Cluster C · India pillar

AI Agent for Indian Elections: ECI, Bhashini and the Bhasha Stack

The India-specific AI election stack — Election Commission rules, the MeitY Bhashini language infrastructure, and the architecture that's uniquely native to Indian electoral conditions.

8 min readUpdated 22 May 20261,598 words

The popular framing — "AI in elections" — implies a global category with global rules. In India, this framing is wrong. India has its own legal stack (ECI advisory, TRAI TCCCPR, DPDP, IT Rules), its own language infrastructure (Bhashini), its own structural conditions (96+ crore voters, 270 dialects, five-phase cycles) and its own emerging vendor ecosystem.

This guide is the India-native picture. What's unique about deploying AI for Indian elections, why generic global playbooks fail, and what the right stack looks like.

The structural conditions that force a different stack

Before tooling, structure. India's electoral conditions are not the US's or the UK's or Indonesia's. The differences compound:

Voter scale. 96.88 crore registered voters in 2024 (ECI). Single state — Uttar Pradesh — has 15+ crore voters, larger than every country except India, China and the US. A serious campaign in UP needs infrastructure that can do crore-scale outreach in weeks.

Language fragmentation. 22 official languages + ~270 dialects with >10,000 speakers each. A voter in a single Indian district might speak any of 3-4 dialect registers. Generic "Indian English" or even "Standard Hindi" tools miss most of the conversation.

Five-phase electoral cycle. Lok Sabha (every 5 years), state assemblies (staggered every 5 years), urban local bodies, panchayats, by-elections. AI infrastructure built for one election gets reused for the next 4-5 over the cycle. The economic unit is the platform, not the single campaign.

Telecom diversity. ~70% of adults own a phone, but only ~55% have smartphones. Older voters, rural voters and lower-income segments overwhelmingly use feature phones. Voice (a phone call) is the only channel that reaches all of them.

Cost sensitivity. Indian campaigns operate at lower per-voter budgets than US or UK equivalents — typically ₹20-200 per voter total budget, against $20-200 per voter in Western campaigns. Per-call AI costs that work in the US ($0.50-2/call = ₹40-160) don't work here. India-native pricing has to come in under ₹1.50 per call.

These conditions force three architectural choices that the global playbook doesn't anticipate:

  1. Voice-first, dialect-aware as the primary engagement channel
  2. India-residency by default for data and inference
  3. Cycle-scale rather than campaign-scale infrastructure economics

Bhashini: the national language stack

Bhashini (officially the National Language Translation Mission) is MeitY's open-source language infrastructure. Launched in 2022, mature by 2024-25, it now provides:

  • ASR (speech-to-text) for 22 scheduled languages, with dialect coverage for the largest ones
  • TTS (text-to-speech) for 22 languages with multiple voice IDs each
  • MT (machine translation) for pairwise translation between Indian languages
  • Indic LLMs — language models tuned specifically for Indian language understanding
  • Universal Language Contribution APIs — open contribution pipelines for community-tuned models

Where Bhashini shines

  • Language coverage breadth. No other provider matches Bhashini's coverage of smaller Indian languages — Konkani, Manipuri, Bodo, Santhali, Kashmiri.
  • Data sovereignty. All processing on India-hosted infrastructure. DPDP residency by default.
  • Licensing. Open-source models with permissive use, including political/government deployments.
  • Continuous improvement. Community contributions feed back into the models; quality improves quarterly.

Where Bhashini is still catching up

  • Conversational quality. The Indic LLMs are still behind global frontier (Gemini, Claude, GPT-4) on nuanced political conversation.
  • TTS naturalness. Some Bhashini voices still have detectable robotic intonation; the global frontier (ElevenLabs, Cartesia) leads on natural prosody.
  • Latency. Bhashini's hosted endpoints currently have higher p99 latency than commercial APIs — important for real-time voice agents.

Hybrid architecture in 2026

Most production Indian voice agents in 2026 use a hybrid stack:

  • LLM: Global frontier (Gemini 2.5 Flash, Claude Haiku 4.5, Qwen 3 30B A3B) for the main reasoning
  • STT: Bhashini for some smaller-language workloads, commercial providers (Deepgram, Google) for the main ones
  • TTS: Commercial (ElevenLabs Turbo, Cartesia Sonic) for the main voices; Bhashini for dialect fallback
  • Translation: Bhashini for India-language pairs; commercial only when needed

The mix shifts every 6 months as Bhashini's models improve. By 2027-28 the calculus likely tilts further toward Bhashini for sovereignty-sensitive workloads.

ECI compliance: what's binding

A short recap of the legal stack covered in detail in the compliance guide.

Binding rules for AI election agents in India (mid-2026):

  1. Disclosure: AI must self-identify in opening line. No impersonation.
  2. No deepfakes of opponents: explicit ban under the 2024 advisory.
  3. DLT registration: outbound calls require TRAI DLT templates.
  4. DND scrubbing: voters on the National DND must not be called.
  5. DPDP residency: voter data and call recordings stored in India.
  6. DPDP retention: documented retention period; right-to-erasure within 7 days.
  7. Silent period: no campaign communication 48 hours before polling.

Most enforcement happens in two ways:

  • Voter complaint → TRAI investigation → sender block (operational pain)
  • ECI Model Code referral → political consequences (campaign-ending exposure)

A well-configured AI platform handles items 3-7 by default. Items 1-2 are choices the campaign makes at the system prompt level.

What Indian campaigns get wrong

Five recurring patterns across the 2024 cycle.

1. Translating American playbooks

A campaign hires a US-based vendor whose playbook is built for Iowa caucuses. The vendor doesn't understand DLT, doesn't have Indian sender pools, doesn't speak Bhojpuri. The campaign spends 6 weeks on integration, then realises the original premise was wrong.

