AI and Healthcare Professionals: What Practical Uses in Clinical Practice?
Reading time: 9 minutes
Guides
Generative AI entered clinical practices before it entered medical training curricula. Many healthcare practitioners are already using it to draft notes, summarize documents, or brainstorm ideas—often without a formal framework—and the French National Authority for Health (HAS) deemed the issue urgent enough to publish its initial guidelines for use in October 2025 (1). Yet the practical question remains: what is AI truly useful for in private practice, and where must its role end?
This guide categorizes AI applications for healthcare professionals into three tiers, details six practical ways to implement it starting this week, and outlines what should never be done with patient health data, whether in France or internationally.
What AI Does Well, and Where It Fails
Explaining, rephrasing, structuring, summarizing: when it comes to medical drafting and administration, large language models save real time. Clinical reasoning and scientific evidence, however, are an entirely different domain, and that is where they are most prone to error.
Non-Existent Citations
A team of orthopedic surgeons in Nice asked several models to retrieve references for eleven systematic reviews on rotator cuff pathology. Out of 471 generated citations, 28.6% of those from GPT-4 and 91.4% of those from Bard were fabricated (2). The titles, authors, and journals looked authentic; the studies themselves did not exist. This phenomenon is known as hallucination, and it is far from rare.
Decisive Opinions Based on Weak Evidence
A language model does not spontaneously clarify the level of evidence supporting its answers. A retrospective study of thirty patients and a systematic review of randomized controlled trials are presented with the same confident tone, yielding definitive conclusions without caveats.
The tool does not say "I don't know." Prompting it with "Which studies are you relying on, and what is their study design?" can help identify fragile claims, but it is no substitute for critical verification. It is up to the clinician to appraise the evidence.
The HAS Rule: A.V.E.C.
On October 30, 2025, the HAS published its "Initial Key Principles for Using Generative AI in Healthcare," developed alongside the CNIL (1). The framework centers on four actions (A.V.E.C.): Learn (Apprendre), Verify (Vérifier), Estimate (Estimer), and Communicate (Communiquer). Its core principle is that every use must be conscious, supervised, and clinically reasoned. It highlights two harmful practices already observed in practice: relying on hallucinated medical content, and inputting confidential patient information into unauthorized tools.
Health Data: What to Never Entrust to a Consumer Chatbot
The reflex is tempting: pasting a patient's consultation report to obtain a clinical summary or referral letter in ten seconds. This is precisely one of the high-risk practices identified by the HAS in the field (1).
In France, Three Cumulative Regulations Apply
Medical Confidentiality. Article L1110-4 of the French Public Health Code covers all information you hold regarding a patient. Disclosing it to an unauthorized third party is punishable by one year of imprisonment and a €15,000 fine (Article 226-13 of the French Penal Code) (3).
The GDPR. Health data constitutes a special category (Article 9), the processing of which is prohibited except under strictly regulated exemptions. You are the data controller for your patients' data, including when software processes it on your behalf (4).
Certified Health Data Hosting (HDS). Whenever a third party stores health data on your behalf, it must be HDS-certified (Article L1111-8 of the French Public Health Code) (5). Consumer versions of conversational AI assistants lack this certification, and their servers are frequently hosted outside the European Union.
To understand what certification levels actually cover and how to verify compliance across your clinical software, consult our guide to health data security.
Removing the Patient's Name Is Not Enough
"A 34-year-old female, primary school teacher in Quimper, presenting with an ankle sprain sustained during last weekend's trail run." There is no name, yet her colleagues could identify her immediately. The GDPR is clear: pseudonymized data that can be linked back to an individual using auxiliary information remains protected personal data (4).
A simple test before pasting text: could a relative, neighbor, or coworker identify the patient? If so, the prompt must not be submitted. Note also that many consumer tools utilize conversation logs to train future models unless you explicitly opt out. Disabling training does not automatically render the tool compliant with healthcare privacy standards.
What About Other Jurisdictions?
