How to read the traffic light

    Your industry: Dentists and healthcare practices

    Tailored to medical, dental, physio and psychotherapy practices: the recommendations of the FMH and SSO require processing in CH/EU/EEA and no provider access to the data, and professional secrecy (Art. 321 Criminal Code) is factored in. The rating is therefore stricter for identifiable patient data.

    Green: proven and recommended

    Works with client data: the setup is proven, contract and governance are in place. This is exactly how you should do it.

    Yellow: unclear, depends on the setup

    Depends on the contract, configuration and industry: feasible with a documented risk assessment, not without it.

    Red: not possible

    Not defensible for client data in our risk assessment: no contract, no control or no documented safeguards. Often still the best choice for research, marketing and anonymised content.

    In the practices view: Azure Switzerland Standard is green only with Modified Abuse Monitoring and store=false; without these approvals it turns yellow, because Microsoft can otherwise store flagged samples that authorised staff may review. The EU options turn red for identifiable patient data: in practice data almost everything is health data, free text can hardly be pseudonymised reliably, and breaching professional secrecy (Art. 321 Criminal Code) carries a custodial sentence of up to three years or a monetary penalty. They stay yellow only for fully anonymised content.

    All AI options at a glance

    Filter the table by traffic-light colour. On a smartphone it scrolls sideways.