2. "We'll do Hindi" without dialect testing

Hindi-only voice agents in regions with strong dialects (Marwari, Bhojpuri, Awadhi) get sub-30% completion rates. The campaign concludes "AI doesn't work in this state" when actually they shipped without dialect tuning.

3. Cloud datacentres in Singapore / Frankfurt / Virginia

Latency penalties of 200-500ms one-way. The voter perceives "the line is laggy". Combined with DPDP residency concerns, this is a hard fail.

4. Skipping the silent period

A campaign keeps calling in T-2 days because nobody told the vendor to pause. Voter complaints → TRAI investigation → senders blocked at the worst possible time.

5. Treating the platform as a campaign cost, not infrastructure

The campaign signs a 60-day contract, runs the platform during the campaign window, then turns it off. The 5-year governance opportunity is wasted. The next campaign (2 years later) starts from scratch.

The right India-native stack

A reference architecture for an Indian campaign serious about AI:

Application layer:

  • Specialist Indian AI election platform (AiSewak, Voxdonna, Saaras, others) — handles system prompt, dashboards, compliance defaults
  • TRAI-DLT integration baked in
  • DPDP-compliant data handling default
  • Sandbox + production environments

Language layer:

  • Hybrid LLM: global frontier for primary reasoning, Bhashini for sovereignty fallback
  • Hybrid STT: best-quality for primary language, Bhashini for smaller languages
  • Hybrid TTS: ElevenLabs/Cartesia for natural primary voices, Bhashini for dialect fallback
  • Dialect tuning per state (Marwari for Rajasthan, Bhojpuri for Bihar, etc.)

Infrastructure layer:

  • All inference + storage in India-region datacentres (Mumbai, Hyderabad, Bangalore)
  • SIP trunks via Indian carriers (Jio, Airtel, Vi, BSNL)
  • DND scrubbing as a daily automated step
  • 99.5% uptime SLA during campaign windows

Governance layer:

  • Campaign-specific compliance officer
  • Audit log retained for 24 months
  • Right-to-erasure pipeline tested before launch
  • Pre-launch mock ECI inquiry to validate readiness

The five-cycle horizon

The single biggest mindset shift: AI infrastructure for Indian elections is not a single-campaign cost. It's a five-year investment that amortises across 4-6 election waves.

In a 5-year electoral cycle, a constituency typically sees:

  • 1 Lok Sabha election (main event)
  • 1 state assembly election (typically mid-cycle)
  • 1-2 by-elections (depending on incumbency dynamics)
  • 1-2 local body elections (Panchayat / Municipal)
  • Continuous governance helpline operation

A platform purchased for one of these and decommissioned after misses 80% of its potential value. A platform sustained across all of them produces a deep, longitudinal dataset on what voters in this constituency actually care about — which becomes a permanent strategic asset for the candidate and the party.

The campaigns that win 2029 will be the ones who saw 2024 as the first election in a 25-year arc, not the last election of a 25-year arc.

What Bhashini will likely deliver by 2029

Trajectory observations for the next 3 years:

  • Conversational quality parity with global frontier on Indian languages: likely by mid-2027
  • Dialect coverage for all major Indian dialects (Marwari, Bhojpuri, Awadhi, Magahi, Maithili, Kumaoni, Konkani, etc.): likely by end-2026
  • TTS naturalness matching commercial frontier: likely by mid-2027
  • Real-time latency matching commercial APIs: 2026-27 dependent on infrastructure investment
  • Open RAG / KB tools specifically for political/civic data: experimental in 2026, mainstream by 2027

By the 2029 Lok Sabha cycle, the Bhashini-based stack will likely be production-grade for most workloads. Commercial vendors will compete on premium quality, latency and integration depth rather than raw language capability.

Where AiSewak fits

AiSewak is a non-partisan, India-only voice agent platform built on top of the hybrid stack above. ECI-compliant defaults, TRAI-DLT integration, DPDP-residency, 22-language coverage with major dialect support, India-hosted infrastructure across multiple regions.

The platform is designed for the five-cycle horizon, not single-campaign cycles — the same infrastructure that runs the 2027 UP campaign continues operating through 2032 as a governance helpline, then re-activates for the next election cycle.

Where to go next

The Indian AI election stack is being built right now. The campaigns that learn to operate on it before mid-2027 will set the standards everyone else has to follow.

Frequently asked questions

What is Bhashini?

Bhashini is the National Language Translation Mission run by MeitY (Ministry of Electronics and IT). It provides open-source ASR (speech-to-text), MT (machine translation), TTS (text-to-speech) and language models for all 22 scheduled Indian languages, hosted on India-government infrastructure with permissive licensing.

Should we use Bhashini or global AI providers?

Hybrid in 2026. Global models (Gemini, Claude, GPT-4) still lead on raw conversational quality. Bhashini leads on data sovereignty, language coverage breadth, and DPDP-residency defaults. Most production deployments use both — Bhashini for some language workloads and sovereignty fallback, global models for others.

What makes Indian elections technically unique?

Three things: voter scale (96+ crore registered), language fragmentation (22 official + 270 dialects), and the five-phase electoral cycle (general + state + local) which means AI infrastructure built for one cycle gets reused across many.

Is the ECI advisory legally binding?

The 2024 ECI advisory has the force of guidelines under the Model Code of Conduct. Violations can trigger Model Code action including warnings, takedown orders and disqualification in severe cases. It's not statute but it's enforced.

Will the ECI update the rules before 2027?

Almost certainly. Expect a 2026 update extending labelling requirements to AI-generated images, possibly establishing a vendor-permission registry, and tightening cross-party AI-generated content rules. Campaigns should build to the strictest plausible 2027 standard, not the current 2024 minimum.