Listen at Scale

The Department of Empowerment of Persons with Disabilities (DEPwD) manages welfare, schemes and certification for ~2.68 crore people with disabilities across 21 recognized types. Existing discovery of needs depended on slow field profiling, overloaded helplines and static registries with incomplete, low-fidelity profiles. Field teams cannot scale to millions; helplines and IVR lack nuance and fail proxy or guardian scenarios. Voice enables accessible, natural multi-turn dialogue, handles mid-call speaker handovers, and verifies intent and consent in culturally realistic settings.

Sarvam and EkStep built an outbound, multilingual, multi-turn voice profiling agent that: ① captures explicit consent; ② collects structured needs, disability type and socioeconomic context; and ③ writes verified, allocation-ready profiles directly into ONEST for downstream matching. The agent called over 14 Lakh people in a single wave to collect 2.8+ Lakh minutes of structured dialogue to create 51k+ profiles.

In a single wave, the campaign built the foundation of an actionable registry larger than most field programs achieve in years. The qualitative depth (specific economic context and aspirations) makes these profiles uniquely useful for supporter matching.

Piramal Foundation partnered with Sarvam and EkStep to solve a classic rural governance bottleneck across Bihar and Jharkhand: thousands of Panchayats and tens of thousands of ward-level functionaries lacked timely, role-specific guidance, producing an "information-to-implementation" gap. Traditional training and one-way broadcasts could not deliver the cadence, two-way clarification or machine-readable feedback needed across 16k+ panchayats; the program therefore required hundreds of thousands of targeted, contextual interactions every year to close that gap.

Piramal Foundation deployed a multilingual, role-aware outbound AI voice agent (AI Sachiv Ji) that converted policy updates into structured, role-specific conversations and logged readable insights. Operationally, the system demonstrated burst capacity (49k+ interactions in a single day) and sustained throughput (1.1+ Lakh interactions across 141 districts and 16k+ panchayats in the campaign window), while neutral guardrails, localized tone, and escalation paths preserved quality. Rapid iteration on the knowledge base delivered immediate gains: policy accuracy jumped from ~48–55% in Phase 1 to 95%+ in Phase 2, and evaluated interactions scored 84% on explanation clarity with 99.8% political neutrality.

When the system identified the role, clarity rose to 93% and accuracy to 80%, with average call duration expanding 4.6× (47s → 215s) showing that identity-first flows are the single highest-yield investment for quality.

India's Ayushman Vay Vandana Yojana promises free health coverage for senior citizens aged 70+, yet awareness and enrolment lagged far behind eligibility. More than three crore seniors, many with low digital literacy and limited mobility relied on sporadic camps or service centre visits to learn about the scheme. Human led outreach at national scale was prohibitively slow and expensive, leaving high-need districts underserved and administrators without a clear view of intent, barriers, or operational gaps.

To bridge this access gap, the National Health Authority partnered with Sarvam and EkStep to deploy a multilingual, outbound conversational voice campaign tailored to elderly users. The system explained benefits in simple language, answered FAQs, captured intent to enrol, and verified identity and location for routing to offline enrolment channels. Designed for real-world conditions, it handled family mediation, guided seniors step-by-step through next actions, and escalated complex cases to human helplines, all while ensuring consent, disclosure, and data protection.

SwaTaleem Foundation is a nonprofit that works to strengthen girls' education by building engagement between Kasturba Gandhi Balika Vidyalaya schools and families across underserved communities in India. For over five years, they relied on IVR to reach parents of girls studying in Kasturba Gandhi Balika Vidyalayas across Haryana and Karnataka. While this ensured wide coverage, the interaction was fundamentally passive: parents could listen, but their understanding, concerns, or engagement could not be measured. In rural contexts where school–family linkage is critical to retention and safety, this created a blind spot where outreach was happening, but its effectiveness was unknown. The challenge was to transform mass communication from one-way broadcasting into meaningful participation without increasing staff burden.

Between 16 January and 10 February 2026, SwaTaleem piloted a multilingual conversational voice campaign with Sarvam and EkStep reaching 6.5k+ parents in Hindi and Kannada on practical digital safety topics including harmful apps, deepfakes, cyber fraud, safe payments and parental controls. Unlike IVR, parents could respond, navigate structured flows and provide feedback in real time. The system handled outreach end-to-end, maintaining scale while generating measurable interaction data. Engagement exceeded expectations: 61% of the parent base was reached, 61.03% actively interacted (more than double the 25% target), and 16% sustained deep conversations of 10 or more exchanges, demonstrating that narrative-driven flows can hold attention even in low-bandwidth rural settings.

