Apply for AI Grants India

What Exotel does in a voice-agent stack

Exotel provides the telephony layer between a phone caller and your application. It can provision or connect business numbers, route inbound calls, originate outbound calls, manage call flows, and send events to your backend. Your application supplies the intelligence: speech recognition, dialogue logic, business-system actions, and synthetic speech.

This separation is important. Exotel does not replace your voice-agent runtime or guarantee that an LLM will produce a suitable conversation. It gives your agent dependable access to the Indian telephone network. If you are new to the category, start with what a voice agent is before choosing an architecture.

A production system normally contains:

• Telephony: Exotel numbers, call routing, recording controls, webhooks, and streaming.
• Conversation gateway: A low-latency service that authenticates events, maintains call state, and coordinates audio.
• Speech layer: Indian-language speech-to-text and text-to-speech models, including support for code-switching.
• Reasoning and tools: An LLM, deterministic policies, retrieval, and actions such as CRM updates or appointment booking.
• Operations: Metrics, transcripts, redaction, escalation, retries, and human handoff.

Reference architecture for Exotel and AI

For an inbound call, Exotel invokes your configured call-flow endpoint or applet. Your backend validates the request, identifies the customer where appropriate, and returns the next call instruction. For a real-time interaction, the call is connected to an audio stream, usually through a WebSocket-compatible integration, and your media gateway forwards audio to the speech pipeline.

The conversation loop should be designed as a streaming pipeline rather than a series of long blocking requests:

1. Receive and decode the audio stream.
2. Run voice activity detection and partial speech recognition.
3. Detect an utterance boundary and send the text, context, and permitted tools to the dialogue service.
4. Start TTS as soon as a safe response is available.
5. Stream the response back while allowing the caller to interrupt.
6. Persist only the events and data needed for support, analytics, and compliance.

Confirm the exact audio formats, WebSocket message schema, authentication method, concurrency limits, and availability of bidirectional media in your Exotel plan and current documentation. Treat these as implementation dependencies, not assumptions copied from an old sample project.

Inbound implementation checklist

Begin with a narrow call flow. A support agent might answer one request type, look up an order, and transfer unusual cases to a human. Configure a dedicated test number and keep production traffic separate until call completion, transfers, recording, and failure paths have been tested.

Your webhook service should:

• Verify Exotel signatures or request credentials where supported.
• Return quickly and move slow work to an asynchronous service.
• Store a call ID and correlation ID on every event.
• Handle duplicate, delayed, and out-of-order callbacks idempotently.
• Apply timeouts to STT, LLM, CRM, and TTS dependencies.
• Provide a fallback message and human-transfer route when the AI is unavailable.

Do not place API keys, customer data, or full transcripts in logs. Encrypt sensitive data in transit and at rest, define retention periods, and redact phone numbers, account identifiers, and financial information before sending traces to an observability platform.

Outbound campaigns and lead qualification

For outbound calling, your application typically selects an eligible record, invokes Exotel’s call API, and listens for answer, failure, busy, and completion events. The agent should not begin its script until the call is actually answered. It should also identify itself clearly, state the purpose of the call, and offer a simple way to opt out or reach a human.

Use a campaign service rather than firing calls directly from a CRM button. The campaign service should enforce rate limits, calling windows, retry rules, duplicate suppression, agent concurrency, and consent status. Store a structured outcome such as interested, callback_requested, not_interested, wrong_number, or needs_human_review instead of relying only on a transcript.

Before scaling outbound traffic, verify DLT registration, approved templates or headers where applicable, consent records, preference checks, and Exotel’s current policy for the intended traffic. Do not assume that a DND check alone makes every campaign compliant; legal and telecom requirements depend on the use case and message type.

Latency, interruptions, and call quality

Call quality is determined by the complete path: handset to carrier, Exotel to your media gateway, STT, reasoning, TTS, and the return route. Measure each segment separately. A useful initial target is a short perceived response delay, but the right threshold depends on language, turn length, and task complexity.

Practical improvements include:

• Deploy the gateway close to Indian telephony and cloud regions, then measure rather than assume network proximity.
• Use streaming STT and incremental TTS instead of waiting for an entire response.
• Keep the first response brief; gather details over multiple turns.
• Use VAD, endpointing, and barge-in support so callers can interrupt naturally.
• Cache stable prompts and frequently used audio where appropriate.
• Route simple intents through deterministic code instead of an LLM.
• Set a maximum silence duration and recover with a concise reprompt.