The terminology varies, but the core principle remains identical. Across the European Union, the GDPR applies uniform standards to health data, and the EU AI Act, which took effect on August 1, 2024, introduces additional transparency mandates (6). Switzerland enforces its Federal Data Protection Act, Quebec its Law 25, and the United States enforces HIPAA, which mandates a formal Business Associate Agreement (BAA) with any vendor handling Protected Health Information (PHI). In all jurisdictions, any platform processing patient data must be legally vetted and certified for clinical use.
Green, Amber, Red: Triage AI Use Cases in Your Practice

Most safe, productive use cases require no protected health information at all. Triage is therefore based first on the data entered into the prompt, and second on how you critically appraise the generated output.
Green (Zero Patient Data): Literature monitoring, drafting course material or patient education workshops, clinic marketing, administrative correspondence, and practicing patient communication scenarios.
Amber (Fully Anonymized Clinical Scenarios): Brainstorming differential diagnoses, cross-checking red flags, or preparing a case presentation for a specialist colleague. AI suggests; the clinician evaluates and decides.
Red (Strictly Prohibited): Pasting identifiable clinical assessments or medical reports into consumer tools, delegating definitive diagnostic or surgical indications to AI, citing references without primary verification, or generating exercise programs outside compliant software to send via unencrypted personal messaging.
6 Concrete AI Clinical Use Cases to Test This Week
Each of these workflows can be trialed in fifteen minutes. The phrases in quotation marks are sample prompts designed to be tailored to your specialty.
1. Streamlining Literature Reviews Safely
Locate the study yourself on PubMed, paste the abstract, and prompt: "Summarize this study in five lines for an outpatient physical therapist. What are the study design, sample size, and primary outcome measure? What methodological limitations affect external validity?" You save time on initial appraisal without delegating scientific rigor. The reverse workflow fails: asking AI to find literature and citing it unverified risks relying on fabricated citations.
2. Preparing a Patient Education Workshop or Seminar
A local athletic club asks you to present on sports injury prevention, or a company requests an ergonomic presentation on office-related musculoskeletal disorders. Prompt the model for a 45-minute lesson plan, three key take-home messages, and anticipated audience questions. Then adapt the content using clinical guidelines and your expertise. No patient data is involved.
3. Drafting Administrative and Practice Materials
Online practice listings, clinic website copy, job descriptions for associates or locum tenens practitioners, clinic policies on late cancellations, or letters to landlords and accountants. This delivers immediate administrative time savings with zero privacy risk, provided no patient information is entered. Always review the text: models tend to default to verbose phrasing.
4. Practicing Consultations with Simulated Patients
Prompt the AI to roleplay as "a patient with chronic low back pain for eight months who exhibits kinesiophobia and fears structural damage." Practice open-ended interviewing, reflective listening, and introducing graduated exposure to physical activity. Ask for objective feedback on your communication approach. The patient case is entirely fictional, keeping the exercise risk-free.
5. Broadening Perspectives on Complex Clinical Cases
This is the Amber use case. Describe a presentation with all identifying markers removed: "Female, early 50s, 6-week non-traumatic shoulder pain. What differential diagnoses and clinical red flags should be screened for?" Cross-reference the output with your objective clinical assessment and established clinical guidelines. Avoid asking: "What treatment should I prescribe?" The AI outlines possibilities; clinical judgment remains yours.
6. Dictated Clinical Notes Under Strict Safeguards
Ambient clinical transcription assistants listen during encounters, generate structured SOAP notes, and draft referral letters. The documentation time saved is substantial, provided four criteria are met:
Certified Secure Hosting (such as HDS certification or HIPAA-compliant infrastructure) paired with an executed Data Processing Agreement (DPA) or BAA. A vendor merely claiming to be "compliant" is insufficient.
Informed Patient Consent, obtained verbally prior to initiating audio capture in each consultation.
Systematic Clinical Review. A transcription error directly compromises the medical record, and the practitioner bears legal liability under their electronic signature.
