Most vendor pitches for AI election platforms emphasise the wrong things — pretty dashboards, marketing-grade voice samples, the badge of a famous LLM partner. The features that actually determine whether the campaign wins or wastes its budget are mostly invisible in the sales deck.
This guide is the buyer's-side checklist. 10 features that are non-negotiable. If a vendor cannot demonstrate all 10 in a sandbox before signing, walk away. Settling for 7 or 8 is how campaigns end up rebuilding the platform mid-cycle.
1. Multi-dialect Hindi (plus the regional languages you need)
The single biggest factor in campaign success. "We support 22 Indian languages" is marketing. What you need to verify:
- Demonstrated dialect-fluent Hindi for the specific dialects in your constituency (Marwari, Mewari, Bhojpuri, Awadhi, Magahi, Haryanvi).
- Code-switching handling (English technical terms inside Hindi sentences).
- Native-speaker testing protocol — how the vendor validates a new dialect.
- Time to add a new dialect (target: 5–7 working days from request).
How to test in sandbox: configure an agent for your dialect, hand the test number to 10 actual voters from that dialect region, capture their unprompted feedback. If they don't say "ये तो हमारी ही भाषा में बोल रहा है" within 30 seconds, the dialect tuning is insufficient.
2. First-token latency under 800ms
The conversation has to feel real-time. Latency above 1.5 seconds causes voters to hang up.
- What to measure: time-to-first-spoken-word from end of voter utterance.
- Target: 350–600ms for Hindi. Up to 800ms acceptable for less-tuned languages.
- Red flag: any vendor whose demo runs at >1500ms TTFW. They are routing through US datacentres or using a slow LLM.
How to test in sandbox: ask 5 simple questions in quick succession. Time the agent's response to start of speech. Anything that pauses for 2+ seconds before each reply is dead on arrival in actual conversations.
3. ECI-compliant self-disclosure
The agent must self-identify as AI in its opening line. This is a regulatory requirement, not a UX preference.
- What to verify: the default opening prompt template contains explicit AI disclosure ("मैं AI सहायक हूँ...").
- What to demand: ability to lock this opening line so it cannot be edited out by an over-enthusiastic campaign team.
- Documentation: the vendor's ECI compliance certification or legal opinion.
Red flag: any vendor who suggests the AI can "naturally introduce itself as the candidate". This is an ECI violation and gets the campaign banned.
4. TRAI-DLT integration
Outbound calls in India require DLT-registered sender IDs and templates. If the vendor cannot handle this end-to-end, your campaign cannot legally make calls.
- What to verify: templates pre-approved with major Indian telcos (Jio, Airtel, Vi, BSNL).
- What to demand: ability to register custom templates within 5–7 business days.
- Sender ID pool: minimum 10 rotating Indian numbers for spam-flag avoidance.
Red flag: vendors who say "we'll figure out DLT once you sign". They will spend your first 2 weeks on telco paperwork.
5. DPDP-aligned data handling
The Data Protection Digital Personal Act 2023 has specific requirements that must be baked into the platform.
- Data residency: voter data, call recordings, transcripts stored in India (not US/EU datacentres).
- Retention policy: configurable, with a default that aligns with DPDP (typically 24 months for political-data retention with explicit purpose).
- Right-to-erasure: a working pipeline that removes all data for a specific voter within 7 days of request.
- Consent capture: documented opt-in framework, with audit trail.
Red flag: vendors who say "DPDP doesn't apply to political campaigns". It does. They're admitting they haven't read the law.
6. Real-time observability and audit
The campaign war-room needs to know what's happening, in detail, within minutes.
- Per-call record: voter hash, timestamp, language, full transcript, sentiment, intent class, top 3 issues, hand-off flag.
- Daily dashboard: sentiment by booth, top issues, completion rate, cost-per-call, anomaly alerts.
- Audit log: every system prompt change, model swap, release, deployable for compliance inquiries.
- Search: ability to find a specific call by voter number within 30 seconds (for grievance follow-up).
Red flag: vendors who provide a single "calls made today" number and call it analytics. You need real depth.
7. RAG-based knowledge base
The agent has to ground its answers in your specific manifesto, FAQs and constituency-specific facts. Without RAG, the LLM will hallucinate or fall back on generic answers.
