Most enterprise buyers start voice streaming discovery by comparing model quality. Which speech engine transcribes fastest, which text-to-speech voice sounds least synthetic, which large language model reasons best. Those questions matter. They also explain almost none of the failures that show up eight weeks after a signed contract. The demo works because the demo runs on clean WiFi, in one language, with one concurrent call, no CRM lookup, no compliance officer, and no telco in the path.
Production is a different animal. A realtime voice deployment touches carrier routing, SIP interconnects, media streaming, turn detection, agent desktops, recording archives, and a dozen regulatory expectations. Failures happen in the seams between those layers, not inside the AI models everyone spent evaluation time on.
What follows is a diagnostic: nine ways realtime voice deployments break after the demo, and for each one, the question that would have surfaced it during evaluation.
The Production Gap: Why Realtime Voice Demos Pass and Enterprise Rollouts Stall
A vendor demo is a controlled experiment designed to isolate the model. Your production environment is uncontrolled by definition. It has to serve customers on 2G-era mobile networks, in code-switched speech, at 4 PM on collections day when 900 calls are dialing at once.
The gap between those two conditions is where budgets die. Teams discover mid-rollout that the voice API streams audio beautifully but cannot originate calls on a local number in a particular telecom circle. Or that latency measured at the model is 400 ms while latency measured at the customer’s ear is 1.4 seconds. Or that a warm transfer to a human agent works only if the human sits inside the same vendor’s contact center, which the enterprise does not use.
Good discovery treats the stack as one system: telephony, media transport, AI, agent handoff, and audit trail. Evaluate any single layer in isolation, and you get a confident purchase followed by a stalled rollout.
Failure 1: The Call Never Connects. Carrier Routing, Circle Coverage, and Virtual SIP Gaps in India
The first production failure is the least glamorous one. Calls do not connect, connect to the wrong region, or land with a number the customer does not recognize and will not answer.
In India, this is a circle-level problem. Number availability, routing quality, and answer rates vary by telecom circle, and a provider sitting on top of somebody else’s telephony inherits whatever coverage that upstream partner happens to have. Virtual SIP and cloud number provisioning look identical on a feature grid until you need 4,000 local numbers across multiple circles with consistent CLI presentation and DLT-compliant outbound. In the Gulf, the equivalent question is whether the provider holds licensed local number infrastructure under the relevant telecom authority, or routes calls through an offshore gateway.
Answer rates are a business metric. A three-point drop in connect rate across a collections portfolio outweighs almost any model-quality advantage.
Ask during evaluation: Which telecom circles or countries do you hold direct carrier relationships in, who owns the numbers we will use, and what connect and answer rates do comparable customers see in our geographies?
Failure 2: The Awkward Pause. Where Latency Really Accumulates Across a Voice Streaming API
Vendors quote model latency. Customers experience conversational latency, the full round trip from the moment a caller stops speaking to the moment they hear a response.
That round trip accumulates across at least seven stages: carrier leg to the media server, packetization and buffering, endpointing (deciding the caller has finished), streaming transcription, LLM inference including any tool calls, speech synthesis, and the return leg back through the carrier. A 200 ms model plus a 200 ms TTS plus a 600 ms CRM lookup plus a sleepy 400 ms endpointer produces a pause long enough for the customer to say “hello?” and derail the turn.
Geography compounds it. When media servers sit on a different continent from the caller, every packet pays that tax twice. Measure latency at the handset, on a mobile network, in the market you will actually serve. Exotel designs its streaming layer for sub-300 ms voice latency, and it can hold that number because the media path and the telecom layer are not separate vendors handing packets to each other.
Ask during evaluation: Give us a per-stage latency breakdown measured end-to-end on a live PSTN call in our primary market, at p50 and p95, with a CRM lookup in the loop.
Failure 3: The Bot Talks Over the Customer. Barge-In, VAD, and Turn-Detection Weaknesses
Turn-taking separates a voice agent people tolerate from one they hang up on. It rests on voice activity detection, endpointing thresholds, and barge-in handling that lets a caller interrupt mid-sentence and have the agent stop cleanly.
Real calls break naive implementations fast. Noise on a factory floor or a busy street triggers false speech detection. A customer pausing to find their policy number gets cut off after 500 ms of silence. Someone says “no, wait” and the bot ploughs through the rest of its script. Echo from a speakerphone gets transcribed as customer speech, and the agent answers itself.
