Healthcare AI in India: What Actually Works (and What's Hype)

The healthcare AI that works in India today is not the AI you see in conference decks. It is not an autonomous diagnostic engine. It is a radiologist's second pair of eyes, a voice bot that triages patients in Hindi, a documentation assistant that saves a busy AIIMS resident forty minutes a day. That is not a disappointing answer. That is what real progress looks like.

What Genuinely Works Now

The clearest wins have come where AI augments a scarce specialist rather than replaces them.

Medical imaging assistance. The Ministry of Health puts Qure.ai's chest X-ray AI at 1,000+ sites serving 15 million patients a year, with TB detection improved by 30% [7]; the company's own materials claim several times that site count, which is the usual gap between a vendor page and a government count. The Ministry of Health reports a 27% reduction in adverse TB outcomes and a 12–16% increase in case detection from AI tools across its national programmes [7]. At Maha Kumbh Mela 2025, AI chest X-ray screening ran at a single site — the Central Hospital in Sector 2 — where it flagged 36.22% of X-rays as abnormal and presumptive TB signs in 12% of those [9]. Worth stating plainly, because this event gets described as mass screening: it was one hospital triaging the pilgrims who walked in, not a crowd of millions scanned in real time. On breast cancer, the strongest evidence for Niramai's AI thermal imaging is a peer-reviewed community study: 6,935 women screened across 25 municipal primary health centres in Bangalore, 5,248 meeting inclusion criteria, 1.18% flagged with abnormalities [5]. The company itself reports deployment across 200+ hospitals in 30 cities — a vendor figure, and a different kind of claim from a study.

One autonomous chest X-ray system, trained on over five million Indian X-rays and deployed across 17 Indian healthcare systems, reports up to 98% precision and over 95% recall across 75 pathologies [8]. Read that with the caveat it deserves: it is a preprint written by the system's own developers, not independent peer-reviewed validation. It is still a meaningful signal, because the deployment is real and the setting is one where radiologists are genuinely scarce — but a vendor's own numbers are a starting point for evaluation, not the end of one.

Telemedicine triage and routing. The eSanjeevani platform has recorded over 47 crore (470 million) consultations since its 2019 launch, and 28.2 crore of those consultations were supported by AI-powered Clinical Decision Support Systems between April 2023 and November 2025 [1]. Note which number is which: 28.2 crore is the AI-assisted subset, not the platform total — a distinction routinely lost in coverage of it. The model is simple: AI assists routing, triaging, and documentation; the doctor remains the decision-maker and the face of care. That separation of responsibility is what makes it work.

Cancer screening. Here the honest picture is thinner than the decks suggest, including an earlier version of this post. The most concrete Indian evidence for AI cervical screening is a 130-woman pilot in Pune: against colposcopy the AI point-of-care test reached 93.46% specificity but only 52.17% sensitivity, for 76.53% accuracy [6]. That is a useful triage tool where the alternative is no screening at all, and it is not the "above 90% accuracy" number that circulates unsourced. A tool that misses half the positives on a small pilot is a starting point, not a result.

Government digital infrastructure. The Ayushman Bharat Digital Mission (ABDM) has crossed 90 crore (900 million) ABHA digital health accounts as of May 2026 [2], and over 100 crore (1 billion) health records linked to them as of the same month [3] — a figure that doubled from 50 crore in fifteen months. This is the foundational plumbing that makes AI in healthcare possible at scale. Without it, every AI system is working from incomplete data.

AI-assisted centres of excellence. AIIMS Delhi, PGIMER Chandigarh, and AIIMS Rishikesh were designated Centres of Excellence for AI in healthcare in March 2025 to lead indigenous solution development [4].

The Hard Problems Unique to India

There is a reason healthcare AI has worked in imaging before primary care, and in cities before rural areas. The hard problems are real.

Data quality and fragmentation. Before ABDM, Indian healthcare had no unified patient record. Most clinical data was on paper, in regional languages, in inconsistent formats. AI trained on this data inherits its gaps. ABDM helps, but interoperability across private and public systems remains incomplete.

