How Does AI Work During a Clinical Visit?

During a clinical visit, AI does not replace the physician; it supports them at different stages of the process. It turns the conversation between physician and patient into a note, brings past patient information to the physician’s attention, and offers suggestions that support decision-making. In Dr.Medkit, all of this happens in real time while the visit is under way; the clinical decision, the review of prepared content, and final approval always remain with the physician.

What exactly does AI do during a visit?

AI-assisted documentation systems used during visits follow an “ambient AI” approach: they listen to the room, turn the conversation between physician and patient into clinical notes, and aim to reduce the physician’s need to take notes separately. Although it varies by system, the process can be described in four core stages.

• Listening: With the patient’s consent, the visit conversation is processed and converted into text. This can be done without storing the audio permanently.
• Structuring: The conversation is organized into clinical note sections such as complaint, history, examination findings, and plan. Irrelevant talk is filtered out and a clean draft is prepared.
• Coding and record preparation: In systems that support it, relevant diagnosis and procedure codes are identified from the clinical notes and presented for the physician’s review.
• Approval: The prepared draft comes to the physician, who reviews the content, makes any corrections, and gives final approval. The final decision and approval on the record always remain with the physician.

One large-scale example of this approach came from Kaiser Permanente Northern California: in an evaluation published in 2024, 3,442 physicians used an ambient AI system in 303,266 patient encounters over 10 weeks. Physicians reported positive experiences, particularly around reduced documentation burden and less after-hours charting¹.

These results, however, rely mainly on user feedback; they do not show that every clinic will see the same level of time savings.

What does it not do in place of the physician?

During a visit, AI can process clinical information, put possible diagnoses forward for consideration, and prepare suggestions that support decision-making. Making the diagnosis, deciding on treatment, and determining the clinical approach for the patient, however, remain the physician’s responsibility.

The World Health Organization’s 2021 guidance on AI in health lists keeping medical decisions under human control and using AI systems under human oversight among its core principles². AI supports the clinical process, but it does not take over decision-making authority.

This approach has a name: “human-in-the-loop,” meaning the decision and final review stay with a person. In a workflow built on this principle, clinical notes, diagnosis codes, and other records are not finalized or sent to health systems without the physician’s approval. Dr.Medkit follows exactly this approach; reviewing prepared content and giving final approval always remain with the physician.

How does Dr.Medkit set up the process?

In ambient AI systems, the visit conversation can be processed during the encounter, but in most products the clinical note is prepared after the visit ends. Dr.Medkit moves this process into the visit itself. It does not just convert the conversation between physician and patient into text; it interprets the clinical information in the conversation and builds a structured history in real time as the visit goes on. In the same process, the patient’s past information and current clinical findings are assessed, and scenario-specific suggestions based on verified medical sources are presented to the physician. The end of the visit is then no longer the moment documentation begins, but the final step where prepared content is reviewed and approved.

Before the visit: the patient’s history is ready

Writing the note for a single visit and knowing a patient’s five-year history are different tasks. Dr.Medkit works with a clinical memory that brings together previous visits, medications, and past clinical findings in the system. Before the patient arrives, the physician can access a summary of past health information prepared from existing records. The physician can prepare for the visit without re-reading the patient’s history, and once the conversation starts, this information provides context for the history and clinical suggestions prepared in real time.

During the visit: notes and suggestions in real time

In most products, the physician does not see what is being written during the visit; the clinical note is presented in one piece after “finish and save.” In Dr.Medkit, documentation progresses alongside the visit. As the patient describes their complaint, the history section starts to take shape, and as examination findings are discussed, the relevant fields are updated. While the physician focuses on the patient, clinical documentation keeps moving in the background.

At the same time, the system assesses the patient’s past information together with the clinical findings that emerge during the conversation. It brings possible diagnoses, differential diagnoses, and additional questions worth asking to the physician’s attention. These suggestions are presented with their sources while the visit is still under way. The physician therefore has the chance to consider points that might be missed and ask any questions they see fit while the patient is still in the room. The system supports this assessment; the physician decides which suggestions to consider and how to proceed.