These are tuning problems with a hardware and network dimension. They are also among the least visible items in most voice streaming comparisons, because vendors rarely demo on bad lines.
Ask during evaluation: How is barge-in implemented, what happens on speakerphone and in high-noise environments, and can endpointing thresholds be tuned per use case and per language?
Failure 4: The Transcript Falls Apart in Hinglish and Arabic. Multilingual and Code-Switching Breakdowns
A borrower in Pune does not speak Hindi or English. They speak both inside the same sentence, with English numerals, Hindi verbs, and product names in a third language. Speech systems trained on monolingual corpora handle each language acceptably and the mixture badly.
Word error rate on a clean single-language benchmark tells you very little about what happens with Hinglish, with Gulf Arabic dialects that diverge sharply from Modern Standard Arabic, or with Bahasa and Tagalog code-switching. Watch entity accuracy rather than overall accuracy. Getting “four lakh twenty thousand” or an alphanumeric application ID wrong matters more than a misheard filler word, because those errors flow straight into API calls and payment amounts.
Language detection and mid-call language switching are separate capabilities again. If a customer answers in Hindi and then switches to English, the agent should follow rather than restart.
Ask during evaluation: Show us transcription and intent accuracy on recordings of our own customers, including code-switched audio, and explain how mid-conversation language switching is handled.
Failure 5: The Cold Transfer. Handoff That Loses Context and Forces the Customer to Repeat
The most common complaint about voice bot use cases in production is the handoff. The AI collects a name, a policy number, and a reason for calling, then transfers the customer to a human who opens with “Hello, how can I help you today?”
Every repeat question erases the goodwill automation earned. A proper warm transfer carries the transcript, the extracted entities, the detected intent, the sentiment trajectory, and any actions already taken, and it puts all of that on the agent’s screen before the agent says hello. It should route by skill and language too, so a customer who has been speaking Arabic does not land with an English-only agent.
Here the architecture question becomes unavoidable. If the AI agent and the contact center are separate products from separate vendors, context transfer depends on an integration someone has to build and maintain. Where conversational context lives in a persistent memory layer shared across bots, agents, and channels, the handoff carries state by default rather than by custom code.
Ask during evaluation: Walk us through a live transfer from AI to a human agent on our own contact center software, and show us exactly what appears on the agent’s screen at the moment of connection.
Failure 6: The Audit Request Nobody Can Answer. Consent, Recording, and Script Adherence Gaps
Six months in, compliance asks for every outbound collections call made to a specific borrower in March, with the consent record, the full recording, the transcript, and evidence that mandatory disclosures were read in the correct language. If that request takes three weeks and a data engineer, the deployment has a problem.
Regulated outbound carries obligations that vary by market: the Reserve Bank of India’s Fair Practices Code and digital lending expectations, OJK rules in Indonesia, BSP guidance in the Philippines, telecom authority requirements in the UAE. No platform makes you compliant. A good one makes compliance provable, which means consent capture as a first-class event, tamper-evident recording with defined retention, encryption and role-based access, and automated scoring of script adherence instead of a manual 2% sample.
Automated quality analysis changes the economics. Scoring 100% of conversations for disclosure language, prohibited phrases, calling windows, and tone means exceptions surface in days rather than at audit time.
Ask during evaluation: Demonstrate retrieving a complete audit package for a single call, and show how consent, script adherence, and calling-window rules are enforced and evidenced at scale.
Failure 7: Scale Day. Concurrency Ceilings, Burst Traffic, and Dialer Pacing
Pilots run at 20 concurrent calls. Production runs at 2,000, concentrated into the four hours when customers actually answer.
Concurrency ceilings hide in unexpected places: media server capacity, trunk limits with the upstream carrier, ASR provider rate limits, LLM token throughput, per-account API quotas. Systems degrade differently depending on which ceiling is hit first. Choppy audio, rising latency, silent call failures, and transcription that arrives after the caller has hung up are all symptoms of the same underlying constraint.
Dialer behaviour is the other half. Predictive pacing tuned for human agents does not translate cleanly to AI agents, and retry logic that ignores regulatory calling-window rules runs straight into them. Retry strategy has to account for time of day, prior outcomes, and consent status.
