Many LMIC countries are releasing artificial intelligence strategies. Some of those AI strategies focus on healthcare. Most of those follow the same template: aspirational language, vague commitments to equity and safety, a precautionary tone borrowed from WHO guidance, and zero operational infrastructure.
They are policy documents in the precise sense of the term. Documents. Not policies.
India’s Strategy for Artificial Intelligence in Healthcare (SAHI), released in February 2026 by the Ministry of Health and Family Welfare, is different. And I think we in the ICT4D community are underestimating just how different it is.
SAHI is a government attempting to govern AI that is already running at population scale inside public health systems. That should change how every practitioner, donor, and policymaker in our sector reads this document.
India Started with DPI Infrastructure
When most LMIC governments write AI health strategies, they are describing a future they hope to build. India is describing infrastructure it already operates.
- eSanjeevani telemedicine platform has logged over 282 million consultations using AI-generated differential diagnosis recommendations.
- Ayushman Bharat Digital Mission has already deployed digital health identifiers, consent-based data exchange, and interoperable health records at national scale.
- Media Disease Surveillance has scanned digital news sources nationwide, publishing over 4,500 infectious disease event alerts to help districts respond early and reduce outbreaks.
When AI and DPI converge properly, they can democratize access to high-quality public services globally. India has reached that convergence. SAHI is the governance layer arriving to manage it, not to imagine it.
That’s the structural reason SAHI is more interesting than it looks on the surface.
Innovation over Restraint: Provocative Direction
SAHI codifies seven governing principles, which it calls “sutras.” The third reads:
“AI innovation in healthcare should aim to maximise overall benefit while reducing the potential of harm. All other things being equal, responsible innovation should be prioritised over cautionary restraint.”
Take a moment with that.
The dominant global framing since 2023 has been precaution-first.
- The WHO’s six principles for AI in health foreground protecting human autonomy and safety.
- The EU AI Act builds an elaborate compliance scaffolding around health AI, and its implementation is already stumbling.
No comparable LMIC health-sector strategy has ever elevated innovation parity with, let alone above, precaution as a named governing principle.
India is explicitly signaling it will not adopt the EU’s precautionary model. This positions SAHI closer to the pre-2025 American “soft touch” regulatory philosophy while retaining the language of risk-proportionality. For the 50+ LMICs watching both models, this is potentially the most influential signal in the document.
The practical implication matters for all of us working in digital health.
This principle gives regulatory and program managers explicit doctrinal cover to approve and scale lower-risk AI applications, administrative tools, operational analytics, claims management, without requiring the same evidentiary burden as high-risk clinical tools. That is the specific kind of regulatory clarity that has paralyzed AI adoption across most of the countries we work in.
The risk is real, though.
“Innovation over Restraint” can only function as intended if Indian regulators can reliably distinguish innovation-enabling governance from regulatory neglect. Without strong institutional capacity to operationalize risk classification, this principle could quietly become a justification for inadequate oversight of harmful applications.
Three Innovative SAHI Features
1. A Tripartite Data-Gap Taxonomy
Most AI-health strategies treat data readiness as a binary. Either data exists and is usable, or it does not. The policy recommendation is always the same: collect more data. This is simultaneously true and useless.
SAHI introduces a three-tier framework distinguishing:
- AI-critical gaps, where lack of data directly blocks priority use cases
- AI-limiting gaps, where fragmentation reduces system intelligence
- AI-enhancing datasets, where high-quality data improves model performance.
This taxonomy is analytically sharper than anything in the WHO’s guidance architecture or the EU AI Act’s data requirements.
If operationalized, this shifts the public investment conversation from “we need more data” to “which data deficits are blocking which AI use cases, and what is the marginal return on closing them?”
That is a procurement and budgeting innovation.
The challenge, and it is a significant one, is that the taxonomy is introduced conceptually but not operationalized. No institution is assigned to conduct the mapping. No protocol links the taxonomy to investment decisions.
For now it is an elegant analytical construct without an implementation pathway.
