What an AI assistant can and cannot do in health

What AI in health actually helps with

With organising: gathering information in one place, finding what you need quickly and not forgetting what you have already planned.

Medical papers pile up over time: test results from different labs, prescriptions, discharge summaries, notes after an appointment. Finding the right line in that pile is an ordinary test of attention, and part of it can be handed to software.

According to the Cleveland Clinic, virtual assistants and chatbots already take on simple tasks: pulling up a medical history faster, helping to book a follow-up appointment, sending medication reminders.¹

Among the uses of large language models in healthcare, the WHO also lists clerical work: for example, recording and summarising visits in the electronic health record.²

All of this is work with what is already known: with the doctor’s prescription, with the date of a test, with the text of a document. There are no new medical conclusions in it.

Why decisions about health stay with the doctor

Because a model can be wrong in a confident tone, while a doctor sees the whole picture: the examination, the medical history, the other medicines.

According to the WHO, large language models carry a documented risk of producing false, inaccurate or incomplete statements – and that can do harm if decisions about health rest on them.²

The WHO separately warns about over-reliance on automation: a software error is easy to miss, and a hard decision is easy to hand over unnoticed, although it should be made by a person.²

The Cleveland Clinic reminds us that such systems were not built or approved for medical use, so health questions are worth taking to a doctor.¹

These are the questions worth asking a doctor or a pharmacist rather than a program:

• Is the value in the test within the normal range?
• Is the treatment working?
• Can these medicines be taken together?
• What dose should I take and can it be changed?

What Nomi will do

Nomi will extract data from your words and documents and find your entries. Everything extracted is shown to you for confirmation first.

1. Builds the profile in the first chat

   You tell it about yourself however you like; Nomi asks clarifying questions one at a time and shows a “Here is what I understood” screen before saving anything.

2. Turns a prescription into a course

   From a prescription you retell in your own words Nomi builds a course with reminders and shows a confirmation screen. If the data is ambiguous it asks again rather than choosing for you.

3. Reads a test result or a prescription

   The date, the lab, the doctor, the values – exactly as printed on the form. You check, correct and confirm.

4. Finds the right entry

   To “which tests did I have in March” Nomi answers from your documents, comments and notes on how you felt – and every answer links to its source.

5. Reminds you about a repeat test

   On the day you name yourself: Nomi does not advise when to have a test.

Where Nomi’s limits will be in MedKeep

Nomi will not diagnose, assess results, advise on doses or change treatment – neither in the beta nor in later versions.

• It will not say whether a value is normal, even if the reference range is printed on the form.
• It will not suggest or correct a dose.
• To “is this normal?” or “did the course help?” it answers that this is a question for a doctor or a pharmacist, and offers a useful step: note the question for the appointment, save the doctor’s contact, set a reminder about a test.
• It will not remember the conversation beyond the session and will not see other people’s data.
• It will only work with your consent to send text and documents to the model provider. Withdraw the consent and Nomi and recognition switch off, while the rest of the app keeps working.

A refusal will not be a dead end. If you ask for an assessment, Nomi will quote your entries as they are and help you prepare for the conversation with the doctor: the question is saved for the next appointment, and the right test is easy to find right there in the consulting room.

A summary for the doctor and a full assistant that answers from your data are planned for V1; they will not be in the beta. The rule for them is the same: no assessing results, no diagnosis, no changing treatment.