Telephony infrastructure is easy to underestimate. A call connects, the conversation happens, the call ends — and the quality of what happened in between rarely shows up in a dashboard.

The real cost lives in the gaps — in that extra 200 milliseconds of delay that makes a customer repeat themselves, in the choppy audio that breaks a critical support call, in the dropped call that ends a conversation before it could close.

Latency, drop rates, and audio failures are not edge cases. They are the everyday reality of poorly built telephony infrastructure, and they have a measurable dollar value.

Businesses globally risk $3.8 trillion in sales due to bad customer experiences. Service delivery and communication problems were the leading causes of bad experiences across all industries.

This blog breaks down exactly what each problem is, where the thresholds are, and what they cost — so you can stop treating call quality as a technical concern and start treating it as a revenue concern.

VoIP latency hurts in different places depending on what type of delay you are measuring — and the threshold for "acceptable" changes significantly depending on whether a human or an AI is on the other end of the call.

One-way VoIP delay, round-trip delay, and AI processing latency each have different thresholds — and most discussions lump them together. Here is how to tell them apart.

The Definition

VoIP latency (also called one-way delay) is the time it takes for a voice data packet to travel from one endpoint to another across an IP network. It is measured in milliseconds (ms). This is distinct from:

• End-to-end (round-trip) latency: The total time for a packet to travel to the destination and back. This is roughly 2× one-way delay and is what you experience as the "echo" effect on bad calls.
• Network latency: The delay introduced specifically by the network path — routers, switches, hops across geographies.
• Processing latency: In AI voice systems, delay is added by transcription (STT), intent analysis (NLU), response generation (LLM), and text-to-speech (TTS) before the voice packet is even transmitted.

When people refer to "acceptable call latency," they are almost always talking about one-way VoIP latency. The ITU-T G.114 standard — the internationally recognized benchmark for voice communications — sets the threshold at 150ms one-way.

The Latency Thresholds (One-Way VoIP Delay)

Source: ITU-T G.114 standard, via Vida AI

The ITU-T G.114 standard recommends a maximum one-way delay of 150ms for acceptable voice quality. Beyond this threshold, users begin experiencing the "talk-over" effect and conversations become noticeably disrupted.

Why AI Voice Systems Demand Even Lower Latency

For human-to-human calls, a person can instinctively adjust to minor delays. For AI voice systems — automated IVRs, AI sales agents, AI support — the tolerance is far tighter because delay compounds across multiple processing steps.

Latency compounds through sequential processing stages in AI call systems: a customer's voice travels across the network (~120ms), is transcribed by speech recognition (~100ms), analyzed for intent (~200ms), generates a response (~150ms), converted to audio (~180ms), and travels back to the customer (~120ms). That stacks to over 870ms before a response is heard — far beyond the ITU threshold — which is why the network transport layer must contribute as little latency as possible. Infrastructure that holds ~80ms at the transport layer gives AI processing stages room to breathe.

What Causes VoIP Latency

• Geographic distance: More hops across regions = higher delay. A call from Mumbai to a US server routed through three intermediary carriers adds latency at every hop.
• Network congestion: Overloaded routes slow packet delivery.
• Codec processing: Compressing and decompressing audio takes time. G.711 is faster but bandwidth-heavy; G.729 saves bandwidth but adds processing delay.
• Jitter buffer: Networks introduce a buffer to smooth out inconsistent packet arrival — this adds a small, consistent delay to prevent choppy audio.

The Definition

Voice data travels across IP networks as packets — small chunks of audio data sent continuously in sequence. Packet loss (also called drop rate) is what happens when one or more of those packets fail to reach their destination. Unlike a file download where a dropped packet can be resent, voice calls are real-time: with voice, packet retransmission is not possible over live calls. The missing audio is simply gone.

The result: robotic voice, clipped words, syllables that disappear mid-sentence.

In most data transmission, lost packets are invisible — the network requests a resend and the user never notices. Voice is different. Every packet represents a slice of real-time audio. When it is lost, that slice of the conversation disappears with it. There is no mechanism to recover it after the fact. This means packet loss in voice is immediately perceptible to the caller — as choppy audio, clipped words, or one-sided silence — at levels that would be completely undetectable in any other type of data transfer.

What Causes Packet Loss

• Network congestion: Too many devices or calls sharing limited bandwidth. Insufficient bandwidth leads directly to latency, jitter, and packet loss.
• Hardware failures: Damaged cables, malfunctioning routers, outdated network switches.
• Software misconfiguration: Incorrect QoS (Quality of Service) settings that fail to prioritize voice traffic.
• Carrier routing issues: When calls transit through multiple carriers with weak interconnects, packets can be dropped at handoff points.

