A serious AI voice agent for an Indian election campaign can be live in 30 days. The campaigns that have done this well in 2024 and 2025 followed a remarkably consistent week-by-week sequence. The campaigns that tried to compress this to 14 days uniformly failed — either the lists were dirty, the DLT templates were unapproved, or the dialect tuning was wrong on day one.
This guide is that 30-day sequence, written for the campaign manager who has signed the contract on Monday and needs lakhs of voters reached by next month.
Week 0 (Pre-kickoff) — the work that has to happen before Day 1
Three things must be in place before the 30-day clock starts.
1. Decision-making clarity. Who signs off on the system prompt? Who approves changes to the script mid-campaign? Who has authority to pause the campaign if it goes sideways? Without explicit decision authority, the team will stall on day 5.
2. Budget approved. A constituency-scale pilot costs ₹25–₹75 lakh depending on volume. A state-level rollout runs ₹2–₹8 crore. Have the budget approved before the team starts — half-funded pilots produce half-baked results.
3. Legal review scope. Identify the law firm or in-house counsel who will sign off on ECI compliance, DPDP, TRAI templates. Confirm they're available throughout the 30-day window.
If these three are not in place, postpone the 30-day clock. Don't try to do them in parallel with execution.
Week 1 (Days 1–7): foundation
Day 1: Kickoff and team alignment
- Two-hour kickoff meeting. Campaign owner, ops lead, AI vendor liaison, legal counsel.
- Define the goal: which constituency, what voter count, what 3-5 use cases (refer to the 12 use cases guide).
- Set up shared workspace (Notion / Confluence / Google Drive).
- Sign the NDA and DPA between campaign and AI vendor.
Day 2: Voter list sourcing
- Pull the ECI electoral roll for your constituency. Each AC has a downloadable PDF; large-format buyers can request the CSV via the CEO's office.
- Match against the National DND registry (TRAI maintains this — your AI vendor will have an API).
- Output: a cleaned, dedupe'd list with phone numbers in international format (+91…) and basic voter metadata (age band, gender, booth number, language).
Day 3: Knowledge base assembly
- Gather the campaign's source documents: manifesto, candidate biography, top 20 issues with proposed solutions, scheme awareness scripts.
- Convert to markdown — one file per topic. Aim for 50–150 KB total. This becomes the agent's RAG knowledge base.
- Identify gaps the candidate's team has not yet decided on. Mark them explicitly so the agent doesn't fabricate.
Day 4–5: Voice strategy
- Decide: candidate's cloned voice, or a generic Hindi (or regional language) voice?
- If candidate's voice: schedule a 90-minute recording session in a quiet studio. Provide the script (manifesto excerpts, 50 conversational utterances in target language, 20 emotional ranges — warm, serious, energetic).
- If generic voice: pick from your AI vendor's voice library. Test 3 voice IDs with 10 native speakers from the target dialect region. Choose the one with highest "feels trustworthy" rating.
Day 6–7: DLT template registration
- Draft 5–8 candidate templates for the AI agent's opening and closing lines.
- Submit to your telco's DLT platform. This is the longest-lead-time item.
- Most telcos turn around 5–7 business days for political templates. Don't delay this start.
End of Week 1: team aligned, voter list cleaned, KB assembled, voice chosen, DLT pending.
Week 2 (Days 8–14): build
Day 8: System prompt v1
- Write the system prompt in Hindi (or target language). Include:
- Identity: "मैं [Agent name] हूँ — [Candidate name] की AI सहायक..."
- Self-disclosure rule: must declare AI in opening
- HARD STOP rules: goodbye, anger, two silent turns, 90-sec time cap
- Allowed intents: which 3–5 use cases the agent handles
- Forbidden behaviours: no promises in candidate's name, no opponent attacks, no Aadhaar/UPI/bank collection
- Length: 600–1500 words. Get the legal counsel to review.
Day 9–10: Agent configuration
- Plug the system prompt into your AI vendor's console.
- Attach the knowledge base (RAG).
- Configure the voice ID and TTS settings (speed, stability).
- Set up telephony sender pool (rotating Indian numbers).
- Connect call logging to your dashboard.
Day 11–12: Pilot test (internal)
- Run 50–100 test calls. Volunteers within the campaign team and trusted karyakartas.
- Listen to 10 full call recordings. Note: stilted phrases, dialect issues, latency problems, hangups.
- Iterate the system prompt. Most agents need 3–5 iteration cycles before they sound right.
Day 13–14: Native-speaker testing
- Get 30–50 actual voters (not campaign team) to receive calls and rate the experience.
- Score: feels-natural (1–5), would-engage (1–5), trust (1–5), language-fit (1–5).
- Target: 4+ average across all four dimensions. Below 3.5 means more system-prompt iteration before launch.
End of Week 2: agent built, tested with real voters, scoring acceptable.
