When it comes to technology for advancements in health, India has often displayed ambition and expertise. Over the last few years, a variety of health-tech solutions have been piloted, including a portable DNA testing device that identifies infections and drug resistance, an AI assistant in delivery rooms, and algorithm-based detection of tuberculosis in chest X-rays. These are more than ideas developed in urban innovation hubs; they have been deployed or adopted in real world settings.
Despite their technical brilliance, many tech solutions often share a common fate. They remain pilots that survive only as long as a grant cycle, serving facilities or geographies selected for convenience or visibility, generating data and metrics that satisfy reviewers, and often lacking institutional ownership needed to endure. Eventually, they stall without reaching their potential, or the populations that need them most. The global health community has come to describe this proliferation of pilots that fail to scale as ‘pilotitis’.
Pilotitis is often diagnosed as a result of ecosystem failure: a regulatory framework unfit for innovation, funding architectures that reward showcase over impact or scale. That diagnosis is correct, but incomplete. In India, pilotitis can also be a failure of innovation design. Technologies that arrive without real user input, treat frontline health workers as implementation levers rather than co-designers, and are validated in controlled conditions that bear little resemblance to the health systems they are eventually asked to inhabit are more likely to fail because they were never designed in consultation with ecosystem players in the first place.
The accountability here is shared by the state that sets the framework and the innovators who build the tools.
The ecosystem limitations are real
1. Regulations are tough to get through
India’s regulatory structure for health technology was largely designed for pharmaceuticals: physical products with fixed dosages, clinical trials, and defined post-market surveillance requirements. Software-based solutions do not fit this mould. A diagnostic app that uses machine learning to flag tuberculosis from chest X-rays has no dosage to standardise and no fixed formulation to approve. Its accuracy can shift as it is used on new devices, in new regions, or with changing datasets. This is something the current approval pathway is not built to monitor or re-certify over time.
Since approval is treated as a one-time event, innovators face high compliance costs.
This mismatch has direct consequences for sustainability and scale. Since approval is treated as a one-time event, innovators face high compliance costs simply to determine which rules apply, with no efficient pathway for monitoring, updating, or re-approving or evolving software. Without a clear, predictable route to scale, investors and government partners hesitate to commit long-term. The result is that even technically sound solutions stay trapped in pilot mode: scaling means stepping into a regulatory grey zone rather than a moving through defined and navigable process.
2. AI innovations are unreliable and have no governance
When a drug is administered at the prescribed dose, its formulation and intended mode of action remain relatively fixed. AI models work differently. They make probabilistic judgments that can change with the data they are trained on, the populations they encounter, and the quality of inputs available in practice. A model trained on high-quality chest X-rays from one set of hospitals, for instance, may perform differently when deployed on different devices or in district hospitals where image quality and patient populations vary.
This raises a fundamental question: who is liable when a TB-detection model misclassifies a scan? The innovator? The health department or facility that procured it? The ASHA worker who acted promptly on its result? Indian law does not yet provide a clear, AI-specific framework for allocating responsibility in such cases. There is no established mechanism requiring developers or users to disclose performance degradation in a model, and no pathway for patients to seek redress when they are harmed by an algorithmic error. While the government has recently released non-binding ethical frameworks, including the ICMR guidelines, MeitY’s India AI Governance Guidelines and the MoHFW’s SAHI, the country does not yet have a binding law or a specific AI Liability Act to address and resolve the accountability gap.
3. Funding landscape prioritises quantity over quality
The dominant funding architecture acts as another impediment. The enduring models that fund health-tech operate within 12-to-36 month grant cycles and review frameworks calibrated for quantitative indicators, with obligatory requirements for a few qualitative responses.
Genuine health-tech development involves testing an idea, identifying what doesn’t work, adjusting, and testing again. That process can often look like failure before it produces something that works. But short funding cycles and metrics-driven reviews reward steady, visible outputs. As a result, funders end up backing solutions that can show numbers quickly rather than those that take time to mature, adapt, and work reliably in real-world conditions. Unsurprisingly, innovators have to optimise for the metrics that secure the next funding window, not the outcomes that would justify population-level deployment.
These structural limitations demand structural responses: adaptive regulatory pathways, longer-horizon catalytic funding, and liability frameworks that apportion accountability clearly across the developer-deployer-implementer chain.
