Cluster B · Pricing

AI Election Agent Pricing: Cost-Per-Conversation Across Platforms in 2026

What does an AI voice agent actually cost in an Indian election campaign? Full pricing breakdown — LLM tokens, TTS minutes, STT, telephony, dialect tuning, voice cloning. Vendor benchmarking framework.

7 min readUpdated 22 May 20261,432 words

Vendor pricing for AI election voice agents in 2026 is a mess. Two vendors with similar-looking pitches can have 3× cost difference for the same workload. The pricing models are designed to be hard to compare on purpose — bundling, peak surcharges, voice-clone markups, regional language surcharges, dashboard fees.

This guide is the de-jargoned breakdown. The cost components, the typical ranges, what to negotiate, and how to spot the vendors who are overcharging.

The cost components

A single voter conversation has 5 underlying costs. Vendors bundle these differently. Make them itemise.

1. LLM tokens (the brain)

The model that processes the voter's words and generates the response. Per call:

  • System prompt + RAG context: ~3,000–6,000 input tokens
  • 4–8 turns of conversation × ~150 input tokens each = ~1,000 input tokens
  • Generated responses × 4–8 turns × ~80 tokens each = ~500 output tokens

At 2026 rates (Gemini 2.5 Flash, Claude Haiku 4.5, GPT-4.1-mini, Qwen 3 30B A3B):

  • Input: ~$0.10–$0.30 per million tokens → ~₹0.04 per call
  • Output: ~$0.40–$1.20 per million tokens → ~₹0.05 per call

Total LLM cost per call: ~₹0.05–₹0.15.

2. STT (speech-to-text)

The model that transcribes voter speech. Streamed across the whole call duration.

  • Cost: typically $0.005–$0.015 per minute of streaming
  • 60-90 second call: ~₹0.10–₹0.20

3. TTS (text-to-speech)

The model that generates the agent's audio responses. Charged per minute of generated speech.

  • Cost: $0.10–$0.25 per minute of generated speech
  • Agent speaks ~40–60% of the call (voter speaks the rest)
  • 60-second call: ~₹0.30 of TTS
  • 120-second call: ~₹0.60

TTS is the single largest variable cost. Reducing call length 10% saves 10% on TTS specifically.

4. Telephony

Carrier costs for outbound calls in India.

  • Per-minute outbound: ₹0.20–₹0.50 (depends on carrier deal, sender pool, volume tier)
  • 60-second call: ₹0.20–₹0.50

5. Platform / operations markup

The vendor's margin on top of raw infrastructure costs.

  • Healthy markup: 30-50% on top of raw costs
  • Predatory markup: 100-200% on top of raw costs

Most enterprise vendors charge somewhere in between. The best clue is whether they'll itemise — vendors with healthy markups will; vendors with predatory markups won't.

The math: what a "₹1 per call" headline actually means

A 75-second average Hindi call costs:

ComponentCost
LLM (Gemini Flash)₹0.10
STT (Deepgram-equivalent)₹0.15
TTS (45s of audio at ₹0.20/min)₹0.35
Telephony (75s at ₹0.30/min)₹0.40
Raw cost₹1.00
Platform markup (40%)₹0.40
All-in cost₹1.40

That's roughly the upper bound of competitive pricing. Vendors who quote ₹2-3 per call for the same workload are taking 100%+ markup or using inefficient infrastructure.

Vendors who quote ₹0.40 per call for the same workload are either subsidising (loss-leader pilot), shortening calls to 30s (UX collapses), or using lower-quality models (engagement drops).

Volume tiers and bulk discounts

Pricing should drop with volume. Reasonable tiers:

VolumePer-call rate (Hindi)
<50,000 calls (pilot)₹1.50–₹2.00
50K–5 lakh calls (constituency)₹1.00–₹1.50
5 lakh–50 lakh calls (state campaign)₹0.80–₹1.20
50 lakh+ calls (national)₹0.60–₹0.90

Bulk discounts come primarily from carrier economics (telephony rates drop with volume commits) and amortised platform costs. LLM/STT/TTS are roughly volume-flat — these are paid per unit.

Regional language markup

Most vendors charge a 15–30% markup for regional language conversations. Reasons:

  • Smaller LLMs less optimised for languages other than Hindi/English
  • Smaller pool of TTS voices, sometimes premium-tier voices
  • Higher engineering cost for dialect tuning

If your campaign is heavily regional (Tamil, Bengali, Marathi), build this into the budget. Don't accept a quote for "all 22 Indian languages at the same price" — that vendor is either subsidising or hasn't actually tested regional quality.

Voice cloning costs

If the campaign wants the candidate's voice:

  • One-time setup: ₹50K–₹3 lakh depending on quality, fidelity, multilingual support
  • Per-call markup: ₹0.05–₹0.15 added to base rate (cloned voices are computationally heavier)
  • Refresh/iteration: occasional re-tuning runs ₹25K–₹75K each

For a 50-lakh-call campaign, voice cloning adds ₹3-8 lakh to total cost. Whether it's worth it depends on the candidate — a well-known voice that voters recognise drives significantly higher engagement.

Peak concurrency

Most vendors set a default concurrency cap (e.g., 5,000 simultaneous calls). Above the cap, two patterns:

  • Queue and retry: cheaper but adds latency, voters get calls minutes later
  • Burst pricing: 1.5–3× markup during peak windows

For rally days, polling-day GOTV waves and pre-poll weekends, plan for peak concurrency in the contract. A flat per-call rate with built-in peak handling is the cleanest negotiation.

