AI prescription software: what the doctor must check Auto-filled medicine details, the rules that still apply, and a review-before-save checklist.

When AI prescription software fills in medicine details or drafts a prescription, the doctor must check the same things as on paper: the generic name, the composition and strength, the form, the dose and quantity, whether the item is a prescription-only drug, and whether the printed prescription carries their own name and registration number. The software is a typist with a good memory. It can pull the wrong salt, the wrong strength or the wrong maker, and it cannot be held responsible for any of it.

Who answers for an AI-filled prescription

The registered medical practitioner does, under rules written for a handwritten pad. Clause 1.4.1 of the Indian Medical Council (Professional Conduct, Etiquette and Ethics) Regulations, 2002 says every physician "shall display the registration number accorded to him by the State Medical Council / Medical Council of India in his clinic and in all his prescriptions, certificates, money receipts given to his patients". Clause 1.5, as substituted by the Gazette notification of 8 October 2016, reads: "Every physician should prescribe drugs with generic names legibly and preferably in capital letters and he/she shall ensure that there is a rational prescription and use of drugs". The NMC repeated that clause in its letter of 15 December 2025 to medical colleges and added that "Compliance of these directions is mandatory". Our post on the generic name prescription rule covers the clause and the court order behind that letter.

The drug side of the law points at the same person. Rule 65(9)(a) of the Drugs and Cosmetics Rules, 1945 says substances in Schedule H and Schedule H1 "shall not be sold by retail except on and in accordance with the prescription of a Registered Medical Practitioner". Rule 65(10) defines that prescription. It must "be in writing and be signed by the person giving it with his usual signature and be dated by him", must "specify the name and address of the person for whose treatment it is given", and must "indicate the total amount of the medicine to be supplied and the dose to be taken". The glossary entries on Schedule H and Schedule H1 carry the rule text. Rule 97(1)(b) is the other half: a Schedule H pack is labelled with the symbol Rx and the words "Schedule H drug- Warning: To be sold by retail on the prescription of a Registered Medical Practitioner only". The pack label, not the software, is the source of truth for whether an item is prescription-only.

Neither document mentions software. It does not need to. The regulations bind the physician, the Rules bind the person giving the prescription and the chemist supplying against it, and a field that a program filled in is still a line the doctor signed.

What auto-filled medicine fields get wrong

A model that fills medicine details from a name works from patterns in text, so its mistakes are those of a hurried reader of pack labels. Six fields fail most often.

The last row matters most. Whether a medicine is prescription-only decides what the chemist may hand over, and it is answered by the Schedule and the label. A model guessing from the molecule name will be right often enough to feel reliable and wrong often enough to matter.

The review-before-save checklist

Run these nine checks on every AI-filled card or drafted line before it is saved to the clinic's medicine list or a patient's prescription. Most take a few seconds with the pack in hand.

1. Generic name: present, spelt the way you want it printed, in capitals (clause 1.5).
2. Brand name: if shown, it is the brand you mean and is marked as a brand, not filed as the generic.
3. Composition: every component of a combination is listed with its strength, matching the pack.
4. Strength and unit: milligram, millilitre or percentage figures match the formulation you prescribe, not a more common one.
5. Type: tablet, capsule, syrup, drops, cream or injection, matching the pack and the patient.
6. Company: the maker on the card is the maker on the pack, or the field is cleared.
7. Prescription status: Rx or OTC read from the label under rule 97, not accepted from the software.
8. Dose and quantity on the prescription itself: rule 65(10)(c) requires the total amount to be supplied and the dose to be taken. No lookup can supply those.
9. Your own details: the print carries your name, registration number (clause 1.4.1), the patient's name and address, the date and your signature (rule 65(10)).

If a card fails any of the first seven, edit it or reject it. Do not accept it to fix later; the next doctor to pick that medicine from the list will trust it. For what the printed document must carry, see what must a prescription contain.

Why nothing should be saved until you accept it

The ICMR's Ethical Guidelines for Application of Artificial Intelligence in Biomedical Research and Healthcare (2023) make the design point. They say that "Just like any other diagnostic tool, AI-based solutions themselves cannot be held accountable for its decisions and judgments", and that "the responsibility of optimal utilization of the technology is on the health professional using AI-based solutions for delivering healthcare". On oversight, the 'Human In The Loop' concept "places human beings in a supervisory role and is more relevant for healthcare purposes", and "Human oversight and manual override mechanisms must be made compulsory for all AI technologies that have the capability to cause severe or bodily injuries".

A tool that writes its guess straight into the medicine list, or onto the prescription, leaves the doctor to catch errors on a screen that already looks finished. A tool that holds every suggestion as a draft until a named person accepts, edits or rejects it keeps the supervisory role where the guidelines and the regulations put it. The test for any product is simple: close the screen without touching a suggestion and see whether anything changed in the record. The right answer is nothing.

A worked example in Vitels

Vitels does not draft prescriptions, has no clinical decision support, and never sets the prescription status of a medicine. Its one AI helper on the medicine side is adding medicines with AI: a doctor or clinic admin types up to 20 medicine names, or records a voice note of up to 60 seconds, and gets one card per name. A card is either already in the clinic's list, a draft "In review" with composition, type, company and a one-line description, or "Not found". The person can accept, reject, edit, skip or add it manually. Accept saves through the same path as a medicine typed by hand. The Rx or OTC status is left for the person to set from the pack label. Nothing is saved until someone accepts it; cards left unreviewed when the drawer closes wait under Requested Medicines as "Approval pending" for a doctor or admin.

Run the checklist on one card. The doctor types a name, gets a draft with a composition and a company, and checks it against the pack on the shelf. If the lookup returned the tablet strength and the clinic uses the syrup, she edits the card. She reads the label, finds the Rx symbol and the Schedule H warning, and sets the status herself. Then she accepts. Only the medicine's details enter the list; any patient's prescription is still written by her, line by line, on the pad that prints her name and registration number.