Image Source: Getty Images
India’s one million Accredited Social Health Activists (ASHAs) anchor the country’s primary healthcare system. Yet evidence points to uneven knowledge, especially in danger-sign recognition and timely referral, where training and supervision are incomplete. Contemporary Indian AI policy has focused on foundation models, sovereign compute, and IT sector disruption. While these supply-side priorities are necessary, this article argues for deploying AI for public good with equal ambition. A voice-first, multilingual protocol AI assistant for India’s entire community health workforce is proposed, following the country’s established digital public infrastructure playbook. The state would own an open protocol and evaluation standard, with mandatory error audits and human-in-the-loop safeguards as binding conditions for deployment. The article thus reframes India’s ambition for health AI around a measurable objective: artificial intelligence should help frontline workers apply approved health protocols consistently at the point of care.
India’s AI Policy Beyond Compute and Models
The India AI Impact Summit of February 2026 anchored its agenda in ‘AI for All‘. Yet Indian artificial intelligence policy has concentrated overwhelmingly on the supply side, with the demand side receiving far less structured policy attention. The IndiaAI Mission, for instance, directs the bulk of its outlay towards compute capacity and model development. NITI Aayog’s 2018 national AI strategy had already identified healthcare as a priority sector, explicitly linking AI adoption to shortages of skilled medical expertise. Yet a dedicated health AI framework arrived only in February 2026, in the form of the Strategy for Artificial Intelligence in Healthcare (SAHI) and the Benchmarking Open Data Platform in Health AI (BODH).
AI’s value in public healthcare should be measured by gains at the point of care, such as referrals correctly made, danger signs caught in time, and complications averted, rather than by compute power provisioned or models trained, where the bulk of public AI funding currently flows.
Recent analysis frames AI as a force multiplier for a staff-scarce health system, not a replacement for clinicians. This article extends that logic from what AI is for to how its success should be judged. AI’s value in public healthcare should be measured by gains at the point of care, such as referrals correctly made, danger signs caught in time, and complications averted, rather than by compute power provisioned or models trained, where the bulk of public AI funding currently flows.
One high-value candidate where AI can make a measurable impact is India’s community health workforce. Launched in 2005 under the National Health Mission, the ASHA programme fields over a million workers who connect households with the public health system. Their record on immunisation mobilisation, institutional deliveries, and family planning is well established. However, evidence on their clinical knowledge beyond these routines reveals a structural gap.
The Frontline Knowledge Gap
Two decades of health systems research converge on a consistent picture of the ASHA programme. A 2025 systematic review and meta-analysis pooled evidence from 37 studies published between 2005 and 2022 and estimated ASHAs’ knowledge at 62 percent for maternal health and 69 percent for neonatal and child health. ASHAs were strongest in maternal and child health, their core programme area. However, the review identified specific weaknesses in child referral for severe conditions such as diarrhoea and respiratory tract infection. Knowledge of danger signs and pregnancy complications was also inconsistent across studies.
An empirical study in Assam found that two-thirds of workers had not received the full recommended 23 days of training. A further weakness, identified in a study from northern India, concerns role comprehension: nearly all workers knew their immunisation and delivery-related duties, yet very few recognised village health planning, counselling, and adolescent health as part of their mandate. Research from Odisha similarly documents job-role confusion among rural ASHAs.
The reviews attribute these gaps to education level, experience, inadequate supervision, and insufficient training, not to any failing of the workers themselves. ASHAs function to fulfil their core mandate, often under difficult conditions and with limited support. Referral is the point at which a frontline worker’s judgement determines whether a person reaches a clinician in time. This is where a well-designed AI protocol layer can help, surfacing danger signs and standardising referral decisions at the point of care, as a complement to — rather than a substitute for — the broader human-resource agenda of better training, supervision, and support.
A Protocol Assistant in Every Hand
The IndiaAI Mission should create a dedicated public-health application window, jointly governed by the Ministry of Health and Family Welfare (MoHFW), the National Health Authority (NHA), and the Ministry of Electronics and Information Technology (MeitY), to fund state-led pilots and scale-up of a voice-first AI protocol assistant. This assistant, an ‘AI Sahayak’, would reach every ASHA worker and every Ayushman Arogya Mandir.
