The Architecture Trap: Why better algorithms won’t fix healthcare

India has arrived at a genuine inflection point. On 17 February, at the India AI Impact Summit in New Delhi, the Union Health Minister launched SAHI, the national strategy for artificial intelligence in healthcare, with its companion platform BODH, built by IIT Kanpur with the National Health Authority to evaluate AI models against real-world data without exposing the underlying datasets. India became the first country in the WHO South-East Asia Region to adopt such a strategy, and strategy and platform arrived the same day. The two are usually separated by years.

Yet beneath the momentum lies a structural risk. Look at where health AI is deployed in markets that started earlier. Bain’s 2025 survey found the leading uses to be documentation, clinical documentation improvement and coding. Rock Health reported AI-enabled companies taking 54 per cent of digital health funding last year, up from 37. And the 2025 CAQH Index found more than half of health plans using AI in administrative workflows against roughly a quarter of providers: payers automating the review of claims faster than providers can produce them.

Faster prior authorisation. Smarter revenue-cycle management. Ambient clinical scribing. Each has value. None changes the architecture of care. The money is going, with great sophistication, into making the existing transaction run more smoothly.

The Pattern That Should Give Us Pause

This mirrors the mistake that defined the digital-health era. Over the past decade, global healthcare committed sums on the order of $100 billion on the premise that layering technology onto existing infrastructure would drive reform. Costs rose, chronic disease grew, care became more fragmented. The technology was never the problem. The architecture within which it was deployed was.

India faces the same choice, but from a rare position of advantage. The digital foundations are laid, and they are national. UPI has made frictionless payment ordinary for hundreds of millions. The Ayushman Bharat Digital Mission has created more than 90 crore health accounts with over 100 crore records linked to them, and a consent architecture that could make individuals the owners of their own data. eSanjeevani has delivered more than 43 crore teleconsultations since 2019. These are the rails for a different model of care, capable of moving from ‘Find It, Fix It’ to ‘Predict It, Prevent It.’

Three Cities, One National Choice

Different cities are building different pieces of this future. In Hyderabad, Telangana’s Next-Gen Life Sciences Policy targets ₹2 lakh crore, some $25 billion, by the decade’s end; the state has already drawn close to ₹73,000 crore over two years. With the world’s largest pharma companies running capability centres beside deep IT strength, Hyderabad is the convergence of biology and software.

In Pune, the front line is access and prevention. The National Health Authority convened a Chintan Shivir there in April to accelerate ABDM and PM-JAY, opened by Maharashtra’s Minister of State for Public Health with a call for early awareness rather than late repair. A planned deployment of more than 2,000 AI-powered health ATMs across rural Maharashtra was announced in March: stations running over forty diagnostic tests in minutes and linking patients to doctors. Once deployed, that is early detection reaching the last village.

In Bengaluru, the story is intelligence itself: a dense biotechnology cluster beside the deepest AI talent pool in the country.

Each is building something essential. The question is whether the pieces are assembled into an architecture, or left as silos.

The AI That Worked. The Architecture That Failed.

To see what is at stake, follow one patient, a composite though nothing in her story is unusual.

Sunita is 58, lives in a district town in Maharashtra, and has had type 2 diabetes for eleven years. In March she stopped at a health ATM near the bus stand, because it was there and her daughter insisted. Six minutes, a finger-prick, a printed slip. Her HbA1c had drifted; her blood pressure was up on her reading two years earlier. The data went into her health account, longitudinal and linked.

In April, a model reads that record alongside her dispensing history and flags her: high probability of hospitalisation within thirty days. The prediction is correct. It reaches a physician forty kilometres away who has never met her, at the top of a queue of two hundred and forty alerts, on a morning when ninety patients wait outside.

He does not ignore it because he does not care. He ignores it because there is nothing behind the alert. No care manager to telephone her. No protocol assigning her to anyone. Titrating her medication would take a consultation, a repeat test and a follow-up call, none of it anybody’s responsibility and none of it generating a rupee. He closes it and sees the ninety patients outside, the only thing the system has asked of him.

On the second of May, Sunita is admitted in hyperglycaemic crisis. She stays nine days. The admission is documented, coded and settled under a package rate. Every institution is paid, correctly and on time, for treating an event a machine had predicted five weeks earlier, and four thousand rupees of coordinated attention would probably have prevented.

The AI worked. The architecture failed. And note what failed: not the model, not the data, not the doctor. The correct answer arrived in a system with no mechanism to act on it, and no one better off if it did.

India is not immune. Its health systems, public and private, are caught in the same trap: the logic rewards response over anticipation, and prevention remains underfunded. Sunita’s hospitalisation has a package rate. The conversation that would have prevented it has no code. The model remains ‘Find It, Fix It,’ whoever writes the cheque.

But India has something the West does not: it never laid six decades of billing-code bureaucracy. Its system is fragmented and entrepreneurial, less protected but far less locked in. That is an opening no other economy can replicate.

The Missing Layer

Hyderabad supplies the research and manufacturing base. Pune supplies access and prevention. Bengaluru supplies the intelligence. What India lacks is the orchestration layer connecting them: a platform turning fragmented providers into a network, a consent framework giving individuals ownership of their data, a payment design rewarding prevention, and standards that keep distributed care safe.

A rupee invested in prevention returns several times its value in avoided treatment, and AI that can flag metabolic risk months before onset carries real economic potential. What it lacks is a policy architecture that rewards it rather than the hospitalisation it would prevent. SAHI sets out the guardrails seriously, with thirty-two recommendations spanning governance, workforce training, data and procurement. What it does not carry, and what no strategy can settle alone, is a financing line: who is paid when the admission does not happen.

The Window Is Open, But Closing

Frameworks alone are not enough. The infrastructure being built today is not neutral. Once a health architecture settles its workflows, data pipelines and governance around a model, the cost of changing course becomes enormous. India has perhaps twenty-four to thirty-six months to direct capital toward the right architecture before the alternative locks in, and that window is closing faster than announcements suggest.

The technology is ready. The rails are laid. The strategy is published, and it is good. Hyderabad, Pune and Bengaluru hold the pieces. Nothing that matters now is a technical problem.

The West built its architecture before anyone knew better, and cannot now dismantle it without ruining the institutions that depend on it. India would build the same thing with the evidence in hand. That is the difference between an inherited mistake and a chosen one.

So the choice is narrower than it looks. We can spend the coming decade teaching machines to process the encounter: to code it, document it, adjudicate it, settle it faster than any system on earth. Or we can spend it building one in which the encounter need not happen. Both are within reach today. Only one will still be within reach in five years, because India will not get to attempt the second once it has finished building the first.

The author is a Boston-based scientist, serial entrepreneur, innovator, and investor