Track time to first transcript, time to first audio, total turn latency, barge-in rate, transfer rate, hang-ups, recognition confidence, and task completion. These metrics are more actionable than average call duration alone.

Indian languages, Hinglish, and conversation design

Telephony audio is compressed and often noisy. Test speech recognition with real accents, regional vocabulary, names, numbers, addresses, and code-switching. Hindi-English conversations may change language mid-sentence; Kannada, Tamil, Telugu, Bengali, Marathi, and other languages also require separate evaluation rather than a single “Indian language” score.

Use explicit language detection only when it improves the experience. A short opening question can be more reliable than guessing from the first few words. Keep prompts conversational, confirm critical values digit by digit, and repeat names or addresses using phonetic clarification. For regulated or high-risk actions, require confirmation and provide a human route.

A provider such as Sarvam AI, Bhashini-connected services, Google, Azure, or another specialist may fit different languages and latency targets. Benchmark providers on your own recordings and obtain the required permissions before using calls for model evaluation or training.

Security, privacy, and compliance

Before launch, document what data is collected, why it is collected, where it is processed, who can access it, and when it is deleted. Tell callers when interactions are recorded or assisted by AI where required by your policy and applicable law. Obtain appropriate consent for recording and promotional communication, and consult qualified Indian telecom and privacy counsel for your traffic model.

Minimum controls should include:

• Role-based access to recordings, transcripts, and customer profiles.
• Encryption, key rotation, secrets management, and audit logs.
• PII redaction in transcripts and operational logs.
• Retention rules for recordings, prompts, tool outputs, and backups.
• Prompt-injection and tool-authorization controls.
• Human approval for refunds, financial commitments, medical decisions, or account changes.
• A documented incident process for leaked data, abusive calls, and incorrect actions.

Healthcare deployments need additional safeguards; review the considerations in HIPAA-compliant voice agents for hospitals, while recognising that Indian obligations still require local legal review.

Testing and production rollout

Test the whole call, not only the LLM prompt. Create scripted and adversarial scenarios for silence, interruptions, background noise, wrong numbers, angry callers, ambiguous dates, language switching, API timeouts, duplicate callbacks, and transfer failure. Have native speakers review a representative sample of calls.

Launch in stages:

• Prototype: one number, one workflow, synthetic or consented test data.
• Pilot: limited traffic, mandatory human escalation, daily transcript review.
• Controlled production: monitor cohorts, cap concurrency, and compare against a human baseline.
• Scale: automate regression tests, alerting, cost controls, and incident response.

Budget for more than Exotel minutes. Total cost includes telephony, phone numbers, streaming infrastructure, STT, TTS, LLM tokens, storage, monitoring, human transfers, and engineering maintenance. A voice-agent pricing and ROI framework helps structure this estimate.

Choosing build versus integration support

Build in-house when your team can operate low-latency media services, secure customer data, evaluate Indian languages, and maintain telecom workflows. Use an experienced integrator when you need a faster pilot, complex CRM or contact-centre integration, or production support across multiple languages. Compare providers on call quality, ownership of prompts and data, observability, portability, escalation design, and incident response—not just quoted per-minute pricing. For a broader shortlist, see voice agent services for Indian businesses.

The strongest Exotel deployments are deliberately narrow at first: one measurable business outcome, one tested language set, clear fallback rules, and evidence from real calls. Expand only after the agent can complete the task reliably, explain its limits, and hand control to a person when the conversation demands it.

FAQ

Does Exotel replace an AI voice platform?
No. Exotel supplies telephony and call-control capabilities. You still need speech, dialogue, tools, security, monitoring, and escalation.

Can I connect my own LLM?
Yes, provided your architecture can receive call events or audio, produce responses within acceptable latency, and meet your privacy and reliability requirements.

Is WebSocket streaming always available?
Streaming depends on the Exotel product, account configuration, and current API capabilities. Confirm supported media direction, codecs, limits, and billing before committing to the design.

How should I handle Hinglish?
Evaluate speech providers using real, consented samples. Add language-aware prompts, confirmation for names and numbers, and a human fallback for low-confidence recognition.

What should I measure after launch?
Track successful task completion, transfer and hang-up rates, time to first audio, recognition errors, latency by dependency, cost per completed task, complaint rate, and customer feedback.