In just four weeks, the pilot proved that large-scale parent outreach does not have to trade depth for coverage. Participation became quantifiable, operational effort decreased exponentially, and engagement quality improved without expanding human teams.

The effective execution of large scale state-run agriculture programs often faces challenges with operational transparency on the ground. The Government of Maharashtra (Agriculture Department) needed fast, reliable answers about whether trainings ran, whether farmers adopted practices, and actually reached beneficiaries; problems that traditional surveys and field visits only exposed months later.

The Department partnered with Sarvam and EkStep to deploy a conversational AI agent to collect large-scale, structured citizen feedback: Farmer Field School (FFS) training feedback and implementation. The initiative converted tens of thousands of outbound calls into measurable program intelligence; from training adoption to corruption flags, thereby enabling rapid, district-level governance insight.

In weeks, the system converted lists into a live statewide pulse check with 2.4 lakh+ call reachouts, 71k+ meaningful conversations, creating one of the largest real-time feedback datasets on agricultural programs in the state. Training effectiveness could be measured and it was observed that 59% of Farmer Field School attendees reported implementing learned techniques.

These signals enabled administrators to move from assumption-based monitoring to evidence-driven action, functioning as a lightweight policy evaluation tool that provides a rapid feedback loop while Maharashtra's policies are being implemented.

The Government of Maharashtra operates Aaple Sarkar Seva Kendras across the state to deliver essential citizen services such as caste certificates, domicile certificates, income certificates, and other revenue services. While the system processes large volumes of applications, administrators lacked a scalable mechanism to verify citizen experience and detect operational issues in real time.

Traditional feedback mechanisms (IVRS, Call Centres) relied on manual audits or sporadic surveys, which provided limited visibility into service delivery quality across districts. This created blind spots around citizen satisfaction, operational friction, and potential integrity risks such as middlemen involvement.

To address this gap, the Government of Maharashtra partnered with Sarvam to deploy a large-scale conversational AI voice campaign to gather citizen feedback directly after service interactions. The AI agent verified beneficiary identity, captured service ratings, audited whether multiple visits were required to complete a service, and flagged reports of unofficial transaction costs.

The system conducted at-scale outreach across 35 districts, enabling administrators to capture structured feedback from thousands of citizens and convert conversations into actionable governance insights. In a week, the campaign created a large real-time citizen experience dataset for public service delivery in the state—transforming feedback from anecdotal citizen signals into actionable governance intelligence.

Civis is a community-driven non-profit platform built and maintained by Civic Innovation Foundation to increase citizen participation in the lawmaking process, to generate public awareness and provide education on issues of civic importance. India's Union Budget affects every household in the country, yet public engagement with it is typically shallow and uneven. Most citizens encounter the Budget through fragmented news coverage or highly technical documents that are difficult to interpret, especially for non-digital and non-English audiences. Call centres cannot absorb the surge in questions following the announcement, and written FAQs exclude those without internet access. As a result, civic engagement remains largely informational rather than participatory. Citizens hear about policy changes but rarely understand how they affect their daily lives.

To bridge this gap, Civis, in partnership with Sarvam and EkStep, deployed a multilingual conversational voice system immediately after Union Budget 2026. The system proactively contacted citizens and enabled two-way conversations in which users could ask questions, seek clarifications, and receive simplified explanations across domains such as taxation, agriculture, healthcare, and education. Each interaction generated structured data on topics of interest, satisfaction, and willingness for further participation, transforming individual conversations into aggregated civic insight while maintaining consent and disclosure safeguards.

Within an 11-day window, the initiative converted budget awareness into measurable engagement at national scale. 1+ Lakh call attempts resulted in 59k+ live conversations, with 21k+ interactions exceeding the meaningful engagement threshold and 24k+ minutes of policy dialogue recorded. Citizens not only listened but actively participated, with 1.3k+ reporting satisfaction and 542 expressing interest in deeper engagement through Civis platforms. Tens of thousands of citizens received personalised explanations, asked questions, and signalled understanding and interest, creating one of the largest real-time datasets on public comprehension during a national policy announcement.