    SolutionClient dataModelsData locationModel qualityEffortCostWhat to watch
    Azure OpenAI Switzerland (Standard)
    Microsoft Azure, Switzerland North region
    GreenGPT-4o (until April 2027), Embeddings
    Processing
    Switzerland
    Storage
    Switzerland
    Solid, several generations behind the frontierMedium, integration by IT or partnerLow, per token, no base feeThe reference choice for patient data, but green only with two approvals: apply for Modified Abuse Monitoring (without it Microsoft stores flagged samples that authorised staff may review; with ContentLogging=false this is verifiably disabled) and store=false. Without both, yellow applies. Only managed customers can apply with Microsoft, we clarify this for our practice customers. Choose the Standard deployment type (not Global or Data Zone), commissioned processing contract, confidentiality of all auxiliaries (Art. 321 Criminal Code). For new customers, GPT-4o (until April 2027) is effectively the only option; GPT-4.1 is deprecated.
    Swiss AI platforms
    e.g. AIgent, Swisscom, Phoenix Technologies (kvant), Infomaniak
    GreenOpen models such as DeepSeek, Qwen, GLM, Apertus
    Processing
    Switzerland
    Storage
    Switzerland
    Good to very good, a step behind the frontierLow, ready-made interface, some with agentsMedium, licence per user, sometimes plus usageA straightforward compliant choice for patient data: processing and storage stay in Switzerland. Name confidential computing as a review point: providers such as Phoenix Technologies use it to protect data even during processing, which comes closest to the FMH requirement of no provider access.
    AI in your industry software
    e.g. practice, law firm, CRM or portfolio systems
    Depends on providerUsually GPT or Claude in the background
    Processing
    Depends on provider
    Storage
    Depends on provider
    Good, tailored to the use caseLow, the system is already in useLow to medium, often as an add-on modulePractice software: clarify in writing which model runs where and pass the FMH requirements on to the vendor. Also clarify the medical device status: diagnostics, such as X-ray analysis, only with certified software (MedDO or MDR); general AI models only for administration, correspondence and drafts.
    Microsoft 365 Copilot
    Microsoft, with EU Data Boundary
    YellowGPT models from OpenAI, operated by Microsoft
    Processing
    EU/EFTA
    Storage
    Switzerland (Data Location or ADR)
    Very good, strong in everyday office workLow, licence and configurationMedium, licence per user and monthOnly with a commissioned processing contract, Flex Routing disabled, Anthropic models switched off and permissions cleaned up. The FMH and SSO recommendations require processing in CH/EU/EEA and no provider access to the data.
    Azure OpenAI Data Zone EU
    Microsoft Azure, resource in Switzerland North
    RedLatest GPT generation, e.g. GPT-5.4 to GPT-6.1
    Processing
    EU Data Boundary, confirm countries with Microsoft before use
    Storage
    Switzerland
    Frontier, latest GPT modelsMedium, integration by IT or partnerLow, per tokenRed for identifiable patient data: the region would comply with FMH guidance, but breaching professional secrecy (Art. 321 Criminal Code) carries a custodial sentence of up to three years or a monetary penalty, and free text can hardly be pseudonymised reliably. Yellow only for fully anonymised content. The EU Data Boundary can also include EFTA countries: confirm the affected countries with Microsoft before use.
    Claude via AWS Bedrock (EU)
    Amazon Web Services, EU regions
    RedOpus 5 and Sonnet 5 in Ireland or Stockholm, Opus 5.5 and Sonnet 5.5 via EU inference
    Processing
    EU
    Storage
    No prompt storage by Bedrock
    Frontier, especially strong for text and analysisMedium to high, AWS account and development neededLow to medium, per tokenLike Data Zone EU: red for identifiable patient data, yellow only for fully anonymised content; also review the AWS sub-processing.
    Gemini via Google Vertex AI (EU)
    Google Cloud, EU regions
    RedGemini models
    Processing
    EU
    Storage
    EU or Switzerland
    Frontier, strong with long documentsMedium to high, Google Cloud project and developmentLow to medium, per tokenLike Data Zone EU; have the processing location verified in writing. Red for identifiable patient data, yellow only for fully anonymised content.
    Claude directly from Anthropic
    Claude API, Team and Enterprise
    RedClaude Fable 5.1, Opus 5.5, Sonnet 5.5
    Processing
    US or global
    Storage
    US
    Frontier, newest Claude models firstLow, ready to useMedium, licence or per tokenUS processing is outside the permitted regions and endangers professional secrecy (Art. 321 Criminal Code): unsuitable for patient data.
    Claude in Microsoft Foundry or Copilot
    Microsoft, operated on Anthropic infrastructure
    RedClaude models
    Processing
    Outside the EU Data Boundary
    Storage
    Not guaranteed
    Frontier, same as Claude directLow, switchable at a toggleMedium, in licence or per tokenProcessing outside the EU: unsuitable for patient data, keep the switch off.
    Copilot with Flex Routing enabled
    Microsoft 365, setting available since April 2026
    RedGPT models
    Processing
    At peak load also US, Canada, Australia
    Storage
    Switzerland
    Very good, same as CopilotLow, a single settingMedium, licence per user and monthPractices must disable this: peak-load processing in the US/Canada/Australia breaches the FMH and SSO region rule.
    Private AI accounts
    ChatGPT, Claude.ai, Gemini with a private or free account
    RedAll of them
    Processing
    Unclear, mostly US
    Storage
    US
    Frontier, which makes it so temptingNone, and that is exactly the problemLow, free or privately paidCriminally punishable with patient data (Art. 321 Criminal Code, custodial sentence of up to three years or a monetary penalty): prohibit by policy, offer a compliant alternative.

    Note on the US CLOUD Act

    Microsoft, AWS and Google are US corporations and are subject to the US CLOUD Act, even with data centres in Switzerland or the EU. This residual risk belongs in every risk assessment.

    The best solutions: three proven setups for dentists and healthcare practices

    Hardly any organisation needs just one tool. What has proven itself is the combination of an AI employee for daily work and a secure interface for automations.

    For practices we recommend the strict setup: a Swiss platform, ideally with confidential computing, plus Azure Switzerland Standard with Modified Abuse Monitoring. Patient data stays in Switzerland end to end, which satisfies the FMH and SSO recommendations most easily. Provisioned with GPT-5.1 is uneconomical for any practice (from approx. USD 13,000 per month).