Jitter: The Cousin of Packet Loss

Jitter is variation in packet arrival timing. Even if all packets arrive, jitter causes them to arrive out of sequence or with irregular gaps — producing the robotic, fragmented audio that is often misidentified as "bad signal." High jitter results in choppy voice or temporary glitches; VoIP devices implement jitter buffering algorithms to compensate, but those algorithms also drop packets that arrive with excessive delay.

What Is a MOS Score?

Mean Opinion Score (MOS) is the industry-standard metric for measuring perceived voice call quality. The scale runs from 1.0 (bad) to 5.0 (excellent), with ratings assigned based on how a human listener perceives the call.

Sources: Balto AI, Scorebuddy

Most VoIP calls fall between a MOS score range of 2.5 to 4.5. A drop from 4.2 to 3.6 may seem small on paper, but it can mean the difference between a smooth support call and a frustrating one.

How Latency and Drop Rates Destroy MOS

For each 100ms latency and 200ms jitter increase, the MOS score drops by one point. Sustained packet loss compounds this further, pushing scores below the threshold where calls can be tolerated.

This is the chain of failure:

• High latency → awkward pauses → talk-overs → lower MOS
• Packet loss → missing audio → choppy voice → lower MOS
• High jitter → fragmented audio → robotic sound → lower MOS
• Low MOS → frustrated customer → churn

A poor MOS value directly impacts business communications, customer experience, and end-user satisfaction in contact center and telephony environments.

Audio Failures in Practice

Audio failures are what happens when the above go uncorrected: one-sided silence, echo that makes the caller sound like they are in a cave, clipped words where every third syllable disappears, calls that drop entirely mid-conversation. In a contact center or AI voice deployment, these are not just annoying — they represent interactions that have failed before any business outcome could be reached.

The Direct Cost: Missed and Dropped Calls

More than half (53%) of consumers say they will cut spending after a bad customer experience, according to Qualtrics XM Institute's 2025 Consumer Trends Report. For businesses running voice-first customer interactions, a degraded call is that bad experience. A call that connects but sounds terrible is, in practice, a missed call.

The customer heard you, but they did not trust you.

The Churn Cost: Silent Switching

Bad call quality does not produce support tickets. It produces quiet exits. 63% of consumers are willing to switch to a competitor after just one bad experience — a figure that has grown 9% year on year. 70% of consumers will abandon a brand after just two negative experiences.

In telephony-dependent industries — financial services, logistics, healthcare, SaaS support — a single call where the audio breaks up at a critical moment can end a customer relationship that took months to build.

The Scale: What Service Failure Costs Globally

According to Qualtrics XM Institute's 2025 Consumer Trends Report, $3.8 trillion in global sales are at risk from bad customer experiences. The research, spanning 24,000 consumers across 23 countries, identified service delivery and communication breakdowns as the leading causes.

When a customer is already experiencing an issue, a choppy, lagging, or dropped call doesn't just fail to solve the problem, it compounds it. The pain shows up as silent churn, lost referrals, and permanently lower lifetime value.

The Compounding Cost of Silent Churn

Acquiring new customers in telecom costs significantly more than retaining existing ones. When call quality degrades the experience silently — without a complaint, without a ticket, without a visible signal — every churned customer represents both lost revenue and the compounded cost of replacement.

Latency, packet loss, and audio quality are not random — they are determined by the quality of the telephony infrastructure underneath the call. Every call you make or receive runs through a stack of components:

• SIP trunking: The protocol layer that carries voice signals over IP networks. Poor SIP configuration or low-quality SIP providers introduce jitter and packet loss at this stage.
• Number provisioning: The quality of DIDs and the carrier interconnects attached to them affects routing quality and drop rates.
• Carrier routing: How a call gets from Point A to Point B across the PSTN and IP networks. More intermediary hops = higher latency and more points of failure.
• Network interconnects: The quality of peering between carriers. Weak interconnects at handoff points are where packets go missing.

Without infrastructure-level observability, the call drop is visible but the cause is not — which makes it nearly impossible to fix.

Setting benchmarks against the ITU-T G.114 standard and industry best practices, here is what telephony infrastructure needs to deliver:

Sources: ITU-T G.114 | NetBeez

Infrastructure built for AI-native telephony — where call latency at the transport layer must stay low enough for AI processing to happen within human conversation tolerances — requires direct carrier-grade connections, intelligent routing across 190+ outbound countries, and real-time quality monitoring that catches degradation before it reaches the caller.

At Vobiz, calls run at ~80ms latency at the transport layer — well within the 150ms ITU-T threshold, and with enough headroom that AI agents built on top can respond in a way that feels natural, not robotic. With DID coverage across 130+ countries and PSTN-grade infrastructure underneath every call, the quality problems described in this post are infrastructure choices, not inevitabilities.