From Answering Questions to Doing the Work
When we first wrote about the assistant, it was mostly a very good reader: you attached a report, asked a question, and the answer arrived with the patient's context already in mind. That part still works. What changed is that it now acts. It fills in the record the way a colleague would, from one sentence in Greek or English.
There are now more than thirty actions behind it, covering patients, appointments and the clinical record. It is not a chatbot bolted onto a medical application — it is inside it.
What It Does Today
- Patients — finds a patient by name, ΑΜΚΑ or phone, creates and corrects the record, and adds allergies, medication and medical history.
- Appointments — checks which slots are genuinely free, then books, moves, cancels or completes one.
- Visits — writes a visit with chief complaint, notes, findings, height, weight and BMI, or appends findings to a visit that is already open, and records the vitals measured and the medication prescribed during it.
- Labs and examinations — records lab values against the clinic's own panels and tests, and corrects results and examination findings afterwards.
- The patient in one answer — allergies, current medication, recent visits, labs, examinations and history as a single summary, before the patient walks in.
Nothing Is Written Until You Confirm
This is the part that matters clinically. Reading happens immediately; writing is proposed:
- A question reads: finding the patient, listing appointments, summarising the record.
- Anything that changes data comes back as a review card in the chat showing exactly what will be written, and is applied only when you confirm it — as the user who confirmed it.
- Every write reaches the audit trail, with the user, the model and the time. The doctor stays the author of the record; the assistant is only the one typing.
It Reads the Paperwork
Attach a PDF or a photograph of a lab report to the chat and the assistant reads the values and files them against the clinic's own tests. A value the clinic has no test for is kept in the result rather than dropped, and the report the values were read from is attached to the result, so an entry written by the assistant can always be traced back to its source. The same holds for a visit note taken from a document.
If the report is already in the clinic's file library, pick it from there instead of uploading it a second time.
How It Works in Practice
Say you are reviewing a patient's recent blood panel and notice elevated liver enzymes. Instead of switching to an external search engine, you ask in the chat:
"Patient has elevated ALT and AST with normal bilirubin. She is on atorvastatin and metformin. What could explain this?"
The assistant already has the medication list, recent lab trends and medical history in front of it and answers with a structured analysis — statin-related hepatotoxicity, metabolic causes, suggested next steps — tailored to this specific patient. And if the conclusion is "stop the statin", it can record that decision in the visit once you confirm it.
Dictation Instead of Typing
The microphone in the chat and in the visit notes turns speech into text in the field you were typing in — useful when your hands are busy. It understands Greek and English.
Models, Credits and Costs
The clinic picks the model, and each user can pick their own; the strongest models from OpenAI and Anthropic are available, and switching does not change how the assistant behaves. Usage is metered rather than bundled:
- AI credits are not part of the subscription. The clinic holds a prepaid balance, and each request is charged for what it actually cost, so a quiet month costs less than a busy one.
- The balance is visible in the header of the chat itself, and the AI usage page lists every request with its tokens and its charge.
- Continuing a conversation reuses the patient context already sent to the model, so follow-up questions cost a fraction of the first one.
Clinical Questions Still Get Clinical Answers
Before answering a clinical question the assistant assembles the patient's context — history, allergies, medication, recent labs, vital signs — and answers from that record rather than from the open internet. It is a second pair of eyes on your patient, not a search engine.
On the Telephone, for Clinics That Connect One
Where a clinic connects its telephone system, the assistant answers calls as well: a caller gives their ΑΜΚΑ and can book, move or cancel their own appointment. The telephone channel is deliberately narrow — a caller is not a member of staff, so no allergies, medication, visits, examinations or lab results are reachable by phone, and an unknown caller cannot be registered.
Privacy, Isolation and Traceability
- Patient data is never used to train AI models.
- Each clinic's data stays isolated, and the assistant works under the same role-based permissions as the person using it: it cannot reach what that user may not open.
- Every interaction is logged, and the GDPR duties — export, anonymisation, erasure — are part of the system rather than an add-on.
What We Are Working On
We keep widening what the assistant can do, one verified action at a time. Every screen in the system that changes data must either be reachable by the assistant or be listed, deliberately, as out of its reach — and our test suite fails when nobody has decided. That discipline is what turns the assistant from a demonstration into something a clinic can trust with the real record.
Try It
The AI Medical Assistant is available to every Atomic Medical user. Contact us at [email protected] to have your clinic set up, and ask it to record something real.