Two extreme attitudes appear when AI in healthcare is discussed: the fear that it "will replace physicians," or the expectation that it "will solve every problem." The reality is outside both, and calmer. Today the most reliable and valuable use of AI in a clinic is not to make diagnoses but to reduce the physician's writing and editing load, while the decision stays with the physician. This article explains what an AI-supported assistant can do in a clinic, what it should not do, and which questions to ask when you evaluate a product.
The line between assistant and decision maker
However impressive an AI feature is, it cannot take over clinical responsibility. This line is ethical, legal and practical at once:
- Assistant: structures what the physician says, reminds of missing fields, summarizes a long history. The output it produces is a draft.
- Decision maker: makes a diagnosis, chooses a treatment, publishes a result. This role belongs to the physician.
In HekimBis the AI assistant is in the first category. No output is finalized in the record without physician approval; the AI does not diagnose and does not decide treatment. This principle is a built-in constraint of the product, not a setting.
Safe areas of use in a clinic
1. A structured note draft from dictation
The physician speaks during the examination; the system turns the audio into text and places it in the specialty template: complaint, findings, assessment, plan. This reduces the time the physician spends at the keyboard and lets them look at the patient more. The draft is edited and approved by the physician. Dictation errors (especially drug names, doses and numeric values) can slip through, so the physician must read and verify the result.
Source and trail stored
The source text, model output, the physician's changes and the approval trail are kept together.
- The process is complete here.
All steps
- 01 Dictation (Physician) After the consultation the physician speaks; the audio is transcribed against a specialty-appropriate template.
- 02 Structured draft (Assistant) The speech becomes a structured note draft for the specialty and visit type. The draft is marked as a "suggestion".
- 03 Missing-field and contradiction warning (Assistant) Empty required fields, contradictory statements and follow-up jobs are flagged as warnings.
- 04 Physician edits (Physician) The physician corrects the draft, resolves the warnings or declines the suggestion.
- 05 Approval and signature (Physician) The physician approves and signs; the record is finalized only now.
- 06 Source and trail stored (Record) The source text, model output, the physician's changes and the approval trail are kept together.
Alternative path from 04 Physician edits; the process ends on this path.
- 04a Rejected: not finalized (Physician) The physician does not accept the draft; it never enters the clinical record and the physician writes the note themselves.
Alternative path from 01 Dictation; it returns to the main flow at 04 Physician edits.
- 01a AI off: continue manually (Record) If the assistant is off, the physician writes the note directly; appointments, signatures and other flows are unaffected.
06 Source and trail stored
2. Missing-field and contradiction warnings
Mandatory fields left blank in a template, or inputs that contradict one another (for example a drug draft incompatible with an allergy while "no allergy" is written), can be flagged to the physician. This is not an "error finder" but an "attention caller"; the decision is still the physician's.
3. Summary of a long history
Summarizing a patient file built up over years before an examination saves the physician time. A summary should be presented with a basis that shows its source; the physician should be able to click and see which record a statement rests on. A summary without a basis is an untrustworthy summary.
4. Administrative paperwork
For tasks such as a patient information letter, a referral letter draft or a report draft, AI can prepare the first draft. Signature and responsibility still rest with the physician.
5. Operational support
In operational uses such as forecasting appointment density, flagging appointments with a high likelihood of no-show, or anticipating stock needs, the risk is lower than clinical use; even so, results should be presented as decision support.
What AI should not do
- Make an autonomous diagnosis or treatment decision.
- Write to the record without physician approval. An output awaiting approval is a draft, not a record.
- Deliver a result directly to the patient. Results reaching the patient pass through physician review and a publishing decision.
- Send patient data to general-purpose model training. Patient data must not enter a training set outside the provider's control.
- Present certainty without evidence. Showing an uncertain output as certain pulls the physician's attention the wrong way.
Eight questions to ask when evaluating
When evaluating a clinic software with AI-supported features, ask:
- Are outputs drafts, or written straight to the record? Is physician approval mandatory?
- Are the source, the output and the physician's changes stored?
- When AI is switched off, does the flow continue unchanged by hand?
- In which country and with which subprocessor is patient data processed?
- Is patient data used in model training?