Test orders are placed within the same flow. The physician can choose frequently used tests from ready-made sets or create a new set for the patient; the ordered tests are written directly into the visit note and the summary screen.

Suggestions come with their sources and scenario

Dr.Medkit builds its clinical suggestions from the patient’s past history, current complaints, and examination findings. Suggestions are based on clinical guidelines and peer-reviewed scientific publications accessed through platforms such as PubMed and Google Scholar, and are presented to the physician together with the relevant clinical scenario and sources. Throughout this process, the system’s internal source verification module separately checks the basis of each suggestion for source type, recency, and consistency with current guidelines. No suggestion is generated from unverified sources, and cases where verification could not be completed are clearly flagged to the physician.

Suggestions are designed to be reproducible under the same clinical conditions: the same history and the same findings produce the same evidence-based suggestion. The physician can see not only which suggestion is offered, but why it was generated, which sources it relies on, and its verification status. The decision stays with the physician.

After “Finish”: a summary and a single approval

When the visit ends, Dr.Medkit presents the history prepared during the conversation, clinical notes, diagnoses considered, relevant ICD diagnosis codes, ordered tests, and procedure details to the physician as a single summary. The content, orders, and codes created during the visit come together on one screen for a final review.

The physician reviews the summary, edits any fields they see fit, and gives final approval. The approved information is written to the patient’s record in the system and becomes ready to send to e-Nabız, Türkiye’s national health record system. The documentation that progressed throughout the visit is completed under the physician’s review and approval.

More than a new interface: a new way of working

Traditional clinic management software moved many paper-based tasks into a digital environment. The need for data entry and navigating between screens, however, remains. A 2018 multi-center study with 55 emergency physicians in the US found that the number of clicks and the time needed for the same clinical tasks varied significantly between systems³. Dr.Medkit automates these workflows with AI while keeping the physician in control.

The physician states what they want to do with written or spoken commands. For tasks such as preparing a prescription, creating a test order, or issuing an invoice, the system carries out the necessary steps from a single command and brings the result to the physician. Instead of moving between different menus for each task, the physician can manage familiar clinical work directly through commands.

Built for Turkish clinical language

Abbreviations, professional terms, and patients’ everyday expressions used in medical conversations can make it harder to interpret conversations correctly. In Turkish medical language, this challenge matters even more because domain-specific language resources are limited⁴.

Dr.Medkit was built around the daily clinical practice of physicians in Türkiye and Turkish medical terminology. The abbreviations physicians use, Turkish drug and procedure names, and the way patients describe their complaints in their own words are all part of the system’s clinical language processing. Conversations are not merely converted into text; structured notes are prepared in the format physicians are used to, taking the clinical context into account.

How is patient data protected?

Under KVKK, Türkiye’s personal data protection law, health data is a special category of personal data and is subject to specific processing conditions and security measures⁵. In AI-assisted clinical systems, where data is stored, how it is processed, and who can access it therefore matter.

In Dr.Medkit, patient data is stored in Türkiye. Conversations during the visit are converted into text, but audio recordings are not kept permanently. Access to patient records is managed according to each user’s permissions.

The structure behind real-time work: agents

In Dr.Medkit, preparing notes, generating clinical suggestions, and carrying out related tasks during the visit are not done by a single model in a single pass. The system consists of AI agents that take on different tasks and work in coordination. Each agent focuses on a specific task while drawing on the information produced by the others. The physician does not have to manage these technical processes one by one.

Five core agents take part in the process:

• Memory agent: Provides the patient’s past records as context for the current visit.
• Listening agent: Converts the visit conversation into text and separates the speakers.
• Note agent: Interprets clinical information and prepares the structured history.
• Suggestion agent: Presents source-backed suggestions to the physician based on clinical findings and past information.
• Record agent: Writes the information approved by the physician to the patient’s record.