Doctor-patient trust and the last mile. The government puts India's doctor-population ratio at 1:834, better than the WHO norm of 1:1,000 — but that figure counts 5.65 lakh AYUSH practitioners and assumes 80% availability of registered allopathic doctors [10]. So it is not a count of available allopathic doctors, which is the number that decides whether a district hospital queue actually moves. Most shortage is in rural areas. But deploying AI where the need is highest — primary health centres, sub-district hospitals — runs into connectivity, device availability, and trust barriers. Farmers and rural families often trust a local ASHA worker more than an app.

Languages and literacy. India has 22 scheduled languages and hundreds of dialects. An AI system that works in English is useful in urban metros. An AI system that works in Bhili or Gondi serves a population that nobody is building for yet. Most current systems cover Hindi and a handful of southern languages.

Regulation is still forming. India does not yet have a dedicated medical AI regulatory framework. The CDSCO regulates medical devices, and AI diagnostic tools fall under that umbrella — but guidance specifically for AI-assisted diagnosis is still evolving. Builders are operating in a grey zone.

Bias and training distribution. Models trained primarily on Western or urban Indian populations may underperform on rural, underserved, or nutritionally distinct populations. A TB screening model is only as good as the training data that reflects the population you are screening.

What My Lens as an Assistive-AI Builder Adds

I built SmartON — an assistive AI for the visually impaired, now with 17,000+ users. The technical problems we solved there overlap significantly with healthcare AI: low-bandwidth inference, voice-first interfaces for users with low digital literacy, and building trust with a population that has been poorly served by technology before.

The lesson I keep returning to is this: the user interface is the product. In healthcare, a diagnosis that arrives in the wrong language, at the wrong reading level, or on a device the user does not understand is not a diagnosis delivered. Building AI for India means designing for constraints from the first commit, not adding vernacular support as a feature in version two.

That same principle applies to healthcare. MIRA, our multilingual voice AI router, handles code-switching between Gujarati, Hindi, and English in a single conversation. That architecture is directly applicable to patient triage in Gujarat — where patients switch languages mid-sentence and expecting them to do otherwise is unrealistic.

Where It Is Heading

The credible near-term trajectory for healthcare AI in India:

• More imaging AI at district and sub-district levels. The TB screening precedent is replicable across radiology shortage areas. Portable AI-assisted diagnostics that work offline or on intermittent connectivity are a genuine near-term opportunity.
• Voice-first documentation assistants. Reducing physician documentation burden is one of the clearest ROI cases for AI in any healthcare system. In India, where a doctor may see 80-100 patients a day, it is especially acute.
• Vernacular patient navigation. AI that helps patients understand their diagnosis, navigate the ABDM system, and find nearby facilities — in their own language — is a large, underserved problem.
• Preventive and chronic disease management. Diabetes, hypertension, and cardiovascular disease are India's primary chronic disease burden. AI-assisted monitoring, medication reminders, and dietary guidance at scale remain largely unbuilt.

The next decade of AI in India will likely be defined not by the most technically sophisticated models, but by the ones that work reliably for the populations most ignored by current technology.

Honest Cautions for Founders Entering This Space

• Regulation will tighten. AI diagnostic tools will face increasing CDSCO scrutiny. Build with documentation and clinical validation from day one, not as an afterthought before launch.
• Clinical partnerships are not optional. An AI health product built without clinician co-design will be rejected by the system it is trying to serve. Doctors are not obstacles. They are the implementation path.
• The DPDP Act applies. Health data is sensitive personal data under the DPDP Act 2023. Every data flow, every third-party API call that touches patient information, needs a lawful basis and proper handling. There are no carve-outs for healthcare.
• Distribution is the hard problem. Getting into AIIMS is not a distribution strategy. Getting to the CHC in a tier-3 city where the need is actually greatest — that is the problem most founders underestimate.