A voice agent demo runs on a laptop, over a clean connection, with a colleague reading a script. Production is a 3 a.m. call from a customer on speakerphone in a moving truck. Nothing about the model changed between those moments — what changed is the phone line underneath it.
That gap is now a product category. Ecosmob Technologies launched a Voice AI Readiness Assessment this week that audits SIP and WebRTC topology, benchmarks speech-to-text accuracy against real call audio, and tests latency, jitter, and CRM responsiveness before a project commits budget.
For agencies selling voice work, that is the useful part. The failure data has been consistent all year: practitioners auditing enterprise deployments put roughly 88% of AI agents failing to reach scaled production, with voice failing faster than text because every defect happens out loud. Only a third of the voice pilots that team audited hit their committed cost-per-resolved-interaction target in 90 days.
Why voice pilots die differently
Three structural facts make voice harder than any chat deployment. First, you are stitching at least four real-time systems together — voice activity detection, speech-to-text, the LLM, and text-to-speech — and any one of them missing its latency budget kills the conversation. Second, output is immediately consequential: a hallucinated refund policy spoken over the phone is treated as a contract, not a draft. Third, voice traffic is bursty: Monday 9 a.m. concurrency can run eight times your Sunday average.
None of those are model problems — they are operations problems wearing a model’s clothes.
Test the plumbing before the prompt
Most readiness work is telephony: audio quality, jitter, codec consistency, API responsiveness, and the hand-off points. Carrier choice is measurable — an independent June 2026 test of 120 outbound calls per provider found a 118 ms p95 carrier round-trip on one stack versus 161 ms on another, at roughly half the carrier cost.
The catch is that vendor latency numbers are not comparable. As one platform comparison notes, one vendor’s headline “sub-200 ms” is audio-layer round-trip, not end-to-end; another’s frequently quoted 600 ms is a full turn measured across production calls. Ask what is measured, from which event to which event, at what percentile — then read the carrier benchmarks against the platform claims.
The cheapest structural win is co-location: pin STT, LLM, and TTS to the same cloud region where the SIP gateway terminates. Teams that have done it measured 90 to 140 ms of pure network savings — a meaningful slice of a sub-second budget.
Score the tail, not the average
The commonly cited budget for a natural-feeling agent is roughly 800 ms end to end: about 50 ms of capture and VAD, 150 ms of STT, 400 ms to first LLM token, 150 ms to the first TTS chunk, and 50 ms of network egress. Conversations feel broken past about 1.5 seconds, and call-completion rates fall materially once the p95 crosses two seconds.
Real pipelines run closer to 750-1,400 ms, because the budget has roughly twelve components and streaming only touches three of them. Endpointing is the largest controllable one: fixed-silence detection waits around 500 ms before deciding the caller has stopped, which is half a second of dead air on every turn before the agent has done any work. Semantic endpointing cuts that to roughly 150-200 ms.
Measure per-turn p95 and dead-air events — pauses longer than two seconds — rather than averages. An average word error rate of 9% can hide 22% on calls with background noise or non-native speech, which is exactly the segment that escalates and damages the client relationship.
The gate: what to require before the number goes live
Before a client’s line is cut over, seven things should have evidence behind them:
- Real-audio accuracy. Test STT against recordings from the client’s actual callers — accents, noise, speakerphone — not synthesised samples. Treat word error rate above roughly 15% as a failure state, not a tuning note.
- Barge-in. Interruption handling tested deliberately: the agent stops, registers the interruption within a frame or two, and flushes queued audio instead of talking over the caller.
- Escalation and hand-off. Transfers tested end to end, with a structured payload arriving for the human agent: identity, recent intents, what the agent said, what is unresolved. A caller should never repeat an account number they just gave the bot.
- Failure injection. Kill each dependency on purpose — slow LLM, STT timeout, CRM 500, TTS outage — and confirm the call degrades to a fallback or a human without dropping.
- Load and soak. Test at three times expected peak concurrency across the telephony layer, the media path, and the APIs, not just the AI vendor’s endpoint.
- Cost at the peak. Recompute per-minute cost at peak concurrency with retrieval, fallback calls, and recording storage included. Pilot pricing rarely survives the wiring-up.
- Adversarial calls. Attempt prompt injection by voice, a cloned-voice impersonation, and a call replay.
A 12-point pre-live checklist covers similar ground, and readiness frameworks package the same tests as deliverables: a scorecard, a gap matrix, a remediation roadmap. That structure turns “the bot sounds weird sometimes” into a prioritized engineering list.
Shadow mode is the cheapest insurance
Do not flip a client’s phone line from demo to autonomous in one step. The rollout that holds up is shadow first: the agent listens to live calls without speaking, generates what it would have said, and that output is scored against what the human actually did. Zero caller risk, real data, and an evaluation harness validated before it is load-bearing. Then narrow-intent assist, then autonomous with a live SLO dashboard.
Advance on numbers, not the calendar. A reasonable gate for a tier-one inbound agent: containment in the 60-75% range, median latency under 600 ms, p95 under 1.5 seconds, escalation accuracy high enough that false transfers are rare. Watch containment and correctness together — chasing deflection is how agents end up guessing on calls they should have handed off.
What this means if you sell the work
Package readiness as a paid fixed-scope phase: one to two weeks, deliverables in units, priced well under the deployment. It is the cheapest line item in the project and the one that protects the rest — unverified pilots do not fail quietly, they become unbilled support work. The market is moving toward agents that execute actions rather than answer questions: platforms now market voice agents that reset passwords, grant access, and provision devices. Every added write action is another integration to test, another idempotency case, another failure path.
The judgment call is whether the work should exist. Voice AI pays off on repeatable volume — the same twenty intents, thousands of times. If a client’s calls are genuinely novel and high-touch, readiness testing cannot rescue the economics, and saying so is cheaper than a pilot that dies in month two.
Start small this week: pull the last fifty calls, listen to ten in full, and count the moments the agent paused too long, repeated a question, or stated something ungrounded. Two or more in ten is an engineering problem, not a vendor problem — and far easier to write into a statement of work than to discover on a client’s busiest Monday.