Design Principles
- Assistive by design. The Sahayak supports triage and leaves diagnosis to clinicians. It walks the worker through standardised protocols drawn from Indian Council of Medical Research (ICMR) and MoHFW guidelines. These protocols cover danger signs in pregnancy, newborn sepsis, and tuberculosis screening. When an ASHA worker meets a patient in the community, she describes what she observes, and the AI tool walks her through the relevant protocol question by question and tells her whether the case needs referral. The ASHA then acts on that prompt by counselling the family, arranging transport, and escalating the case to the auxiliary nurse midwife (ANM), the community health officer (CHO), or the medical officer, who make the clinical judgement and decide on diagnosis and treatment.
- Voice-first and multilingual. The interface operates by voice on the worker’s existing smartphone, and the design must account for the conditions ASHAs actually work in, from low-end and often shared handsets to limited battery and the cost of mobile data. It must function in the scheduled languages and major dialects, and work offline-first, synchronising periodically to match the connectivity realities of rural India.
- Built on the DPI playbook. India’s comparative advantage lies in open, population-scale digital public goods such as Aadhaar, UPI, and CoWIN. The Sahayak should be developed accordingly. A public unit under the National Health Authority and the Ministry of Electronics and Information Technology would publish an open protocol, a clinical evaluation benchmark, and data governance rules. Private model providers would then compete to certify against the benchmark.
Accountability Architecture
Deploying AI into clinical pathways at population scale demands guardrails commensurate with the risk, operationalising the lifecycle governance direction set out in the SAHI strategy. The strategy’s accountability principle allocates liability across developer, deployer, service provider, and the user to identify the point of failure that may have caused harm. Three provisions are non-negotiable.
First, mandatory error audits, published quarterly and disaggregated by protocol module and language, so that model drift or bias surfaces early rather than as silent harm in the field.
Second, a liability framework built around two commitments. It protects the ASHA from tool error, so that a worker who follows the Sahayak’s protocol faithfully does not bear responsibility for a fault in the model, its data, or its guidance, which instead attaches to the developer or deployer as SAHI’s tiering envisages. It keeps a qualified clinician — the ANM, CHO, or medical officer — as the accountable decision-maker for every referral and treatment action, so that the human in the loop is not just a formality but the point at which errors are caught before they reach the patient.
Third, a kill-switch requirement, under which any protocol module whose audited error rate crosses a pre-specified threshold is suspended automatically pending review. Together, these provisions convert ‘responsible AI’ from a communiqué phrase into an enforceable operating condition.
Anticipating the Objections
“The accuracy risk is unacceptable.” The relevant comparison is not the Sahayak against a physician, but the Sahayak-assisted worker against the status quo of memory-reliant recall by a worker operating with variable training and supervision.
“This substitutes technology for investment in workers.” The proposal is explicitly an augmentation mandate. The Sahayak sits alongside the pending agenda of ASHA remuneration reform, training completion, and payment-system reliability documented in the literature.
“Connectivity and power constraints will defeat it.” Offline-first design addresses connectivity at the device level. The deeper constraint is infrastructural: India’s AI ambitions ultimately run through its power grid and data-centre buildout. The Sahayak’s planning should be integrated into the National Electricity Plan alongside broader AI demand.
India’s AI moment will be judged not by the summits it convenes or the models it trains, but by whether the technology measurably improves the lives of citizens whom the formal economy and health system reach last.
Conclusion
India’s AI moment will be judged not by the summits it convenes or the models it trains, but by whether the technology measurably improves the lives of citizens whom the formal economy and health system reach last. The ASHA programme is where India’s last-mile health capacity meets its digital public infrastructure strength. The bottleneck is timely access to usable, locally intelligible protocols; the strength is population-scale digital infrastructure. A single, well-guarded policy decision can join the two. If even a fraction of frontline interactions move from memory-dependent practice to protocol-supported decision-making, the welfare returns will exceed those of any prestige project in the national AI portfolio.
Ajey Pai Karkala is a PhD researcher in artificial intelligence, working on AI for oncology.
Disclosure: Claude Opus 4.8 was used for language refinements.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.
