When the pilot works, but the scaled deployment fails
A few months ago, you declared victory: your voice AI pilot was successful. It took time to roll it out, dial in how the AI agent responds, and ensure smooth transitions between your AI and human agents.
Fast forward, you’ve gone from two test markets to rolling it out across 25 countries. In most markets, performance holds. In one or two, it does not. Containment drops below target, transcripts show misheard account numbers, and callers start talking over the bot or repeating themselves after a delay. The first questions usually focus on the model, the context, or the vendor.
What often gets overlooked is whether the audio reaching the model in that specific market is good enough for the AI to perform consistently.
81% of organizations have already AI-enabled their voice environments, and the single most-cited concern, named by 45%, is the quality of AI output. That concern is real. But "quality of AI output" is often read as a model problem, and it isn't always the case. A capable model receiving degraded audio across a fragmented international voice path can still produce poor output. In those cases, the issue may be the input rather than the intelligence layer.
If you own the infrastructure supporting a voice AI program, this is one part of performance you can control and measure.
AI is unforgiving of bad voice infrastructure
People are exceptional error-correction machines. If a consonant clips, jitter interrupts a word, or an unfamiliar accent changes the sound, listeners often use context to fill the gap without thinking about it.
Speech recognition engines have less room to compensate. If the audio arrives degraded, transcription quality can suffer, and those errors can carry into downstream stages.
Three factors account for most of it.
1. Latency breaks turn-taking
Human conversation moves quickly. Research on conversational turn-taking across languages indicates that the typical gap between speakers is roughly 200 to 300 milliseconds, about the length of a blink. People often begin planning a response before the other person has finished, so delays become noticeable quickly.
That leaves a tight response-time budget for voice AI:
| Under 300 ms | Excellent | Feels human. Matches the natural gap in face-to-face conversation. |
| 300 – 500 ms | Good | Fluid and responsive. The standard production goal. |
| 500 – 800 ms | Acceptable | Passable, but callers notice and start slowing down their own speech. |
| 800 – 1,200 ms | Poor | Awkward. Callers begin talking over the agent. |
| Above 1,500 ms | Broken | Feels like a walkie-talkie. Abandonment climbs. |
Source: Ashutosh Singh, AVOXI Analysis
Two things about that table matter if you own infrastructure rather than models.
First, the budget is end-to-end. It covers everything between the moment your customer stops speaking and the moment they hear a reply: endpointing, speech recognition, model inference, speech synthesis, and network transit at both ends. Most voice AI teams optimize the middle of that chain because that's where a vendor dashboard actually shows them.
Second, network transport can consume a meaningful share of that budget. On a domestic call over a well-provisioned route, network transit accounts for a small share. On an international call crossing multiple carrier hops with transcoding at each boundary, it can consume a budget whose entire "good" range is 500 milliseconds.
A team can spend a quarter tuning prompts and reducing inference time without recovering the milliseconds lost in the route. In that case, the model is not the slow part.
2. Audio clarity sets a higher bar than you think
Mean Opinion Score is the standard measure of perceived call quality. AVOXI treats 4.0 and above as good — a call a person will describe as clear and easy.
That threshold was set for human listeners. For AI-mediated conversation, the bar moves. AVOXI's engineering guidance is that voice bots need a 4.3 or higher score to perform reliably.
The gap between those two numbers matters because humans can compensate for small audio imperfections that speech models may not handle as well. A 4.0 call your agents would never complain about can still be the reason your intent recognition is underperforming.
| Human conversation | 4.0 and above | AVOXI's "Good" threshold. Humans reconstruct meaning through clipping, jitter, and accent variation. |
| Voice bot / ASR pipeline | 4.3 and above | Speech models depend on the quality of the audio they receive. Higher audio quality gives the model cleaner input to interpret. |
3. Jitter and packet loss compound downstream
Packet loss can remove enough audio to change a word or phrase. In an AI pipeline, that can affect the transcript, which can then affect intent classification, routing, and the customer’s next interaction.
Errors can compound across stages. Each stage passes its output to the next, so a problem that begins with degraded audio can eventually surface as lower containment, longer handle time, or softer CSAT. By then, the original network issue may be several layers removed from the business metric, which is why these programs can be diagnosed as model problems.
AI is driving more traffic to voice, not less
Some infrastructure plans still treat voice as a channel in decline, with investment focused on keeping it running rather than preparing it for growth.
The data says the opposite. 82% of companies are seeing or expect AI to drive more traffic to their voice channel. As AI reduces the cost of handling a call, organizations stop rationing access to the channel that customers preferred all along.
The voice layer may need to carry more traffic for AI systems that are less tolerant of degraded audio. That makes the underinvestment more consequential as volume grows.
Why your current setup probably isn't ready
96% of enterprises use more than one voice provider,usually as the result of regional purchasing, expansion, and acquisition. Every provider brings its own codecs, transcoding behavior, routing paths, and hop counts, and the differences vary by market.
As a result, AI performance can vary by geography.
Here is what that looks like in practice. The bot performs well on US calls. The same model and prompt underperforms in APAC, and the team spends a quarter tuning the model. If the model and prompt are unchanged, the route may be the variable. Standard CCaaS reporting often does not surface that difference because it reports on the call, not on the route it took.
77% of enterprises already report difficulty achieving consistently high voice quality across markets. That variance was more manageable when human listeners could compensate for audio issues, but it becomes more consequential when a model cannot.
The point is to understand what your voice layer is delivering, market by market. For an AI workload the criteria that matter are the low latency, high audio clarity, coverage in the markets you operate in, integration paths that don't add hops, and routing you can change without a carrier ticket.
An IT Director's checklist for an AI-ready voice layer
Five questions worth answering before the next pilot, not after it.
- Can you measure call quality per route, per market, and per call, or only in aggregate? Averages hide exactly the regional degradation that breaks AI.
- Do you have continuous monitoring and testing in place, or do you learn about degradation from customers and agents? In the first annual State of International Voice study, 85% of enterprises reported periodic contact center downtime, and roughly 60% experienced outages more than once a quarter. If that's being discovered reactively, your AI program is inheriting an unknown baseline.
- How many carrier hops sit between your customer and your AI? Every hop is transcoding, adding latency, and another place where quality can be lost without anyone owning it.
- Can your provider expose the telemetry your AI team needs — MOS, jitter, packet loss, post-dial delay — at a granularity that maps to individual conversations?
- When quality does degrade, who diagnoses it? A provider that hands you a dashboard and a ticket queue has moved the work to your team, not removed it.
92% of organizations are already measuring the outcomes of their AI-enabled voice systems. That measurement discipline is the right instinct, but it often stops at the application layer. If you're tracking containment, sentiment, and resolution rates but can't see MOS by route, you're measuring the symptom and not the cause.
Sequence matters
The instinct with voice AI is to buy the intelligence first and address the voice layer only when something breaks. That order can waste a pilot and a quarter of work while leading teams to blame the technology when the transport is the actual issue.
Evaluate and improve the voice layer before or alongside the AI investment.The model’s performance depends on the quality of what reaches it.
Is your voice infrastructure ready for AI?
Learn how to evaluate across network architecture, call quality, routing, integrations and visibility to build a voice foundation that can support AI at scale.