Hidden costs to watch for

Five line items that vendors often hide:

  1. Setup / onboarding fee. Reasonable: ₹5-15 lakh. Predatory: ₹50+ lakh. The setup work is real (DLT registration, KB ingestion, system prompt tuning) but should be capped.
  2. Dashboard / analytics subscription. Some vendors price the dashboard separately. Insist that dashboard access is included in the per-call rate.
  3. Call recording storage. Storage is ~₹0.01 per call per month. Vendors who charge ₹0.10+ per call for storage are gouging.
  4. Voter list ingestion charges. Reasonable: free for up to 10 cr voters. Predatory: ₹0.05 per voter ingested.
  5. War-room support fees. Reasonable: included for 4-8 hours of operations support per week. Premium hours billable.

How to benchmark vendors

A simple framework. For each vendor under consideration:

  1. Demand itemised pricing. Vendors who refuse are signalling they're overcharging.
  2. Calculate raw cost from itemisation. Compare to the ~₹1 raw cost we calculated above.
  3. Compute platform markup. Anything above 70% markup is a yellow flag.
  4. Test the demo. Run 10 calls with your dialect, your KB, your system prompt. Measure latency, completion, satisfaction. Bad demos at premium prices are a hard no.
  5. Check references. Talk to 2 prior campaigns that ran with this vendor. What did they actually pay? What broke?

Sample budget: 50-lakh-voter Lok Sabha constituency

Working from the per-call costs above, a complete campaign budget:

Voice AI:

  • Wave 1 (T-90 to T-60): 50 lakh × 1 call × ₹1.20 = ₹60 lakh
  • Wave 2 (T-60 to T-30): 35 lakh × 1 call × ₹1.20 = ₹42 lakh (target undecided + persuadable)
  • Wave 3 (GOTV T-7 to T-0): 20 lakh × 1 call × ₹1.50 (higher peak rate) = ₹30 lakh
  • Inbound helpline: 6-month operations × ₹2 lakh/month = ₹12 lakh

Total voice AI: ~₹1.44 cr

Adjacent costs:

  • DLT registration: ₹2 lakh
  • Voice cloning: ₹5 lakh
  • WhatsApp Business API (50 lakh templated messages × ₹0.25): ₹12.5 lakh
  • SMS (50 lakh × ₹0.15): ₹7.5 lakh
  • Dashboard, analytics, integrations: ₹8 lakh

Total AI campaign cost: ~₹1.80 cr

For comparison, the equivalent human call-centre coverage (3 waves × 50 lakh × ₹5 per call) would be ₹7.5 cr — over 4× more, with worse audit trail and lower throughput.

What ChatGPT-equivalent direct-to-vendor costs would be

Some campaigns ask: "can we just build this on top of OpenAI / Anthropic / Google directly and skip the platform fee?"

Theoretical costs (raw model + telephony, no platform):

  • LLM + STT + TTS + Telephony: ~₹0.70 per call
  • Engineering team to build the orchestration, observability, integrations: ~₹3-5 cr/year fixed cost
  • Compliance work (DLT, ECI, DPDP): ~₹50 lakh of legal + ops setup

For a single campaign of <2 cr call volume, the engineering team alone overwhelms the per-call savings. Direct-to-LLM only makes sense for permanent infrastructure used across many cycles — and at that point, you've become a vendor.

Where AiSewak fits

AiSewak's pricing model:

  • ₹1.00–₹1.30 per Hindi conversation, all-in (LLM + STT + TTS + telephony + platform)
  • 15% markup for regional languages
  • No setup fee for campaigns >5 lakh calls
  • No dashboard / analytics fee
  • Voice cloning ₹1.5 lakh one-time + ₹0.10/call markup
  • Volume discounts: 15% off above 25 lakh calls, 25% off above 1 crore calls

Pricing is itemised by default; we want you to audit it.

Where to go next

Cost matters because the same budget allocated differently produces different election outcomes. A clean per-call price comparison protects you from vendors who are betting you won't do the math.

Frequently asked questions

What's the all-in cost per voter conversation?

₹0.50–₹1.50 inclusive of LLM + STT + TTS + telephony + platform. The variance comes from call duration (45s vs 120s), language (Hindi cheapest, smaller languages more expensive), peak concurrency, and platform markup.

Why does call duration matter so much?

TTS minutes are the largest variable cost. A 60-second call costs ~₹0.45 in TTS alone; a 120-second call costs ~₹0.90. Most campaigns over-engineer the conversation length — shorter calls with tighter scripts cost less and engage better.

Should we negotiate per-call or per-minute pricing?

Per-conversation is what you actually consume. Per-minute is what vendors prefer because it lets them earn on longer calls. Insist on per-conversation pricing with a max-duration cap; this aligns incentives.

What about peak concurrency surcharges?

Some vendors charge a multiplier (1.5-3×) during peak windows. Rally days and pre-poll weekends spike concurrency. Negotiate a flat per-conversation rate that includes peak handling, with a reasonable concurrency cap (10,000-50,000 simultaneous calls).

Are voice cloning costs separate?

Usually yes. Voice cloning is a one-time setup fee (₹50K–₹3 lakh depending on quality and vendor) plus a small per-call markup (typically ₹0.05–₹0.15 added to base rate). Budget for both.