Back to Blog

    Agentic AI

    AI agent for medical practices: building and connecting Linda

    Linda, our AI agent for medical practices: architecture on Copilot Studio and Azure OpenAI, connecting to practice software and checks before a pilot.

    June 22, 20245 min read
    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    marcel.haas@cnext.ch
    CNEXT AI Agent
    In this article

    An AI agent can take administrative work such as appointment requests off a medical practice's hands, provided the practice team controls every relevant step. We built Linda for this: a virtual agent that sits on top of the existing practice system instead of replacing it, and connects to the practice software via REST/OData, SQL/ODBC or RPA. Diagnoses, treatment recommendations and medication decisions do not belong in such an agent.

    What we built

    Linda is our virtual AI agent for medical practices. She is not practice software in the classic sense: we provide the agent and the AI technology, which build on the practice's existing data and systems.

    • For patients: an app for iOS and Android, or a web application that can be embedded in the practice's existing website. Patients sign in with a personal login, so Linda is not a general chatbot but works with the data of the signed-in person.
    • Communication: by text chat, by voice input through the microphone, by phone call or by digital audio call. Linda mainly speaks German and can generate speech output in Swiss German or other languages.
    • For the practice team: Linda runs in Microsoft Teams or in a web browser.

    Architecture

    Linda is based on Microsoft Copilot Studio and other Power Platform and Azure services:

    ComponentRole in Linda
    Microsoft Copilot StudioFoundation of the agent and its dialogues
    Power AutomateAutomation of workflows and routine tasks
    Microsoft DataverseData storage within the Power Platform
    Model-driven Power AppBack end for managing and analysing the data
    Azure OpenAILanguage models for dialogues and generated content
    Azure AI SearchSearch index for retrieval-augmented generation (RAG)

    Linda does not train a model of her own. She uses existing language models and brings in the practice's data through retrieval-augmented generation, so that answers are based on the practice's specific, current data. The data sits in the practice's Microsoft cloud tenant, not with CNEXT. For operations and support we need access to some of the data; the practice grants that access.

    Connecting to the practice software

    A practice usually already works with practice management software. Linda builds on it and exchanges data with this core system through interfaces, some of them AI-assisted. For example, appointments can be managed in the agent and transferred to the core system in near real time. We offer three ways to connect:

    InterfaceWhen it fits
    REST/ODataThe practice software offers a modern REST interface that Linda connects to.
    SQL/ODBCThe system offers an ODBC-based interface.
    Robotic process automation (RPA)No other interface to the core system is possible; RPA still reads and writes the data.

    Features

    Linda is organised into modules:

    • Appointments: booking and rescheduling by chat or voice in natural language; appointment length based on a predefined service catalogue and past experience; scheduling of cleaning time and patient changeovers; reminders for the practice team and for patients; synchronisation with the calendar in the practice management system.
    • Staff and rooms: choice of clinician by availability and range of treatments, replacement appointments when bookings move, room allocation by staff availability and required equipment, notifications about changes.
    • Patient-related modules (history, treatments, record, vaccinations): patient access to selected health data such as vaccinations, treatments and medication, plus reminders for vaccinations and check-ups.

    We recommend starting with a narrow, non-clinical use case such as appointment requests. The patient-related modules process health data, so the checks in the next section matter most for them.

    What to settle before a pilot

    Patient data are highly sensitive. Confirm the functions, data flows and limits of the specific deployment in writing, and have the practice owners responsible review them:

    • What inputs, conversations and documents are processed, and for what purpose?
    • Which providers and subprocessors receive data? Where is it processed and stored?
    • How long are inputs, logs and outputs retained, and who can access them?
    • Which interfaces write back to practice or patient records, and who approves those changes?
    • How are outages, incorrect answers, complaints and security incidents handled?

    Data protection

    A data-location promise should cover the complete processing chain, not just the selected cloud service. The Swiss Federal Data Protection and Information Commissioner (FDPIC) states that a practice remains responsible for its data protection obligations as controller, even when it uses a cloud service. This covers the contract, technical and organisational measures, subprocessors and any cross-border transfers. Health data are sensitive personal data, and professional secrecy obligations may also apply. A data protection impact assessment may be required where processing is likely to result in a high risk. Linda also requires a data protection agreement between the practice and the patient, or an extension of the existing agreement.

    Clinical boundaries

    Do not let a system independently make diagnoses, recommend treatment or decide on medication. Swissmedic assesses software by the intended purpose the manufacturer defines, and a medical purpose can bring software under medical device regulation. Language models can also misunderstand requests. The practice team should therefore control every relevant step, especially where data flows back into the practice system.

    How we help

    • Consulting: analysis of your practice's requirements
    • Development: virtual assistants built on Copilot Studio and tailored to your workflows
    • Implementation: integration into your existing systems
    • Training and support: training for the practice team and ongoing support

    For data protection questions in the Microsoft cloud, we also offer our own product, DSG next. We are also interested in partnerships with practice software vendors in German-speaking countries.

    Sources

    Agentic AI
    Teilen:

    AI tools may have supported the creation of this article. CNEXT is responsible for the published content. Product features and prices can change; check the linked primary sources for current details.

    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    6x Microsoft Applied Skills

    See Linda in a demo

    We show you Linda and discuss how she can be connected to your practice software.