2. Outcome-Oriented Procurement as a Strategic Pillar
SAHI dedicates an entire strategic pillar to what it calls “market stewardship.” It explicitly diagnoses the pathologies of current procurement:
“When procurement is fragmented, opaque, or narrowly cost-driven, it can incentivise stop gap solutions, vendor lock-in, and misalignment with health system needs.”
This is something development practitioners know viscerally.
We see it constantly, as we’ve explored in discussions about sustainable digital health incentives. The supply side gets all the attention. Regulation gets discussed. The demand side, how governments actually buy and adopt AI tools, is almost universally absent from LMIC AI strategies.
India’s public health expenditure, channeled through the National Health Mission and AB PM-JAY, represents an enormous potential demand signal for health AI vendors across South and Southeast Asia. If India actually reforms procurement toward outcomes, it reshapes the competitive landscape for the entire region.
The gap is the same as elsewhere: procurement reform requires changes to treasury rules, not just health ministry guidelines. SAHI does not specify the institutional pathway for that reform.
3. BODH: Strategy and Implementation Tool
This one genuinely surprised me.
SAHI was launched simultaneously with BODH, the Benchmarking Open Data Platform for Health AI, developed by IIT Kanpur with the National Health Authority. BODH allows innovators to train and validate healthcare AI models against anonymized, real-world health data from a centralized repository.
National AI strategies in LMICs almost never launch with a complementary operational instrument.
Strategy documents and implementation platforms are typically separated by years, if the platform arrives at all. The concurrent release of SAHI and BODH represents an unusual attempt to close the strategy-to-implementation gap at the moment of policy announcement.
If BODH proves functional, it offers a replicable model for countries that cannot afford the regulatory overhead of the EU’s conformity assessment system or the FDA’s pre-market approval pathway. A soft regulatory mechanism through benchmarking, requiring developers to validate against standardized datasets before public deployment, is faster and more adaptive than statutory approval.
The critical caveat is dataset quality.
SAHI’s own tripartite taxonomy acknowledges that data adequate for administrative purposes may be insufficient for clinical AI applications. BODH’s utility depends entirely on whether its underlying datasets are representative and current enough to validate high-risk tools, not just low-risk administrative ones.
What SAHI’s Gaps Tell Us About the Sector
SAHI has three structural weaknesses practitioners should flag immediately.
1. No dedicated financing mechanism.
The strategy mentions blended finance in passing but contains no budget allocation, no cost estimates, no identification of funding sources. Thirty-two recommendations require sustained investment across governance, data infrastructure, workforce training, and procurement reform.
The silence on financing is a structural weakness that undermines every other ambition in the document.
2. No implementation timeline or accountability metrics.
SAHI contains no phased roadmap, no milestones, and no key performance indicators. It describes itself as a “living document,” which provides flexibility and creates a risk of indefinite deferral.
3. No patient voices in the governance architecture.
The consultative process involved government, academia, industry, and professional bodies. The governance architecture itself contains no structural role for patient organizations, disability rights groups, or community health advocates.
For a strategy that foregrounds “People First,” this is a significant omission.
What We Should Do With SAHI
Watch SAHI closely, because it is a live test case of a specific hypothesis: that a large, federal, mixed-economy LMIC can operationalize risk-based AI governance in healthcare atop existing DPI, without new legislation and without a centralized regulatory agency.
No other country has attempted this combination at comparable scale.
If it works, it offers a replicable model. If it fails, the failure will be instructive in ways that theory-heavy frameworks cannot replicate. It will reveal precisely which institutional bottlenecks derail even the most analytically sophisticated LMIC health AI strategies.
For donors, the immediate question is whether SAHI’s financing gap becomes your entry point.
Outcome-oriented procurement reform, BODH dataset development, and centre-state coordination mechanisms all represent investment opportunities that would convert a governance blueprint into operational infrastructure.
The countries that should be watching SAHI most carefully are not the ones that can replicate India’s DPI stack. They are the ones that can adapt SAHI’s tripartite data taxonomy, its market stewardship framing, and its “innovation over restraint” doctrine to their own institutional contexts, without waiting for the full infrastructure to be in place first.
If governance doctrine travels faster than digital infrastructure, then SAHI has already exported its most valuable asset.
