The doctor decides. The AI prepares. The patient saves time.

Ärztliche Dokumentation im Praxisalltag: Bis zu 44 Prozent der Arbeitszeit entfallen auf administrative Aufgaben

A patient comes to the consultation. They have filled out a questionnaire, brought imaging, and their lab values are in the system. The doctor is trained for precisely this moment: classifying findings, weighing risks, recommending a therapy that suits this person. What they do instead: They manually transfer questionnaire answers, repeatedly formulate the same finding text, and dictate a doctor's letter that is still not finished two hours later.

This scenario is not an isolated case and not a specialty-specific problem. It affects internal medicine as much as orthopedics, rheumatology as much as endocrinology, geriatrics as much as neurology. 44 percent of hospital doctors' working time is spent on documentation. One third for office-based doctors. 93 minutes daily, and even 120 minutes for internists. Over 90 percent of office-based doctors feel overburdened by bureaucracy. These are not subjective complaints. These are measured data from German hospitals and practices.

By 2040, the Central Institute for Statutory Health Insurance Physician Services expects a shortage of 30,000 to 50,000 doctors in Germany. More patients, fewer doctors, more documentation requirements. The equation is structurally unsound, unless the workflow is fundamentally rethought.

This article describes an AI-supported clinical workflow that does exactly that. It is transferable to any specialty that has the same basic structure: anamnesis, examination, imaging or measurements, laboratory, findings, doctor's letter. And it is explicitly aimed at those who make decisions: medical directors, clinic managers, practice owners, medical controlling, and IT managers in healthcare facilities.

The Basic Problem: Media Discontinuities Cost Money, Time, and Personnel

Analyzing the clinical documentation workflow in most specialties reveals the same pattern: data is generated in one place and manually transferred to another. The patient answers questions on paper, and the receptionist types them in. The doctor reads a finding and reformulates the classification, even though the structured raw data is available digitally. The lab provides values, which are then manually entered into the doctor's letter. Each of these transfers costs time. Each is a potential source of error.

For clinic decision-makers, this is not an abstract problem: it ties up doctor hours that are then unavailable for patient care, increases the risk of errors during manual transfer, and makes it harder to retain qualified staff who spend a significant portion of their working time on administrative tasks for which they were not trained.

The actual problem is not documentation itself. It is the missing connecting layer between the steps. Data that has been structured does not need to be formulated again. Patterns that a system can recognize do not need to be identified anew each time. Finding texts that do not differ significantly in similar constellations do not need to be rewritten each time. This is precisely where AI comes in, not as a decision-making authority, but as an intelligent connecting layer.

This pattern works in dozens of specialties

The following AI-supported clinical workflow is not specialty-specific. It follows a universal basic pattern that differs from specialty to specialty only in content, not in structure. The AI architecture behind it is the same. What changes are the underlying guidelines, relevant parameters, and specialized terminology.

The pattern is identical: questionnaire, examination, measurements or imaging, lab values, guideline-compliant classification, doctor's letter. Only guidelines, parameters, and specialized terminology change. The AI architecture behind it remains the same.

AI as an intelligent integration layer: questionnaires, imaging, lab results, and physician reports in a seamless workflow

The AI-supported clinical workflow: step by step

📋

1. Digital Patient Questionnaire AI Support

The patient fills out the questionnaire digitally before the appointment, on a smartphone, tablet, or terminal in the practice. No paper forms, no manual transfer, no legibility issues. The questionnaire is configurable for specific specialties: risk factors, pre-existing conditions, medications, current complaints, medical history.

In addition to yes and no, there is explicitly an 'Uncertain' option. This is clinically important: patients often don't know if an event in their history "counts," if a relative really had a specific illness, or if a symptom is relevant. These answers are not counted as 'no' but are marked as a clinical hint for the medical consultation. The AI analyzes the completed answers and creates a prioritized summary for discussion before the doctor enters the room.

AI: Identifies risk constellations, prioritizes discussion points, marks uncertainties. The doctor sees a structured summary, not a raw form.

🩹

2. Anamnesis and Consultation Doctor Decides

The medical consultation remains what it is: irreplaceable and not automatable. Facial expressions, context, the quality of a symptom, the answer to a follow-up question, the clinical judgment after years of experience: these are pieces of information that no system can capture or replace. The AI has prepared. The doctor conducts the conversation and evaluates the answers in the overall context.

The structured discussion guide based on the questionnaire evaluation reduces the time for routine questions and creates space for the clinically important discussion moments that are not on a form. This is the real gain: not automation of the conversation, but better quality through better preparation.

AI: Supports conversation preparation. Clinical assessment remains entirely with the doctor.

The doctor-patient consultation remains irreplaceable: AI prepares the ground, while the doctor makes the decision and takes the lead

🔬

3. Automatically Classify Imaging and Examination Findings AI Support

X-rays, MRI, ultrasound, ECG, spirometry, bone density measurement: findings are available digitally. The system captures the structured measurements and automatically assigns them to clinical categories, based on the respective guidelines for the specialty. Normal values, thresholds, classifications do not need to be looked up manually. The preliminary assessment is presented to the doctor for validation.

What AI does not provide: clinical contextualization. Whether a finding, in combination with anamnesis, medical history, and current symptoms, takes on a different urgency than when viewed in isolation, remains a medical decision. The AI provides the framework. The doctor fills it with clinical judgment.

