The upstream bottleneck for health AI capital in India

In a packed outpatient clinic in Punjab (or anywhere else in the subcontinent), the consultation ends the way most consultations do: in a rapid mix of local language and English medical terms. While the clinician has a clear clinical picture, a missing structural record stunts development of a downstream artificial intelligence (AI) model that can assess clinical staging, generate real-time evidence-based practices, flag potential treatment toxicities or develop a recurrence risk engine.

The binding constraint sits upstream, in capture and structuring. It is an unbuilt market with an identifiable buyer and a capability match already sitting in the Indian services industry.

The workflow constraint

Hospitals still lean on partial digitisation. Clinical data is generated at the point of care in the spoken exchange and a handwritten notes while admission and discharge details are digital. The structural bridge for long horizon retrieval is missing. As one detailed examination of the ecosystem has shown, the scarcity of accessible, reliable health data forces startups towards synthetic case sheets or overseas partnerships. Investors miss the capture layer completely. The result is a recurring misdiagnosis. Deployment failures are read as model failures and surprisingly the next funding round goes to the same layer again.

Why the data does not exist

The clinical diagnosis stack is precise: capture → structuring → longitudinal linkage → outcome adjudication. Indian players own fragments of the first two. No one owns linkage at scale or even attempt systematic outcome adjudication.

An identity and exchange layer is not an outcomes layer. The Ayushman Bharat Digital Mission has delivered +genuine public good—more than 90 crore ABHA accounts, registries for facilities and professionals, and interoperable exchange rails. It deserves credit for creating the digital identity and consent architecture. But ABHA enables sharing of whatever records exist. It does not generate structured clinical narratives from spoken consultations. This compounds the “chicken-and-egg” problem.

Why the obvious buyer is not building

Pharma has the most to gain from longitudinal real-world outcomes. Post-marketing surveillance in India functions largely as a compliance exercise. Phase IV produces filings and Periodic Safety Update Reports under the Pharmacovigilance Programme of India; it does not systematically produce evidence assets that can guide indication expansion, pricing, or value demonstration. The consequence, therefore worth naming as such: no outcomes infrastructure means no demonstrated value and hence no investment. The absence of the data is itself the evidence.

Where the capability match actually sits

The language layer is underpriced. Consultations happen in spoken mother tongues but surprisingly the formal record is expected in English. The loss in translation is where the data problem originates. Voice transcription in Indic languages is a genuine capability match for the Indian IT services industry as services pivot and not a deeptech moonshot.

Open work on Indic speech models has lowered the entry cost enough that this is a build decision, not a research bet. Production systems for medical transcription tuned for code-switching, regional accents, and clinical terminology have demonstrated that the technical barrier is lower than the organisational one. Eka Care’s Parrotlet models and similar efforts by Sarvam and others have revealed that the constraint is linguistic and procedural, and India happens to hold the relevant capability.

The allocation guidance

Returns will accrue to whoever owns structured longitudinal capture, not to whoever ships the best model on top of nothing. Sovereign compute is arriving at scale and will sit under-utilised for health applications precisely because there is nothing structured to run on it.

The claim is straightforward: until capital and capability shift upstream to capture and structuring, model-layer investments will continue to under-deliver relative to their valuation, while the services firms that solve the language-to-structure problem will quietly build the actual asset.

(The author is a practising Radiation Oncologist and founder of Agravaani Research. Views are personal.)