestation.
Section 391 SGB V and NIS2: Under Section 391 SGB V, hospitals must implement IT security in line with the state of the art, regardless of their size. In addition, Germany’s NIS2 Implementation Act has been in effect since December 6, 2025. Hospitals at or above the threshold of an “important entity” must meet the risk management obligations under Section 30 BSIG. This threshold is generally 50 employees or 10 million Euro in annual turnover and balance sheet total. The obligations cover all IT systems used for patient care, including AI systems.
Depending on how they are used, AI systems themselves are subject to additional requirements, for example under the EU AI Act or medical device regulations.
This overview does not replace a legal assessment of individual cases.
Myra AI Workspace: Data Protection You Can See
Myra AI Workspace is a central AI work environment for organizations with sensitive data. It combines a familiar chat and project interface with administrative controls: guardrails, PII masking, model governance, logging, and budgets.
Two Protection Tiers for Different Data Classes
Tier 1—Open-weight models hosted in the EU for highly sensitive data: Myra hosts open-weight models on its own infrastructure in the EU. In projects with the “Local only” access tier, only these models are available; cloud models, web search, and external tools are disabled there. For organizations with air-gap or data residency requirements, an on-premises version is also available.
Tier 2—PII masking for frontier models: With external models, the workspace replaces personal data with placeholders before sending it and restores the original values only in the response. In addition to names, addresses, and contact details, it detects German identifiers via checksum, including the health insurance number on the electronic health card (eGK). In projects with the “PII protection required” tier, masking is mandatory.
Important for hospitals: Masking replaces identifiers, not medical content. Diagnoses, findings, and case histories remain in the text. Patient-related content therefore belongs in Tier 1.
Next to the model picker, the privacy pill shows how the next message will be processed before it is sent—color-coded like a traffic light:
🟢 Green—local model: A model hosted by Myra is selected for the message.
🟡 Amber—masked: An external model is selected; personal data is masked before it is sent.
🔴 Red—unmasked: An external model without PII protection is selected. A notice warns before sending and, where available, offers to switch to a protected route.
Administrators can enforce masking for the entire organization; the privacy pill can then no longer be switched off. Before sending, a masking preview shows which details will be replaced.
Additional Safeguards
EU region routing: The workspace blocks requests to model providers outside the EU before they are sent and records the jurisdiction in which the model ran for each request.
Request logs with optional content logging and ghost mode without storage or logging.
Tamper-evident audit log for admin changes and security events (cryptographic hash chain).
Configurable retention periods and self-service data subject rights under Articles 15, 17, and 18 GDPR.
Protection against prompt injection and jailbreaks as well as inspection of outbound tool and search calls for identifiers.
Single sign-on (OIDC, SAML), SCIM provisioning, roles, and groups.
SIEM integration (including Splunk, Elasticsearch, OpenSearch, Syslog).
Labeling of AI-generated content in line with the EU AI Act.
Full exit export in open formats (JSON, CSV).
Which AI Use Is Acceptable in Hospitals?
The following traffic light indicator classifies typical AI usage scenarios in hospitals. It does not replace a legal assessment, but it shows where the data class determines the permissible processing path.
Public or private AI accounts used with discharge letters, findings, or patient names
🔴
Health data leaves the clinic without a data processing agreement, without a confidentiality obligation under Section 203 of the German Criminal Code (StGB), and without a review under Section 393 of Book V of the German Social Code (SGB V).
Freely available AI tools without approval, policy, role-based access, or logging—even for content not related to patients
🔴
Shadow IT eludes the control of information security and data protection, as well as the documentation requirements under Section 391 of SGB V and the BSIG.
Enterprise AI from a cloud provider with a contract, used for patient-related content
🟡
This is only viable if it can be demonstrated that the requirements under Section 393 of SGB V regarding data processing on behalf of a client and the confidentiality obligation for the entire processing chain have been met.
Central AI workspace with mandatory PII masking for non-patient content, such as quality management, administration, or research
🟢
Suitable, provided that internal policies and training clearly define the permitted data classes.
Central AI workspace, external models with PII masking for patient-related content
🟡
Anonymization removes identifiers, not the medical context. Use only after conducting a data protection impact assessment.
Central AI workspace with “Local only” projects on open-weight models in the EU, EU region routing enabled, audit log, and retention periods—or on-premises
🟢
A strong approach to patient-related documentation, provided that data processing on behalf of a client, a confidentiality obligation, and a data protection impact assessment are in place.
AI outputs feed into diagnosis, therapy, or medication without physician review
🔴
Clinical decisions require expert review; AI intended for medical use is also subject to medical device and AI laws.