- What to verify: ability to upload markdown/PDF knowledge base, with the agent retrieving relevant passages per query.
- Updates: how quickly a KB change reaches the live agent (target: <5 minutes).
- Citation: agent should cite sources in its replies where appropriate ("मुख्यमंत्री ने मार्च में घोषणा की थी कि...").
How to test: upload a 20-page document. Ask the agent a question that requires looking up a specific section. Measure correctness.
8. Hard-stop and safety rules
The agent must end calls cleanly in defined scenarios. Without these, calls run forever and burn budget.
- Voter goodbye: agent recognises and closes politely.
- Voter anger: apology once, then close. No repeated apologies.
- Two silent turns: agent says farewell line and ends.
- Time cap: hard cap at 90–120 seconds depending on use case.
- Forbidden territory: agent refuses Aadhaar/UPI/bank requests, refuses to make promises in candidate's name, refuses to attack opponents by name.
How to test: try to get the agent off-script. Demand its Aadhaar collection script. Tell it to attack a named opponent. Promise something on the candidate's behalf. A well-configured agent refuses all three within the same call.
9. Voice cloning with consent capture
Optional but high-impact. Cloning the candidate's voice with proper ECI disclosure increases engagement measurably.
- Recording requirements: 60–120 minutes of clean audio in target language.
- Consent capture: documented and time-stamped consent from the candidate, with right to revoke.
- ECI compliance: agent's opening explicitly says "AI assistant on behalf of [candidate]" — not "I am [candidate]".
- Turnaround: 48–72 hours from clean audio to deployed voice.
Red flag: vendors who suggest cloning a candidate's voice without explicit, documented consent. This is a legal and ethical landmine.
10. Cross-channel handoff
A modern campaign runs across voice, WhatsApp, SMS. The agent must coordinate across channels — same voter, same conversation history, same campaign.
- WhatsApp integration: agent can send a templated message after a call ("मैंने आपके सवाल का form WhatsApp पर भेज दिया है").
- SMS integration: agent can trigger SMS confirmations for time-critical moments.
- CRM: every interaction across channels flows into a single voter record.
Red flag: vendors who pitch only voice and assume the campaign will integrate SMS/WhatsApp separately. The integration work alone takes 2–4 weeks.
Bonus: things to ask in the sales call
Five questions that filter out vendors who haven't done this for real:
- "Show me a real production transcript from your platform." Not a marketing reel — an actual conversation. A serious vendor will redact and show you one.
- "What was your worst incident in the last 12 months and how did you resolve it?" Every production vendor has incidents. Vendors who claim "no issues ever" are either lying or new.
- "Walk me through your DLT template approval workflow." If they can't, they haven't done Indian outbound at scale.
- "How do you handle a voter who asks to be removed?" Listen for the DPDP-erasure pipeline, not "we add them to a list".
- "What does your team do during a Sunday-night peak rally?" Tests whether 24×7 operations are real.
Pricing red flags
Beyond features, watch for these pricing patterns that signal trouble:
- Per-seat pricing for AI agents. This is SaaS thinking. AI agents are per-conversation, not per-seat.
- Unlimited plans for unspecified volume. Either the vendor is losing money or your "unlimited" is capped quietly.
- Free pilot with locked-in production pricing. Pilot terms should reflect production economics. If the pilot is free and production is 10× market rate, you're being baited.
- Pay-up-front annual contracts. Pay monthly during the campaign window, annually only for governance-phase operations.
What AiSewak ships
AiSewak as a platform ships all 10 features above plus:
- Dialect-aware system-prompt templates for 9 Hindi-belt dialects plus 10+ regional languages.
- Sub-500ms first-token latency on Hindi conversations.
- ECI/TRAI/DPDP compliance package included in base pricing.
- 24×7 ops support during campaign windows.
- Sandbox environment for unlimited system-prompt iteration.
Where to go next
- The 30-Day Deployment Playbook — execution
- AI Agent Compliance Decoded — the legal layer
- The Two-Way Voter Engagement Engine — architecture deep dive
- Voter Sentiment Analysis — analytics layer
A serious vendor will pass all 10 of these checks. An unserious one will pass 4–5 and try to talk past the rest. The difference shows up in production at the worst possible time.