AI voice agent for social-impact programmes in India
An AI voice agent reaches the households a development programme is designed for and usually cannot call twice — parents, smallholder farmers and forest-dwelling families — in their own language, on an ordinary phone, for five minutes at a time, week after week.
- 22+ languages
- Hindi, regional and tribal — Santhali in Ol Chiki included
- No app
- A feature phone and a voice are the only requirements
- 5 minutes
- The length of a coaching or advisory call a family will take
- Every call
- Pulse survey and reading assessment data as a by-product
Most social-impact programmes in India are not short of good design. They are short of contact. A literacy programme that works when a parent is coached weekly reaches that parent twice a year, because one field worker covers forty villages. A farmer advisory that would change a sowing decision arrives after the sowing. A tribal livelihood scheme with a guaranteed minimum price for forest produce is worth nothing to a household that does not know the price on the day the trader arrives at the door.
A voice call is the only channel that closes that gap at programme scale. It needs no smartphone, no data pack, no app store, and no ability to read a form — which is precisely the profile of the households these programmes exist for. An AI voice agent can make that call to every family on the roll, hold a real two-way conversation in the language spoken at home, and do it again next week.
AiSewak builds and runs those agents across three areas: education, agriculture and tribal livelihoods. The work is mostly language, listening and follow-up design rather than model choice — a Santhali-speaking household hears textbook Hindi as an outsider, and a parent who is asked a question they cannot answer hangs up and does not pick up next week.
What a programme runs through the agent
Parent coaching for foundational literacy
One five-minute activity a week, in the language spoken at home: what to read with the child tonight, how to hold the book, what to ask afterwards. Built for a parent who may not read themselves.
Oral reading fluency assessment
The child reads a passage over the phone and the agent scores it. An assessment round stops being a field trip and becomes something a district can run every month.
Teacher and field-worker microlearning
Short coaching calls for the people delivering the programme, on the same line the households are on — the cheapest way to keep a cadre consistent across a district.
Farmer advisory and scheme status
Sowing-window advice, PM-KISAN instalment and e-KYC questions, and the eligibility conversation a farmer would otherwise queue for at a common service centre.
Minimum support price lines
The day's declared price for the crop or the forest produce, spoken plainly, before the trader arrives. Price information is the part of a livelihood scheme that decays fastest.
Tribal livelihood and Van Dhan outreach
Minor forest produce procurement, Van Dhan Vikas Kendra guidance and scheme entitlements, in Santhali and the other languages a state helpline does not staff.
Measurement stops being a separate exercise
The expensive part of impact measurement is not the analysis. It is getting a person to a household to ask the questions — which is why baseline and endline usually happen twice in a five-year programme, on a sample, and tell you very little about the middle.
When the programme already calls every household weekly, the measurement rides along. Two pulse questions at the end of a coaching call produce a district-level read every month rather than every second year. A reading assessment is a call the child was going to get anyway. The agent records refusals and drop-offs as honestly as it records answers, which matters more than it sounds: a block where nobody is picking up is a finding, not a gap in the data.
What this does not give you is a causal claim. Call-derived data tells you what is happening across the whole roll, at monthly resolution, at almost no marginal cost. It does not replace an evaluation design, and any vendor implying otherwise — including us — should be pushed on it.
Reaching a household in the language it actually speaks
"Supports 22 languages" is a much weaker claim than it sounds in a social-impact context, because the households furthest from a programme are usually furthest from its official language too. A Santhali-speaking family in Jharkhand, a Gondi-speaking family in Bastar, a Bhojpuri-speaking parent in Purvanchal — each hears standard Hindi as the voice of an institution that has not previously been useful to them.
Two things follow. The agent has to speak the language of the home rather than the language of the scheme document, including when the two are not the same. And the scheme's own vocabulary — the name of the entitlement, the produce, the form — has to be pronounced the way the household says it, not the way it is written. Agents here run on a mix of Indic-first models and the MeitY Bhashini stack, chosen language by language rather than platform-wide.
The rules we hold a programme line to
- 1
The agent says it is an AI, first sentence
No soft wording, no exceptions. A household that later discovers it was a machine stops trusting the programme, not the vendor.
- 2
TRAI TCCCPR, DLT and DND
Registered headers, calling-window limits and DND scrubbing before every wave. A welfare programme has no exemption from this and should not ask for one.
- 3
DPDP Act 2023, including children's data
Named purpose, consent captured in-call, a stated retention window and a deletion path that is actually run at programme close. Reading assessments involve children, which raises the bar rather than lowering it.
- 4
The agent never asks for money, OTPs or bank details
Hard-coded refusal. Beneficiary households are the most heavily targeted group in India for exactly this fraud, and a programme line that behaves like one trains people to be defrauded.
- 5
It says when it does not know
A wrong answer about an entitlement gets acted on. The agent routes to a human rather than guessing, and the handover carries the transcript.
Related work on this site
Social-impact programmes usually sit next to a government department and a published helpline. These pages cover the parts that are not programme-specific.
Government departments and scheme delivery
Where the programme is run by a department rather than a foundation: eligibility, guided applications, grievance routing and adoption dashboards.
Read the page →Indian-language coverage in depth
Register versus language, code-switching, pronouncing scheme names, and when Bhashini is the right choice over a commercial model.
Read the page →Van Dhan in Santhali — what we actually built
The Hindi-to-Santhali bridge behind the tribal livelihood line, written up with the parts that did not work.
Read the page →Official citizen helpline numbers
What each statutory line handles, so a programme line hands off to it instead of becoming a slower route to the same place.
Read the page →The platform overview
Elections, government and social impact in one place, if you are not yet sure which of them describes your work.
