AI is gaining ground in healthcare

The first AI wave fascinated with specialized diagnostic algorithms, but mostly failed to scale. Example: Kidney cancer is one of the 10  most common types of cancer, but there are fewer than 15,000 new cases in Germany every year. Even with usage costs of €500 per case, the market potential of around €7.5 million per year remains modest and unattractive for typical investors.

GenAI is now fundamentally changing the playing field in the second wave: instead of complex special cases, it is focusing on everyday time wasters – anamnesis, reporting, and coding. Every hospital visit and every piece of care documentation is a potential use case. As a result, the number of possible uses is growing from a few thousand to several million per year, while the basic AI technology remains the same. The new winning formula: productivity before precision – with consistently high quality.
 

For me, the boom in ambient listening solutions was the most exciting finding at this year's DMEA, Europe's leading event for digital health, in Berlin. AI-based writing assistants can listen in the treatment room, pick up contextual information from patient files, and create suitable notes or a draft report in a matter of seconds. In the largest field study to date at Kaiser Permanente in the USA, over 7,000 doctors used an AI Scribe system on more than 2 million visits; on an average, a time saving of 1 minute per case was achieved.  Does one minute sound short? In total, the time saved amounted to 15,700 hours and, more importantly, 84% of participating doctors reported that AI helped them to focus more on their patients. 

Ambient listening is also finding its way into care: voice apps run on smartphones, transforming dictations into structured entries and thus saving valuable minutes in every care shift. The result: fewer screens, more care – across all professions.

The second wave of AI in healthcare requires more than just technical access and APIs to large language models (LLMs). Market-leading providers are establishing foundation model layers with healthcare-specific security, multilingualism, and seamless tool integration. One example of this is the AI Foundation Services from T-Systems. This offering includes a model hub, vector databases as the basis for “Retrieval Augmented Generation (RAG)”, intelligent prompt control, and comprehensive security mechanisms – hosted securely and confidently in the Open Telekom Cloud (OTC). This allows many application scenarios such as ambient listening, coding agents or patient chatbots to be set up as microservices, even on-premise or hybrid. Raw data remains in the hospital network, while data prepared for the LLM context, such as embeddings, moves to the secure layer in the OTC. The platform concept – instead of a small, specialized solution – provides the turbo boost for scaling and innovation.

A clear example of AI Foundation Services in action is the T-Systems SmartChat. The SmartChat brings GenAI securely into public institutions, hospitals, and health insurance companies. It taps into internal knowledge bases, answers patient and insurance queries, assists in preparing cost estimates, translates medical findings, and pre-fills forms—all 24/7 in the correct professional language. Open APIs can connect it to TI-Messenger, iMedOne®, or contact center platforms; an integrated user management system controls role and token limits. Hosted in the OTC, health data remains in German data centers; a model-agnostic stack of open-source components prevents vendor lock-in. The OTC provides all prerequisites for sovereign AI: certification for social data according to § 35 SGB I, certificates for GDPR, EU Cloud Code of Conduct, ISO/IEC 27017 and ISO/IEC 27018, as well as BSI C5 Type II and others ensure compliance. In this way, the SmartChat and the AI Foundation Services can evolve from an FAQ tool to the automation hub of the healthcare ecosystem.

AI can only be effective with trust. This is particularly true for the second AI wave with the large language models: hallucinations, data protection, and liability issues temper the excitement.  Providers and users need to tackle these issues together. There are technical answers to only some of the challenges: of course, the infrastructure should be located in Germany or the EU, and of course the operating company itself should have its headquarters in the EU, and thus be fully subject to European jurisdiction. In addition, there are technical guardrails for AI systems and control steps such as “LLM-as-a-Judge” to control and validate the output of the language models. However, some risks require organizational solutions. A governance board could, for example, exclude no-go use cases (such as autonomous diagnoses) or define areas of use (“green zones”) such as documentation or billing that are permitted and clearly regulated. The EU AI Act provides important guidance here with its risk classes. In this way, governance can be transformed from a mandatory regulatory program into a scaling booster that fully exploits the strengths of the second wave of AI – a wide range of possible applications and a high level of acceptance.

¹ Zeitfressern auf der Spur – Dokumentationsaufwand in Münchner Kliniken, Greiling M., 2022, BibliomedManager, Fachwelt Pflege & Krankenhauswirtschaft  

² Ambient Artificial Intelligence Scribes: Learnings after 1 Year and over 2.5 Million Uses, Tierney A.A. et al., 2025, The New England Journal of Medicine  

³ The imperative for regulatory oversight of large language models (or generative AI) in healthcare, Meskó B., Topol E.J., 2023, npj Digital Medicine