16 November 2026 · Medows
India's AI Scribe Is Now Public Infrastructure
India's health ministry demoed a national ambient AI scribe. The doctor approves every record before it enters the patient's national file.
A district hospital OPD, mid morning. The queue is forty deep. The consultation runs in two languages, sometimes in the same sentence. The doctor listens, examines, decides, and then turns to the register, because the record is what the system will actually see later. The deciding took four minutes. The writing takes two more, and it happens forty times before lunch.
That is the workload India's health ministry just pointed an ambient AI scribe at, and the way it did it is worth reading closely.
What was actually announced
The National Health Authority (NHA) convened a two day national review meeting on AB PM-JAY and the Ayushman Bharat Digital Mission in Kolkata on 5 and 6 November 2026. Among the demonstrations was NHA's Ambient AI-enabled Voice Scribe, described in the government's own release as "a technology that converts doctor-patient conversations into structured clinical records in multiple Indian languages."
Then, the sentence that matters: "The records are reviewed and approved by doctors before being linked to ABHA."
A week later, trade coverage reported that NHA had selected Bengaluru based EkaCare (Orbi Health Private Limited) as technology partner, following an Expression of Interest issued on 2 September 2026 for open source ambient AI voice-to-text solutions for digital health record creation under ABDM. Per that reporting, the solution is designed to run inside NHA's air-gapped environment on NHA infrastructure, notes come out in FHIR using speciality templates, and doctors sign in through the Healthcare Professionals Registry.
The vendor's own claims sit on top of that, and should be read as vendor claims: support for more than 15 Indian languages including code-mixed speech, an open-weight speech model (Parrotlet-A 2.5 Pro) with weights published on Hugging Face and AIKosh, a reported lowest error rate across five benchmarked languages on 6,550 real Indian consultations, and more than 1,000 doctors already using its private-practice product. Rollout dates, facility counts and commercial terms were not disclosed.
The approval gate is the design decision
Most scribe deployments treat review as a courtesy. The note lands in the draft folder, the clinician signs it at 9 p.m. along with nineteen others, and the signature means "I did not have time to find a problem."
Writing the gate into the architecture is different. Here the AI output does not become part of the patient's national record until a doctor, authenticated as that doctor, approves it. The model drafts. The record waits. That is the correct order, and it is the thing we keep arguing for: AI that proposes, a human who disposes, and a trail that shows which was which.
It is also the thing that will be hardest to hold. An approval gate under forty-patient load degrades into a tap. The question nobody answered in Kolkata is what the gate looks like on the eleventh hour of a shift.
They shipped a way to test it, too
The same meeting included a presentation of BODH, the Benchmarking Open Data Platform for Health AI, from Prof. Phaneendra Yalavarthy of TANUH, IISc Bengaluru. Per the release, it lets states evaluate health AI models on Indian datasets, and states were encouraged "to incorporate BODH-based benchmarking into their healthcare AI procurement processes."
Benchmarking written into procurement is rarer than it sounds. It means a state can ask a vendor to perform on data that looks like its own patients, before signing, rather than reading a slide about accuracy measured somewhere else. Transcription that does well on American clinic audio tells a Tamil district hospital almost nothing.
The part a scribe cannot reach
A structured note is one artifact from a shift that produces many. The scribe handles the consultation. It does not handle the ward round, the handover at change of shift, the list of twenty-eight patients with four things pending each, or the question of what happened to the sodium that was flagged yesterday. Documentation was never the whole burden. It was the visible part.
So this is real progress, narrowly. India is treating ambient capture as public plumbing (standards, identity, an approval gate, an open benchmark) rather than as a product each hospital buys separately. That is the right layer for transcription. The rest of the shift is still waiting for a workspace.
Sources
- PIB, Ministry of Health and Family Welfare: Union Health Ministry Reviews Progress of AB PM-JAY and ABDM with States and UTs (6 November 2026)
- Elets CIO: NHA AI Voice Scribe: EkaCare Selected (11 November 2026)
- Elets eHealth: NHA Shows Sandbox 2.0 AI for Faster ABDM Onboarding (7 November 2026)
- Outlook Business (syndicated release): EkaCare Selected as a Technology Partner for NHA's Ambient AI-enabled Voice Scribe Solution for Doctors (13 November 2026)
Medows is a clinical AI workspace for the doctor on rounds. Learn more or write to alapan@medows.ai / alapanx@gmail.com.