    Proven for everyday office work

    Pragmatic: Copilot plus Azure Switzerland

    • M365 Copilot for emails, documents and meetings, with storage in Switzerland and Flex Routing disabled
    • Azure OpenAI Switzerland North for automations with sensitive client data, such as document review or reports: for new customers, GPT-4o (until April 2027) is effectively the only option there
    • A clear policy: what may go into Copilot, what only goes through the Swiss pipeline
    Recommended for practices

    Strict: Swiss platform plus Azure Switzerland

    • A Swiss AI platform as the working surface, ideally with confidential computing
    • Azure OpenAI Switzerland North (GPT-4o) with Modified Abuse Monitoring for automations
    • External models in the platform selectively disabled per use case
    When quality matters above all

    High performance: frontier models in the EU

    • Azure OpenAI Data Zone EU or Claude via AWS Bedrock EU
    • Pseudonymise where possible before data reaches the model
    • Risk assessment, contract and client information with extra care

    What professional secrecy and data protection require

    Patient data is especially sensitive personal data under the revised DPA and is subject to professional secrecy under Art. 321 of the Criminal Code, which since 1937 has grown to cover far more professions, including physio- and psychotherapy. The FMH and SSO recommendations on IT in the practice and the documentation and retention duties under cantonal health law apply regardless of the technology.

    • AI inventory: which applications, which data, which model, which data location
    • Risk assessment per application: process patient data preferably in Switzerland
    • Policy for the practice team: approved tools, no patient data in private AI accounts
    • Contractually secure the confidentiality of all auxiliaries and processors (Art. 321 Criminal Code), following the FMH guide on confidentiality agreements that explicitly names cloud providers
    • Conclude cloud contracts on the basis of the FMH framework agreement for cloud services
    • Exchange patient data only in encrypted form (e.g. the factsheet of the Zurich Health Directorate)
    • Comply with the documentation and retention duties under cantonal health law, also for AI-supported records
    • Inform patients about the use of cloud services and where their data reaches third parties
    • Diagnostics, such as X-ray analysis, only with certified medical device software (MedDO or MDR); general AI models only for administration, correspondence and drafts
    • Remove DICOM metadata (name, date of birth) before any AI processing of X-ray images

    This overview does not replace legal advice. Model availability and regions change constantly; October 2026 edition. Verify details with the provider before use.

    Sources

    • FDPIC: Guide to checking the admissibility of transferring personal data to countries without an adequate level of data protection
    • FDPIC: Factsheet on commissioned processing of personal data under the revised DPA
    • Microsoft Learn: Model availability by region in Azure Foundry
    • Microsoft Learn: Deployment types (Global, Data Zone, Standard, Provisioned)
    • AWS: Model availability by region in Amazon Bedrock
    • Google Cloud: Data residency for generative AI
    • Revised DPA: patient data as especially sensitive personal data (Art. 5)
    • Criminal Code, Art. 321 (professional secrecy)
    • FMH: Technical and organisational requirements for cloud services (2020)
    • FMH: IT baseline protection for medical practitioners (2023), in cooperation with the SSO
    • FMH: Framework agreement for cloud services and guide on confidentiality agreements for auxiliaries
    • Medical Devices Ordinance (MedDO) and EU MDR: certification duty for AI with a medical intended purpose
    • Health Directorate of the Canton of Zurich: factsheet on encrypted data exchange (example of one canton)

    We implement the recommended setup for you

    Taskstreams builds and operates the pipeline that fits your policy: Switzerland only with Azure OpenAI Switzerland North, or the frontier model in the EU while your client data stays in Switzerland. Contract and governance are included; you receive finished AI employees for your processes, from document review to client reporting, instead of your own infrastructure project.

    • Processing and storage in Switzerland (Azure Switzerland North)
    • Contract, policies and documentation for data protection and supervision included
    • Fixed-price pilot, then ongoing operations by us
    • Modified Abuse Monitoring and a confidentiality agreement under Art. 321 Criminal Code included