Ask during evaluation: What is our contracted concurrency, what happens at 3x that number, and how does the platform degrade rather than fail? Ask for load-test results, not a capacity claim.
Failure 8: The Integration Tax. CRM, Payments, and Real-Time Action Latency
A voice agent that can talk but cannot act is an expensive IVR. The value in voice bot use cases comes from doing things mid-call: fetching an outstanding balance, generating a payment link, rescheduling a delivery, verifying a KYC field, logging a disposition.
Every one of those actions costs time inside the conversational turn. A CRM that responds in 900 ms turns an otherwise fast agent into a slow one. The design answer is a mix of pre-fetching known data before the call connects, streaming filler speech while a call is in flight, and setting hard timeouts with graceful fallbacks. None of that happens automatically.
Build-versus-configure also drives timelines. Pre-built connectors into common CRMs and helpdesks, open REST APIs, and a no-code builder shorten the path from pilot to production considerably compared with writing every integration from scratch. Exotel ships 150+ pre-built integrations for exactly this reason.
Ask during evaluation: Which of our systems do you already integrate with, what is the measured added latency of a live API call inside a turn, and what does the agent say while it waits?
Failure 9: Finger-Pointing Between the Bot Vendor, the CCaaS Vendor, and the Telco
The nine-month failure is organizational. Call quality degrades on Tuesday afternoons. The AI vendor points at the telco. The telco points at the contact center platform. The contact center vendor points at the AI vendor’s media handling. Nobody owns the customer’s experience, and your team becomes the integration layer.
Fragmented stacks also fragment data. Interaction records live in three systems with three timestamps and three definitions of a “call,” which makes containment rate and cost-per-interaction genuinely hard to measure. Reporting arguments turn into vendor arguments.
Consolidation is the structural fix. When AI agents, contact center, and the telecom network run on one architecture, there is one root-cause analysis, one SLA, and one set of numbers. That is the practical case for treating cloud telephony discovery as part of AI evaluation rather than a separate procurement exercise conducted a year earlier.
Ask during evaluation: In a voice quality incident spanning telephony and AI, who runs root-cause analysis, what is the single SLA that covers the whole path, and can you show a postmortem from a real incident?
Turning Failure Modes Into a Voice Streaming Discovery Checklist
Compress the nine failures into a scoring sheet and apply it consistently across vendors. Weight the items by what your business actually risks.
- Telephony ownership: Direct carrier relationships, circle and country coverage, number provisioning, licensed local infrastructure, and connect-rate benchmarks in your markets.
- Measured latency: Per-stage breakdown at p50 and p95, on live PSTN calls, with an integration in the loop, measured in your region.
- Conversation mechanics: Barge-in, noise resilience, tunable endpointing, echo handling, and behaviour on speakerphone.
- Language reality: Accuracy on your own recordings, code-switching, dialect coverage, entity-level precision, and mid-call language switching.
- Handoff quality: Warm transfer with full transcript and context, skill and language routing, and supervisor visibility across concurrent AI conversations.
- Compliance evidence: Consent capture, audit-ready recording, retention and access controls, automated script-adherence scoring, and alignment with the regulations governing your market.
- Scale behaviour: Contracted concurrency, burst headroom, degradation profile, load-test evidence, and dialer pacing with compliant retry logic.
- Integration surface: Existing connectors, open APIs, added turn latency, fallback behaviour, and deployment flexibility across public cloud, private cloud, on-premise, and hybrid.
- Accountability: One SLA across telephony and AI, named incident ownership, and uptime history.
Most published voice streaming comparisons rank vendors on model quality and pricing. The nine items above decide whether the rollout works, and they rarely appear in a feature table.
Stress-Test Scenarios to Run Before You Sign: A Practical Pilot Script
Run these against every shortlisted provider, with the same script and the same recordings, so results are genuinely comparable.
- The bad-line test: Place calls from a moving vehicle on mobile data in a tier-2 city, on speakerphone, with background noise. Score intelligibility, barge-in accuracy, and dropped audio.
- The code-switch test: Have testers switch languages mid-sentence and mid-call, including numerals and product names. Score entity accuracy, not just word error rate.
- The interruption test: Interrupt the agent five times in one call, including during a long disclosure. Measure how quickly it stops and whether it resumes at the right point.