Read the page →What the published evidence actually shows
Two figures a programme office can check, published by the Government of India about government deployments. Both are here because they establish what voice AI has demonstrably done for Indian households — not what it did for anyone we worked with.
Read this first. Every figure below was published by the organisation named beside it, about that organisation's own deployment. Neither is an AiSewak client, neither used AiSewak, and neither result is ours. They show what has been demonstrated in the category — the honest basis for what a programme might achieve, not a promise of what yours will.
- 93 lakhOfficial figure
The Government of India stated that Kisan e-Mitra, a voice-based AI chatbot for PM-KISAN queries built on the Bhashini stack, had answered more than 93 lakh queries and was handling over 8,000 farmer queries daily in 11 regional languages.
India's flagship government AI voice deployment for farmers, with figures placed on the public record. One inconsistency worth knowing: an April 2025 Lok Sabha reply gave “over 20,000” queries daily against 92 lakh cumulative, while this December 2025 figure gives “over 8,000” daily against 93 lakh. The two government statements do not reconcile, so we cite the later one and date it rather than presenting either as a trend.
Press Information Bureau, Government of India · “Artificial Intelligence Transforming Indian Agriculture” backgrounder (figures as of December 2025) · 2026
- 10 languagesOfficial figure
MeitY announced that UIDAI is deploying Sarvam AI's models for voice-based resident interactions and fraud detection across 10 Indian languages, hosted on-premise inside an air-gapped UIDAI environment, with no data leaving that environment.
Cited for one purpose only: it establishes that a central Indian authority has accepted AI voice under strict data-sovereignty conditions, which is the question a funder asks before a programme touches beneficiary data. No outcome, savings or accuracy figure is published on that page, and none should be inferred from it.
Press Information Bureau / MeitY · PIB release on UIDAI–Sarvam AI · 2025
Call a live agent before you decide
These are running agents, not recordings. Open one, press call and speak to it in Hindi or English — the same stack that runs the deployments described above.
FLN Saathi — parent literacy coaching
The weekly five-minute activity call, in Hindi, built around the NCERT foundational literacy and numeracy goals.
Open the demo →Kiran Didi — farmer voice line
PM-KISAN instalment and e-KYC questions, and the advisory a farmer would otherwise queue for.
Open the demo →Van Dhan Vani — minor forest produce
Minimum support prices for forest produce and Van Dhan Vikas Kendra guidance.
Open the demo →Van Dhan — Santhali
The same line in Santhali over the Bhashini bridge, for households a Hindi helpline does not reach.
Open the demo →Anand — programme follow-up
Post-programme participant check-in with escalation to a human teacher, the pattern any cohort programme needs.
Open the demo →पठन साथी — reading fluency
Oral reading fluency assessment over the phone, scored and rolled up to a district view.
Open the demo →Go deeper
AI for the Kisan Call Centre 1551
Why India's farmer helpline answers 45.7% of calls at peak sowing, and what a voice layer changes.
PM-KISAN scheme status by voice
The instalment-cycle demand spike, and why no dedicated voice line exists for it yet.
Multilingual voice AI and Bhashini
Using the MeitY stack when the household's language is not one a vendor sells.
Voice agents for surveys at population scale
Outbound measurement: what a call-based pulse survey can and cannot tell you.
Related pages: AI voice agents overview · For government departments · Indian-language coverage · Live deployments · Citizen helpline numbers · हिंदी में पढ़ें
Frequently asked questions
What is an AI voice agent for a social-impact programme?
A conversational AI that phones the households a programme serves and holds a real two-way conversation in the language spoken at home — coaching a parent, answering a farmer's question about an instalment, or reading out the day's minimum support price. It is not a recorded message: the person can interrupt, object, or ask something the script never anticipated, and the agent routes to a human when it does not know.
Which Indian languages does the voice agent speak?
Hindi plus the 22 scheduled languages, and the regional variants that decide whether a call actually lands — Bhojpuri, Awadhi, Marwari, Mewari, Magahi, Santhali and others. Agents run on a mix of Indic-first models and the MeitY Bhashini stack, and a single agent can switch language mid-call when the person on the other end does.
Does a household need a smartphone or internet?
No. The agent works over an ordinary phone call, which is the whole reason it fits this work — it reaches a feature phone on a weak network, and it reaches someone who cannot read a form. Where a household does have WhatsApp, the same programme can run on both and the agent will use whichever the family answers on.
How is a programme line kept compliant?
The agent discloses that it is an AI in its opening sentence. Outbound calling follows TRAI TCCCPR 2018 with DLT-registered headers, calling-window limits and DND scrubbing. Personal data is processed under the DPDP Act 2023 with a named purpose, in-call consent, a stated retention window and a deletion path that is executed at programme close — and reading assessments involve children's data, which raises that bar rather than lowering it. The agent never asks for money, an OTP or bank details.
Can it work with our existing MIS?
Yes. Beneficiary lists come out of your MIS or programme database and call outcomes go back in, so the record a field worker sees is the same record the agent updated. Where a state MIS has no write API, outcomes are exported on whatever cycle the department accepts.
Will this prove our programme works?
No, and be careful of anyone who says it will. Call-derived pulse data and phone-based assessment give you monthly, whole-roll visibility at almost no marginal cost, which is a large improvement on a biennial sample. They are not an evaluation design and they do not establish causation. Use them to see where a programme is stalling, not to claim it worked.
Start with one block, not one state
One block, a few thousand households and four weeks of calls, with your programme team reading the transcripts. That is enough to see whether the agent holds up in the language your families actually speak.
Every AiSewak agent identifies itself as an AI at the start of the call, never asks for an OTP or a payment, and honours DND. Election deployments require ECI / state CEO registration as a political advertiser.