Public or private AI accounts used with discharge letters, findings, or patient names
🔴
Freely available AI tools without approval, policy, role-based access, or logging—even for content not related to patients
🔴
Enterprise AI from a cloud provider with a contract, used for patient-related content
🟡
Central AI workspace with mandatory PII masking for non-patient content, such as quality management, administration, or research
🟢
Central AI workspace, external models with PII masking for patient-related content
🟡
Central AI workspace with “Local only” projects on open-weight models in the EU, EU region routing enabled, audit log, and retention periods—or on-premises
🟢
AI outputs feed into diagnosis, therapy, or medication without physician review
🔴
Bottom line: Hospitals should neither ban AI across the board nor allow it without oversight. The viable path is a central work environment in which the data class determines the protection tier and in which all employees see the protection status before sending.
A Fact-Based Comparison of AI Options for Hospitals
This table compares three categories of healthcare AI solutions based on publicly available information gathered through web research (as of October 2026).
Automatic masking of personal data
❌ Not part of the enterprise services; requires an upstream layer
❌ Only via additional open-source components for PII detection
✅ Reversible tokenization; detection of German identifiers including the health insurance number
BSI C5 Type 2 attestation
🟡 Not consistent available
❌ No external provider; audits of in-house operations must be commissioned separately
✅ BSI C5 Type 2 attested Myra infrastructure
Technical block for model providers outside the EU
🟡 EU data residency assured by contract
✅ Data stays within your own network
✅ EU-Region-Routing blockiert Anfragen an Modellanbieter außerhalb der EU vor dem Versand
Auditability
✅ Admin and compliance interfaces
🟡 Conversations in your own database; audit-proof storage must be built in-house
✅ Request logs and tamper-evident audit log for admin changes and security events
Model choice
🟡 The provider’s own proprietary models
🟡 Self-hosted open-weight models
✅ +20 provider integrations plus open-weight models hosted by Myra on its own infrastructure in the EU; “Local only” project tier; on-premises option
Cost control per team and user
❌ Chat products billed per seat, no granular budgets per team or user
❌ No usage-based token costs; fixed costs for hardware and operations can only be allocated by calculation, not by actual usage
✅ Enforceable budgets, applied before the model call; cost allocation by tenant, model, and user
Operational effort for hospital IT
✅ Fully managed service
❌ Hardware, updates, patches, monitoring, and on-call duty in-house
✅ Managed service; alternatively on-premises with support from Myra
Pricing
🟡 Depends on provider, plan, and usage
🟡 Capital and operating costs
✅ From €13 per user per month, excl. VAT
Automatic masking of personal data
❌ Not part of the enterprise services; requires an upstream layer
BSI C5 Type 2 attestation
🟡 Not consistent available
Technical block for model providers outside the EU
🟡 EU data residency assured by contract
Auditability
✅ Admin and compliance interfaces
Model choice
🟡 The provider’s own proprietary models
Cost control per team and user
❌ Chat products billed per seat, no granular budgets per team or user
Operational effort for hospital IT
✅ Fully managed service
Pricing
🟡 Depends on provider, plan, and usage
Conclusion: When hospitals evaluate AI solutions for healthcare, model quality is only one criterion. Just as important is whether the processing path can be controlled by data class, enforced technically, and proven.
Why Myra Is Relevant for Hospitals
Myra Security GmbH is a German company without a parent company, headquartered in Munich. Since 2012, Myra has protected websites, web applications, APIs, and DNS infrastructures—with a focus on critical infrastructure, public authorities, and regulated industries. With Myra AI Workspace, Myra brings the same security and compliance standards to the use of AI.
BSI C5 Type 2 attested Myra infrastructure: Myra’s infrastructure has been audited against the C5 cloud criteria catalog of Germany’s Federal Office for Information Security (BSI) with a Type 2 attestation. Type 2 assesses the operating effectiveness of controls over an audit period.
Listed by the BSI as a qualified DDoS mitigation provider: The BSI lists Myra Security as a qualified DDoS mitigation provider.
Audited and operated on Myra infrastructure in the EU: The platform, guardrails, and PII detection run on Myra infrastructure; for external models, processing takes place at the respective model provider.
German company without a parent company: Myra Security is subject to European law.
Audited and operated on Myra infrastructure in the EU
Introduce AI in Hospitals—With Control Over Patient Data
AI for healthcare providers must reduce documentation workload while respecting data protection, medical confidentiality, and evidence requirements. Myra AI Workspace provides a central AI work environment with protection tiers per project, PII masking, a visible protection status, and operation on Myra infrastructure in the EU.
Discover Myra AI Workspace—from €13 per user per month, excl. VAT
Go to Myra AI WorkspaceFAQ: AI in Hospitals and Patient Data
In principle, yes, if the requirements are met: a legal basis under Article 9 GDPR, a data processing agreement, a confidentiality obligation for all service providers involved under Section 203 of the German Criminal Code (StGB), compliance with Section 393 of Book V of the German Social Code (SGB V) for cloud services, and usually a data protection impact assessment. Private or freely available AI accounts generally do not meet these requirements.