Week 3 (Days 15–21): scale-up
Day 15: DLT templates approved (hopefully)
- If approved: connect the templates to the agent's opening lines. Run a 100-call smoke test through the registered sender.
- If pending: chase the telco. This is the most common slip point.
Day 16–17: Compliance package
- Document everything for ECI compliance:
- System prompt (versioned)
- Voice clone consent form (if applicable)
- DLT template registration certificates
- DPDP-compliant data flow diagram
- Audit log access procedures
- Bind into a single PDF; share with legal counsel.
Day 18–19: Soft launch (1% volume)
- Day 18 morning: 500 calls
- Day 18 afternoon: 2,000 calls
- Day 19: 10,000 calls
- Monitor in real time: completion rate, hangup rate, sentiment, complaints.
- Have someone listen to 20 random recordings per day for the first week.
Day 20–21: Ramp to 10%
- 50,000 calls/day
- Watch for telephony spam-flagging (carriers may de-prioritise your sender pool if hangup rate is high).
- Watch for booths where completion rate is anomalously low — usually a dialect mismatch.
End of Week 3: live in production at 10% intended volume, telemetry healthy.
Week 4 (Days 22–30): full scale
Day 22–24: Ramp to 100%
- 5–10 lakh calls/day depending on constituency size.
- Daily war-room review: top issues mentioned by voters, top complaints about the agent, top booths needing extra attention.
Day 25–27: Wave 2 planning
- Based on Wave 1 sentiment data, identify the segments that need a follow-up persuasion call.
- Adjust system prompt for Wave 2: less introduction, more specific issue engagement.
- Schedule for T-30 to T-15 window.
Day 28–30: Stabilise
- Bug fixes and edge cases discovered in production.
- Documentation handover for the rest of the campaign team.
- Plan for Wave 3 (GOTV) which will run in the last 7 days.
End of Week 4: full-scale production AI voice campaign running, with capacity to iterate through 2–3 more waves before polling day.
What goes wrong (and how to recover)
Five most common failure modes, ranked by frequency.
1. DLT template approval delayed (Week 2–3)
Symptom: telco hasn't approved your templates by Day 14.
Fix: have a backup sender ID and template set with your vendor (most AI platforms maintain a pool of pre-approved generic templates that can be used as fallback). Use these for the soft launch while your custom templates are approved. Don't delay the 30-day timeline waiting.
2. Dialect mismatch (Week 2 testing)
Symptom: native speakers from a specific district rate the agent <3.5/5 on language-fit.
Fix: add dialect-specific examples to the system prompt. Have a karyakarta from that district record 10 reference utterances and feed them in. Re-test with the same group. Most dialect mismatches close within 2 iterations.
3. Voter complaints about being called (Week 3)
Symptom: 5%+ of voters explicitly say "stop calling me, I don't want this".
Fix: ensure the opt-out is being respected — voters marked DND should not be receiving calls. If they are, your DND scrub is broken. Investigate immediately; this is an ECI/TRAI risk.
4. Spam flagging by carriers (Week 3–4)
Symptom: connect rate falls from 65% to 35% within 48 hours.
Fix: rotate the sender ID pool. Reduce calls/hour per sender. Check that your sender numbers are properly DLT-registered. If still happening, escalate to the telco — usually a misconfiguration on their end.
5. Off-script behaviour (any time)
Symptom: the agent says something the campaign didn't approve. Could be a promise, an opinion on an opponent, a misstatement.
Fix: tighten the system prompt's forbidden-behaviours section. Add explicit examples. Increase model temperature down to 0.0–0.2 to reduce variance. Re-test.
What the campaign owner does in 30 days
For the campaign owner (not the ops or vendor staff), the time commitment is:
- Week 1: 4 hours (kickoff + decisions)
- Week 2: 6 hours (system prompt review + native testing review)
- Week 3: 6 hours (compliance review + soft launch monitoring)
- Week 4: 8 hours (war-room reviews + Wave 2 planning)
That's ~24 hours of campaign-owner time across 30 days — feasible alongside the rest of the campaign. The vendor and ops do the bulk of the day-to-day work.
What this scales to
A successful 30-day deployment for a constituency-scale pilot is the foundation for:
- Wave 2 and 3 persuasion calls in T-30 to T-7 window
- GOTV wave in T-7 to T-0 window
- Result-aftermath wave in T-0 to T+7 window
- Governance helpline starting T+15 and running for 5 years
If the 30-day deployment is solid, the rest of the cycle is incremental. If it's shaky, every subsequent wave makes the cracks bigger.
Where to go next
- Top 10 Features Every AI Election Campaign Agent Must Have — what to demand from your vendor
- AI Agent Compliance for Indian Elections — the legal layer
- The Two-Way AI Voter Engagement Engine — full architecture
- AI Election Agent Pricing — budget benchmarking
30 days is enough to launch. 60 days is enough to launch well. Most campaigns wait until 14 days, which is enough to fail.