Large-scale Minimum Support Price (MSP) procurement programs succeed or fail on last-mile readiness, yet governments rarely know in advance how much crop is actually available, who intends to sell, or what barriers farmers face. In Karnataka's maize-growing districts, officials needed rapid visibility into farmer readiness, inventory, and preferred procurement channels before operations began. Traditional field surveys would have taken weeks and significant manpower, risking missed harvest windows and farmers selling to private traders first.

To close this gap, the Agriculture Department with Sarvam and EkStep deployed an outbound conversational voice campaign to nearly 50k+ registered farmers over just four days. Across 90k+ call attempts to 49k+ farmers, the system connected with 35k+ farmers (72% reach) and generated 26k+ meaningful conversations, verifying crop status, sale intent, preferred Agricultural Produce Market Committees (APMC), and key concerns. The outreach identified 5k+ farmers ready to sell immediately and mapped 884k+ quintals of maize availability, of which 676k+ quintals were intended for MSP procurement. Retry strategies proved critical, recovering thousands of connections, while call analytics surfaced practical issues such as outdated records, crop diversification, and farmers who had already sold before outreach began.

The campaign provided insights into procurement planning, revealing concentrated demand in early February, specific APMC preferences, and major information gaps around payment timelines, transport costs, and process clarity. These insights enabled targeted follow-up prioritizing ready sellers, scheduling callbacks for not-ready farmers, and refining communication to address concerns while highlighting structural fixes needed for future cycles, such as earlier outreach and database updates.

Lend A Hand India (LAHI) is a nonprofit organisation that integrates vocational education in secondary schools across India. It works with 50k+ students each year, primarily in underserved and rural communities, helping them build practical skills and giving exposure to future education and career pathways. The organisation wanted a way to holistically connect with their students after they left school: informal follow-ups and spot checks left LAHI guessing whether vocational training translated into real career mobility, and where support should be focused. Without readable outcome data, program design and donor reporting remained anecdote-driven rather than evidence-led, a blind spot across tens of thousands of alumni that blocked targeted interventions and impact measurement.

The program used a multilingual, voice-based agent to verify identity, validate academic linkages, and classify current status across education, employment, skill training and disengagement at scale.

Outreach attempted 1+ lakh alumni and processed 1.1+ Lakh call interactions, producing 14k fully verified alumni records with confirmed identity, education/employment status, college/course or employer details, income bands and not in education, employment, or training (NEET) reasons. The tracer delivered real-time validation, rich contextual fields and a readable dataset.

The Government of Odisha needed timely evidence on whether benefits were actually reaching farmers under two major initiatives: the Comprehensive Rice Fallow Management (CRFM) Scheme, which supports cultivation of pulses and oilseeds on fallow paddy land during the Rabi season, and the Mission for Atmanirbharta in Pulses (MAP), which aims to boost pulse production through seed distribution, inputs, and farmer support.

Traditional surveys (IVRS/Call centres) and field visits are too slow and costly, and beneficiary databases contained outdated or shared phone numbers, leaving administrators without reliable insight into delivery gaps, adoption barriers, or awareness levels while the programs were underway. Manual surveys and field checks are slow and costly, and beneficiary contact data is noisy, where nearly 30% of targeted farmers remained unreachable in initial waves.

To address this, the Department deployed conversational voice agents to conduct structured feedback calls at population scale. Across 1.1+ Lakh call outreach targeting 45k+ unique beneficiaries, the system ultimately reached 31k+ farmers, collected 13k+ completed surveys, and generated over 21k+ meaningful conversations. The outreach verified 5.1k+ seed receipts and 2.5k+ input procurements, while retry campaigns successfully connected hundreds of farmers who were unreachable initially. Call-level analytics also surfaced operational realities such as identity confirmation gaps and information asymmetry in terms of low explicit scheme awareness (≈21–22%).

The project confirmed that seed distribution is broadly functioning (≈77% receipt among respondents) but revealed an adoption gap in input procurement (63%), indicating barriers such as dealer access, cost, or awareness. The findings provide a roadmap for immediate improvements: expand dealer networks, bundle inputs with seed distribution, strengthen communication campaigns, clean beneficiary databases, and institutionalize retry-based outreach as a continuous monitoring tool. In effect, the initiative gave Odisha a fast, scalable feedback loop thereby enabling course correction while schemes are active.