Confined to Clinical Documentation. Home exercise programs must be prescribed through dedicated exercise prescription software featuring clinical video libraries, adherence tracking, and secure messaging—never as raw text copied into unencrypted email.
From Clinical Evaluation to Exercise Prescription: Maintaining a Compliant Workflow

In exercise prescription, data security depends entirely on infrastructure design. In non-compliant workflows, clinical notes are fed into consumer chatbots, outputs are copied into unsecured PDFs or WhatsApp chats, and clinicians lose oversight of data governance and patient adherence. In a dedicated medical workflow, every phase—from baseline assessment to therapeutic follow-up—remains within a unified, compliant ecosystem.
This is the engineering model behind Andrew®. Andrew® AI analyzes structured assessment data (anatomical region, rehabilitation phase, pain scoring, functional goals) to suggest evidence-based exercises from its clinical video library, aligned with international clinical guidelines (e.g., NICE, APTA, HAS). No protocol is assigned without direct clinician validation. Programs transfer straight to the patient app, complete with compliance tracking, clinical check-ins, and end-to-end encrypted messaging hosted on certified HDS servers in France. Tailoring an evidence-based home program takes under one minute, compared to five to ten minutes manually.
Would you like to review this clinical workflow on a typical case? A peer clinician from our clinical specialist team can demonstrate how AI streamlines documentation and prescription without compromising compliance. Book a free live demo.
For related reading, see our articles on digital continuing professional development and how healthtech helps clinicians save clinical and administrative time.
Where to Start This Week
Select one Green use case, test it for fifteen minutes, and evaluate its clinical utility. Proceed to Amber workflows using strict de-identification standards. Red practices remain prohibited regardless of potential time savings.
The broader debate surrounding AI's place alongside clinical expertise is explored in our article Will AI replace healthcare professionals? To evaluate your current practice software stack with an experienced clinician, schedule a consultation with a clinical specialist.
Article Sources
Your questions
Your questions, answered
Primarily to save time on non-clinical tasks: summarizing a scientific paper, preparing a prevention workshop, drafting practice documentation, or role-playing a consultation with a simulated patient. With an anonymized case, it can also suggest avenues for differential diagnosis. It replaces neither the clinical examination nor the practitioner's medical judgment.
No, not in a consumer-grade version. In France, healthcare data is protected by medical confidentiality, the GDPR, and the requirement for certified health data hosting (HDS certification), which these tools do not have. Removing the name is not enough: if a relative or a colleague can identify the patient, the data remains personally identifiable.
No. It can suggest hypotheses, but it responds with the same confidence whether the evidence is robust or weak, and it sometimes fabricates citations. The HAS (French National Authority for Health) requires that every use be conscious, supervised, and reasoned, and that the practitioner verify the tool’s output.
Yes, provided that the tool is hosted by an HDS-certified provider (or the legal equivalent in your country), that a data processing agreement is executed, that the patient gives their consent prior to each recording, and that the clinical note is reviewed. As for the exercise regimen, it must be prescribed through prescription software with compliance monitoring.
In its "Initial Key Recommendations for the Use of Generative AI in Healthcare," published on October 30, 2025, in partnership with the CNIL, the HAS introduces the A.V.E.C. method: Apprendre (Learn), Vérifier (Verify), Estimer (Assess), Communiquer (Communicate). It cautions against AI-generated hallucinations and the disclosure of confidential information.
Andrew® AI uses clinical assessment data (anatomical region, phase, pain level, therapeutic goals) to suggest exercises from the video library, based on international guidelines. The practitioner reviews and validates everything prior to sending. The rehabilitation program is delivered directly to the patient app, complete with compliance tracking and encrypted messaging, and all data remains hosted in France on HDS-certified (Healthcare Data Hosting) servers.
Are you still not listed among Andrew® therapists?
- Video prescription of exercises and health education
- Content created by the best French trainers
- Teleconsultation and billing
- Customized program planning with Andrew® Coach
- Patient-centered follow-up with hundreds of hours of sports, health, and wellness content