Innovation must learn too
It can be safely said that health-tech in India has systematically underinvested in two critical dimensions of innovation: community/user embeddedness and testing whether a technology fits the public health system. For any technology to be accepted and scaled, it needs to perform well on both.
Many ASHA workers now juggle seven or more apps, sometimes entering the same data twice because the systems don’t talk to each other.
Consider India’s own ASHA digital-health rollout. It introduced solid apps meant to streamline recordkeeping and support nearly a million community health workers. In practice, however, many ASHA workers now juggle seven or more apps, alongside numerous WhatsApp groups, sometimes entering the same data twice because the systems don’t talk to each other. A digital transformation meant to ease their work has instead added an unpaid second shift.
The literature on implementation science is unambiguous on this: interventions that work in a controlled pilot rarely translate seamlessly into real-world settings. Fidelity drops and adaptations become necessary. The gap between design and delivery can be reduced by including user, communities and frontline workers in the innovation process from the outset so they do not perceive the tool as imposed but co-owned.
Contextual validity is equally critical. A model trained primarily on data from a well-resourced urban facility performs differently in a primary health centre in a tribal and forested block of Jharkhand, where population characteristics, disease burden, belief system, input quality, and workforce capacity may differ substantially. Public health systems in many parts of India continue to operate with infrastructural deficits, including intermittent electricity and poor internet connectivity, alongside language and literacy barriers in user interfaces for frontline workers and behavioural resistance among some communities. Extrapolating performance claims across these contexts without continued and triangulated validation is a scientific failure, and the innovation community must prepare or bear responsibility for it.
What can be done with this shared pre-condition
Governance infrastructure, funding architecture, and user- and community-embedded innovation design are co-constitutive preconditions. This means both policymakers and innovators have concrete work to do. Philanthropy can also play an enabling role, helping build the technical capacity and the human expertise needed to sustain regulatory literacy on one side and community-embedded design on the other.
Policymakers need to introduce regulations that are proportionate and adaptive to evolving algorithmic innovations. India still has no binding AI liability law for health algorithms, only soft-touch frameworks which cannot be enforced. A sector-specific notification under existing IT or health law that assigns provisional liability across developer, procurer, and implementer, before a tool reaches a public facility’s procurement list is a viable first step that does not require new legislation.
It also means transforming the Benchmarking Open Data Platform for Health AI (BODH), which provides a structured mechanism to test AI models for performance and bias before real-world deployment, ensuring they meet clinical and public health standards. But such assessment should not function as a one-time stamp of approval. Clearance should be renewed when a model is deployed in a substantially different geography or context, with the results of these assessments made publicly available. Defined turnaround timelines should be established for such revalidation, allowing it to function as a routine but fast checkpoint rather than a fresh bureaucratic bottleneck every time.
Innovators themselves often mistake early traction for success. A successful pilot is simply an important milestone, not the destination. The real challenge lies in scaling responsibly, sustaining impact, navigating regulation, generating evidence, integrating with health systems, and ensuring long-term adoption.
Innovators therefore need to understand the scope of the journey before they reach the starting line. Before launching a pilot, they should have a credible plan for what comes after it. Technology product profiling at every stage of development, not just at the point of procurement, should become standard practice. Before claiming scale-readiness, innovators should publish evidence of how a tool performs outside the context in which it was trained: a different state, a different disease burden, a different level of infrastructure. Funding models need to match this ambition, with explicit resources for user co-design and revalidation across contexts.
The Ayushman Bharat Digital Health Mission, growing state-level appetite for AI-assisted public health tools, and the extraordinary density of health-tech innovation in this country create conditions that will not persist indefinitely. Technology is coming whether we are ready or not. Achieving readiness would involve establishing governance mechanisms capable of evolving as fast as the technology they oversee.
What ultimately matters is whether accountability and impact will serve as key levers for the technology that is to come. Will the health worker in a primary health centre view it as support, surveillance, or an additional burden? Will the communities it touches have any say in how it was built? For India’s health-tech innovators, the answer starts with a shift in posture: build for the health system as it actually is, not the one that is easiest to pilot in.
*with inputs from Dr Anant Bhan.
—
Know more
- Read this article to learn more about the role chatbots are playing in filling gaps to reproductive healthcare access.
- Learn more about why AI health interventions fail to scale.