The coordination layer that manages these agents splits the physician’s commands into the relevant tasks and organizes the flow of information between agents. It checks the outputs and shows the physician any missing or inconsistent information it detects. This makes it easier to pinpoint at which stage a possible problem arose.

For the physician, this structure has two main benefits: less need to move between screens to complete tasks, and information that needs review becomes visible at the relevant stage. Instead of carrying out each step separately, the physician manages the process through a single structure, and clinical records are finalized with their approval.

Impact on the process: where is time saved?

AI aims to save time by reducing the documentation and administrative workload the physician carries in the clinical process. In a 2016 observational study of 57 physicians across four specialties in the US, about 27% of working time was spent face to face with patients and 49% on electronic records and desk work. The 21 physicians who kept after-hours diaries reported one to two hours of additional work each day⁶.

The effect of this workload is not limited to lost time. In a 2016 survey of 6,375 US physicians, those using electronic records and computerized order entry systems were less satisfied with the time they spent on administrative tasks. Use of computerized order entry was also associated with a higher risk of burnout after accounting for other factors⁷. These findings point to the effect of clinical documentation and administrative burden on physicians’ professional lives.

Dr.Medkit aims to lighten the daily workload by reducing the manual tasks physicians carry out before, during, and after the visit. The goal is not only to save time but also to reduce the work that spills into after-hours and the pressure it creates.

Although research highlights the link between administrative burden and physician burnout, reducing this burden does not on its own mean burnout will be prevented⁷. Dr.Medkit’s goal is to help physicians spend their time not only on clinical practice but also on themselves.

Key takeaways

• During a visit, AI interprets clinical information, prepares documentation, and supports decision-making. Diagnosis and treatment decisions belong to the physician.
• Dr.Medkit brings the patient’s past information together before the visit; the history, clinical suggestions, and test orders are prepared in real time during the visit. At the end, all content is presented to the physician for approval in a single summary.
• Clinical suggestions are based on the patient’s current condition and verified medical sources, and are reproducible under the same clinical conditions.
• The agent-based structure lets clinical and administrative tasks be carried out with written or spoken commands. Patient data is stored in Türkiye, and visit audio is not kept permanently.
• Reducing documentation and administrative workload can help physicians save time and ease the work that spills into after-hours. The extent of this gain varies with clinical conditions.

References

1. 1Tierney AA, Gayre G, Hoberman B, et al. Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation. NEJM Catalyst Innovations in Care Delivery. 2024. DOI: 10.1056/CAT.23.0404.
2. 2World Health Organization. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. WHO. 2021. https://www.who.int/publications/i/item/9789240029200
3. 3Ratwani RM, Savage E, Will A, et al. A Usability and Safety Analysis of Electronic Health Records: A Multi-Center Study. Journal of the American Medical Informatics Association. 2018;25(9):1197–1201. DOI: 10.1093/jamia/ocy088.
4. 4Oğul İÜ, Soygazi F, Bostanoğlu BE. TurkMedNLI: a Turkish medical natural language inference dataset through large language model based translation. PeerJ Computer Science. 2025;11:e2662. DOI: 10.7717/peerj-cs.2662.
5. 56698 Sayılı Kişisel Verilerin Korunması Kanunu (Personal Data Protection Law No. 6698). Official Gazette (Resmî Gazete), No. 29677. 2016. https://www.mevzuat.gov.tr/mevzuat?MevzuatNo=6698&MevzuatTur=1&MevzuatTertip=5
6. 6Sinsky C, Colligan L, Li L, et al. Allocation of Physician Time in Ambulatory Practice: A Time and Motion Study in 4 Specialties. Annals of Internal Medicine. 2016. DOI: 10.7326/M16-0961.
7. 7Shanafelt TD, Dyrbye LN, Sinsky C, et al. Relationship Between Clerical Burden and Characteristics of the Electronic Environment With Physician Burnout and Professional Satisfaction. Mayo Clinic Proceedings. 2016. DOI: 10.1016/j.mayocp.2016.05.007.