- The integration test: Trigger a live CRM lookup and a payment-link generation inside a turn. Time the silence the customer hears.
- The handoff test: Escalate to a live human agent and record what the agent sees at connection, then ask the tester whether they had to repeat anything.
- The burst test: Fire three times your expected peak concurrency for thirty minutes and log latency, failures, and audio quality throughout.
- The audit test: Pick one call from the burst test and request the full compliance package. Time how long it takes to arrive.
Scenario seven is the one buyers skip most often and regret most reliably.
How Exotel Designs Around These Failure Modes on One Telecom-Grade Stack
Exotel’s answer to these failure modes is architectural rather than feature-by-feature. AI voice agents, the cloud contact center, and the telecom network run on a single stack, with presence across 11 telecom circles in India and licensed local number infrastructure in the UAE. That removes the upstream-carrier dependency behind both connect-rate and finger-pointing problems.
The AgentStream voice streaming infrastructure handles the realtime path: noise-resilient speech recognition, barge-in and interruption handling, low-latency media streaming designed for sub-300 ms voice latency, and intent and sentiment signals available mid-call. GenAI voice agents on top of it work in English, Hindi, Hinglish, Arabic and other regional languages, with CRM and payment gateway integration so actions happen during the conversation rather than after it.
Handoff runs through the same platform the human agents already use. A persistent conversational context layer carries customer profile and conversation history across bots, agents, and channels, so a transfer arrives with the transcript and extracted details in place. One supervisor can monitor several AI conversations at once and step in with full context, and each intervention feeds back into the AI in a continuous improvement loop. Exotel reports up to 75% containment on routine queries and up to 40% agent productivity gains from AI Assist in customer deployments.
On the compliance side, conversation quality analysis scores 100% of interactions for script adherence, sentiment, and disclosure language instead of a manual sample. Audit-ready recording, consent capture, encryption, and role-based access sit under ISO/IEC 27001 and PCI DSS certification, with controls mapped to the RBI Fair Practices Code, OJK, and BSP expectations for regulated outbound. The platform carries 25B+ interactions a year at 99.99% uptime for 7,000+ enterprise customers across 60+ countries, including banks, lenders, marketplaces, and mobility platforms where a dropped call has a direct revenue cost.
No single capability here is unique. What matters is that when telephony, AI, and the contact center share one architecture, the seams where real-time voice deployments usually break become internal engineering problems instead of vendor boundaries.
FAQs
Hold vendors to end-to-end conversational latency measured on a live phone call in your primary market, not to model inference time. Roughly 500 to 800 ms from end of customer speech to start of agent speech feels natural to most callers, and anything past about 1.2 seconds reads as a broken connection. Exotel designs its streaming layer for sub-300 ms voice latency on the media path, which leaves headroom for model inference and any API call inside the turn.
Only if your AI vendor does not own the telecom layer, which is the common arrangement with bot-only providers sitting on third-party telephony. Splitting them adds a second contract, a second SLA, and an ambiguous boundary during voice quality incidents. Running AI and telephony on one stack removes that seam and gives you a single owner for root-cause analysis when call quality degrades.
Supply your own call recordings, ideally 200 to 500 real conversations covering your customer mix, and have every shortlisted vendor transcribe the same set. Score entity-level accuracy on the things that drive actions, such as amounts, dates, account identifiers, and product names, rather than overall word error rate. Include code-switched audio deliberately, because Hinglish and dialect Arabic are where monolingual benchmarks stop predicting real performance.
Consent capture recorded as a first-class event, audit-ready call recording with defined retention, encryption with role-based access, and automated script-adherence scoring across all conversations rather than a manual sample. Calling-window and retry controls matter equally, since pacing logic is where automation most easily drifts out of regulatory alignment. No platform can guarantee compliance on your behalf, so evaluate how quickly it can produce evidence when a regulator or internal audit asks.
Timelines depend far more on integration and compliance sign-off than on building the conversation flow. A no-code bot builder with pre-built CRM and helpdesk connectors can compress the build phase to weeks, while number provisioning, security review, and regulatory approval typically set the real critical path. Budget explicitly for load testing and an audit dry run before go-live, because those two steps catch the failures that otherwise surface after launch.
