AI medical scribes in India: do they work for Hindi-English consultations?

An AI medical scribe, also called an ambient scribe, listens to the conversation between doctor and patient and writes the clinical note for the doctor to check. Several are now offered to Indian doctors. The first question doctors ask is the right one: will it understand a real Indian OPD, where the patient says “सुबह से चक्कर आ रहे हैं” (I’ve been dizzy since morning) and her son adds the sugar reading in English?

What an ambient AI scribe does, step by step

1. It asks first. The patient is told that software will listen and why, and can say no.
2. It listens, through a phone or a microphone on the desk.
3. It turns speech into text, in the language spoken, or translated.
4. It works out who said what, separating the doctor’s voice from the patient’s, a relative’s and a nurse’s. This step is called speaker separation, or diarisation.
5. It drafts the note and the prescription: complaints, history, examination, diagnosis and plan, and the medicines the doctor prescribed.
6. It waits for the doctor, who reads, corrects and approves the draft. Only then should anything be saved, printed or sent.
7. It keeps, or deletes, what it heard. Products differ on whether they store the audio and how long they keep the transcript.

Steps 3 to 5 are where Indian consultations get hard.

Why Indian consultations are hard for a scribe

They are short. A systematic review of 67 countries found Indian primary-care consultations measured at 1.5 to 2.3 minutes, and about 2 minutes in the most recent study, though it rated the Indian studies only fair to poor in quality. In the medicine OPD of a New Delhi public teaching hospital, the average consultation took 284.6 seconds, under five minutes. There is little repetition to recover from, so every missed sentence matters.

They mix languages mid-sentence. Patients describe symptoms in Hindi and give readings, medicine names and test names in English. Here is an example written for this guide, with made-up people and numbers:

A good note records that the son gave the reading and the history of stopping the medicine, and that the doctor restarted metformin 500 mg twice a day after food. A weak one puts every word in the patient’s mouth, or drops “बंद कर दी थी” (had stopped) and records a patient who was taking her medicine all along.

Several people talk. A relative answers for the patient, a nurse reads out the BP, the next patient looks round the door. The scribe has to know whose words are whose: when the son says “मुझे भी” (me too), he is not the patient.

The room is noisy. Fans, a busy corridor, phones ringing and token announcements all reach the microphone.

Medicines are named by brand. Doctors and patients say “Dolo”, “Pan-D” or “Glycomet”, often clipped, and many brands sound alike. The scribe has to map the brand to the right medicine and strength, or ask.

There is no Indian yardstick yet. A September 2026 preprint, not yet peer reviewed, found no publicly available, large-scale, real-world benchmark for testing ambient scribes in India, and argued that Indian consultations are likely to multiply transcription and note errors. Until such a benchmark exists, each hospital has to test for itself.

The errors to test for

What the evidence says so far

The most rigorous evidence so far is from the US. In a randomised trial at UCLA Health, published in NEJM AI in November 2025, 238 physicians in 14 specialties were assigned to one of two commercial ambient scribes or to their usual way of writing notes, across about 72,000 patient encounters between November 2024 and January 2025.

• Time: with one of the two scribes, average time per note fell by an estimated 41 seconds, from 4 minutes 30 seconds to 3 minutes 49 seconds, against 18 seconds (4:22 to 4:04) in the control group: a reduction 9.5% larger than usual care. The other scribe showed a smaller decrease that was not statistically significant.
• Burnout: doctors using either scribe saw burnout scores improve by about 7% compared with the control group.
• Errors: physicians reported that the notes “occasionally” contained clinically significant inaccuracies, most often omissions or pronoun errors, and one mild patient-safety event was reported.
• Patients: fewer than 10% declined.

Those were consultations in the US. In India the need is plain: in a 2025 FAIMA survey with over 2,000 responses from medical students, teachers and professors across 28 states and union territories, 73.9% reported excessive clerical workload. But we know of no published trial of an ambient scribe in Indian OPDs, so treat any time saving as something to measure, not to assume.

Consent and patient data: what to ask

The patient’s questions are simple: is something listening, what happens to it, and can I say no? Answer them before the consultation starts.

India’s DPDP Rules, notified on 14 November 2025, say consent must be clear, informed and easy to understand, and that people can withdraw it at any time. The duties they place on hospitals, including notice and consent, apply from 13 May 2027. ICMR’s 2023 ethical guidelines on AI in healthcare are more specific: patients should be told about the AI and its risks, must be free to choose or reject it, and should face no pressure from the healthcare professional to accept it.

Here is one way to ask, for a scribe that keeps no audio and deletes the transcript after 30 days. Change the details to match yours:

Before you sign up with a vendor, ask:

• Is the audio stored? Where, and for how long?
• How long is the transcript kept, and when is it deleted?
• Where is the data hosted and processed?
• Who at the vendor can see consultations, and is that access logged?
• Is anything used to train or improve the product?
• Can you erase one patient’s consultation on request?

A one-week test you can run

You don’t need a research study. You need about 30 consultations, a sheet with one row per consultation and a column for each error in the table above, plus one for the minutes spent correcting.

Day 1: set up

Agree the consent script with your team, decide which OPD sessions to include, and practise once with a colleague playing the patient.

Days 2 to 4: run it alongside your own note

With consent, run the scribe on about ten consultations a day. Write your own note as usual, then compare it with the draft line by line. Count every omission, added medicine, wrong dose and wrong detail, and time how long the draft takes to fix.

Day 5: try the hard cases

Choose consultations that will stretch it: a relative doing the talking, an elderly patient with a soft voice, a child whose dose depends on weight, the noisiest hour of the morning, and a patient who switches between Hindi, English and a local language.

Days 6 and 7: add up and decide

Set your bar in advance. Any medicine you never said, or any wrong dose, in a draft is a serious finding, because a tired doctor at the end of a long OPD may approve it. Ask the vendor how the product flags such lines. And if fixing a draft takes longer than writing the note yourself, the scribe is not saving you time, whatever the demo showed.

Where DrMate’s doctor’s room fits

DrMate’s doctor’s room is built and ships with DrMate, but it is not yet in daily use at a hospital, so we have no real-world results to report: run the test above on it as you would on any other. It listens only after the patient’s recorded yes, and respects a no. It listens in 11 Indian languages (Hindi, English, Bengali, Marathi, Gujarati, Punjabi, Tamil, Telugu, Kannada, Malayalam and Odia), or detects the language itself. It keeps the doctor’s words apart from the patient’s, a relative’s and a nurse’s.

In the prescription draft, medicines you never said are removed or marked “AI suggestion”, and nothing is final until you approve it. No audio is kept, and the transcript is deleted 30 days after the visit; the approved note and prescription stay with the hospital. If you would rather type, DrMate Rx reads the shorthand you already write. Either way, the printout should meet the prescription format India’s rules require.