- How long is the voice recording kept, and when is it deleted?
- How is the patient informed and, if needed, how is consent managed?
- Are usage quota and cost included in the plan, or extra?
The first three are the most critical. A system that works completely with AI off shows that the feature is an assistant, not a dependency. In HekimBis, AI is optional; when it is off, the flow continues the same way manually.
An example day: where does the assistant step in?
To make the abstract principles concrete, look at a fictional day of examinations. The physician opens the morning agenda; the first patient is here for a checkup of a chronic condition. Before the examination, the system summarizes the patient's last three records with sources: the latest test values, medications in use, the plan from the previous checkup. The physician clicks the source of a drug in the summary, goes to the record and verifies it.
During the examination the physician talks with the patient and leaves short voice notes. When the examination ends, the system offers a note draft placed in the specialty template. The "blood pressure" field in the draft is empty; the assistant reminds the physician of this missing field. The physician enters the value, notices that a drug name in the draft is misspelled, corrects it, then approves and signs the note. The system stores the voice recording, the draft and the physician's changes as a trace.
In this scenario the AI did three things: it summarized, structured and reminded. At no point did it decide, write to the record on its own or send anything to the patient. The physician kept control at every step. A good assistant is this ordinary, because being ordinary is the condition of being reliable.
Data protection and consent
Using AI means processing personal data and concerns special-category data. So three topics must be clear: the purpose and legal basis of processing, who the data is shared with, and how the patient is informed. If a voice recording is taken, this information should be given to the patient before the examination. The privacy notice and explicit consent are separate records; which one is required depends on the legal basis of the operation and should be settled with an adviser. See our KVKK and health data guide for the general framework.
Also, access rights to AI outputs and voice recordings must follow the same role-based rules as the patient record. If a front-desk employee can hear a physician's dictation, that is a security gap.
Caution in imaging and diagnostics
AI-assisted tools in radiology and imaging are under discussion. Extra caution is needed here: algorithms aimed at diagnosis may fall under medical device regulation and have their own evidence and registration requirements. If a clinic software says it offers an image-reading algorithm, ask what approval and validation it rests on. HekimBis presents images in a viewer for clinical and reference use and does not offer a diagnosing algorithm. For the imaging workflow see our DICOM and PACS article.
Trust of physicians and patients
For physicians, AI gains trust only when it makes its errors visible. Why was a draft produced this way, which record does it rest on, which field is uncertain? Without this transparency the physician must re-check every output and the time saving disappears. For patients, the source of trust is knowing the process is controlled: whether AI is used, where the records are kept and that the physician makes the final decision should be told openly. A deceptive or hidden use can damage institutional trust even in a single incident.
Cost and quota
AI features incur cost according to use, so usage quotas and pricing should be transparent on the product side. For a small practice a limited quota is usually enough; a multi-physician clinic needs organization-level quota and policy management. You can check which plan has which quota on the pricing page. Not stopping the flow when the quota is used up, and keeping the manual route open, is also a basic requirement.
Mobile and in-day use
The most concrete benefit of AI for physicians shows up in small time gains during the day: leaving a voice note from the phone in the corridor, having that note attached to the patient as a draft, and then having the physician edit and approve it in a free moment. In the mobile app flow the voice note draft is designed for this purpose. The same rule applies on mobile: the record is not finalized until the draft is approved.
Setting expectations correctly
When rolling out AI-assisted features, set expectations in the team correctly:
- In the first weeks physicians correct drafts more; the corrections fall as the system adapts to the physician's style and templates.
- The gain will not be equal for every physician; a physician who talks a lot gains more, one who writes short notes gains less.
- Measure use: note-writing time per physician, correction rate and number of unclosed drafts.
- Record the errors you encounter and report them to the provider.
Conclusion
AI in a clinic is not a decision maker that replaces the physician but an assistant that reduces the writing and editing load and stays bound to physician approval. The foundation of safe use is draft status, physician approval, source and change trail, optionality and data protection. When evaluating products, give no place in the clinical process to any feature that does not follow these principles, however impressive it looks. For HekimBis's approach see the AI assistant page, and for the structure that shapes the physician's screen by specialty, the doctor software page.