AI: Classifies measurements, prepares finding text, calculates reference values. Doctor validates and contextualizes.

🧪

4. Lab Values: Recognize Patterns, Not Just List Values AI Support

A single lab value is a number. Clinical significance arises from the interplay of multiple values in the patient's context. An elevated CRP alone is non-specific. Elevated CRP together with elevated ESR, positive rheumatoid factor, and anemia is a pattern that can indicate active inflammatory rheumatic disease and requires a different therapeutic decision than each individual value alone.

The AI compares all lab values with normal ranges, recognizes clinically relevant constellations, and marks them as treatment-relevant. It prepares the lab classification for the doctor's letter in a structured way. No findings report leaves the system without medical approval. But the doctor no longer has to manually assemble the constellation.

AI: Recognizes patterns, marks treatment-relevant constellations, prepares classification. Medical assessment is mandatory.

🎙

5. Dictation with GDPR-compliant AI Transcription AI Support

Whisper is an open-source speech recognition model that can be operated on local infrastructure without transferring speech data to external cloud services. The doctor dictates, Whisper transcribes precisely, even with medical terminology, and the transcript is structured and assigned to the correct patient. For clinics with data protection requirements, this is a crucial difference from cloud-based dictation solutions: no speech data transfer outside, full GDPR compliance.

This replaces the typist or the queue between dictation and typing in many practices and clinics. At the same time, the dictation remains entirely the doctor's: no text is generated that was not spoken. Whisper transcribes. It does not formulate. This is the decisive difference from AI-generated free text.

AI: Transcribes, structures, assigns. What is said remains the doctor's; nothing is added or changed.

📄

6. Automatically Generate Doctor's Letter AI Support Doctor Decides

Once all data is available, the system generates a complete doctor's letter at the push of a button. With salutation, diagnosis, ICD codes, detailed presentation of findings, lab classification, guideline-compliant therapy recommendation with specific preparation and dosage, follow-up recommendation, and closing formula. Specialty-specific, structured, according to the current guidelines of the respective specialty.

The letter is created in the personal style of the doctor. The system learns this style from stored sample letters, which are saved individually for each doctor and are not accessible to any other person in the system. According to Handelsblatt, hospital doctors spend an average of ten hours per week just writing doctor's letters.⁵ The UKE reports that its AI system ARGO drafts the epicrisis in less than a minute.⁶ The doctor reviews, adjusts, and approves.

Result: Medical responsibility lies entirely with the doctor. The documentation burden no longer does.

The AI Assistant as a Clinical Sparring Partner

In addition to the structured documentation workflow, the system includes an interactive AI assistant. It is not a generic chatbot. It is trained on the guideline knowledge of the respective specialty and is capable of answering specific clinical questions at a specialist level. Directly within the system, without switching systems, without waiting time.

The area of application is deliberately conceived differently from the rest of the workflow: not automation, but sparring. If a question arises during a conversation with a patient for which the doctor has no immediate, certain answer, if they are at the borderline of the guidelines for a therapy decision, if they want to know what evidence supports a particular recommendation, or if they are looking for a second opinion on an unusual constellation of findings, then the AI assistant is the contact person.

The assistant distinguishes between stable guideline knowledge, which is answered directly from the stored knowledge base of the specialty, and dynamic product knowledge, such as current availabilities or new study data, for which it researches in real-time. This distinction is not optional in a medical context, but mandatory. Stable guideline knowledge and potentially outdated product knowledge must not be mixed. The system knows this boundary and communicates it transparently.

Data Protection: Architectural Decision, Not Package Insert

No topic raises skepticism faster among clinic IT managers than AI in conjunction with patient data. Rightly so. The question is not whether data protection is relevant, but how it is implemented technically and contractually. General statements are not sufficient at the decision-maker level. Here is the current status.

Zero-Data-Retention: What that specifically means

Zero-Data-Retention (ZDR) is a contractually defined agreement with Anthropic, the manufacturer of the underlying AI model. It guarantees that requests and the patient data contained therein are not stored, not logged, and not used for training future models. The data is processed exclusively for handling the respective request and is not persisted thereafter. This is not a marketing statement, but a contractually enforceable standard.

Client Segregation Down to Vector Level

The system uses a RAG (Retrieval-Augmented Generation) knowledge base for the guideline knowledge of the respective specialty. RAG systems use embedding vectors for semantic search. A known security problem with poorly implemented multi-tenant systems is cross-tenant leaking: Institution A could theoretically access knowledge vectors from Institution B. The system prevents this through strict tenant ID filtering at the database level before any retrieval even takes place.

Three Operating Models for Different Requirements

What this specifically means for clinic decision-makers

A continuous AI-supported workflow not only saves time. It reduces sources of error due to media discontinuities, improves documentation quality through structured data collection, relieves doctors of tasks below their qualification level, and creates capacity for what they were trained for: clinical decisions and patient consultations.

The German Hospital Federation (DKG) has specifically calculated the savings potential: If doctors and nurses spent just one hour less on documentation daily, it would free up approximately 21,600 doctors and 47,000 nurses in full-time equivalents.⁷ Not hypothetical, but through efficiency gains within the existing system.

Given the projected shortage of 30,000 to 50,000 doctors by 2040, this is not a marginal efficiency optimization. This is a structural contribution to securing healthcare provision. The question is no longer whether. The question is how, and who retains control.