How does call quality differ between SIP-based and PSTN-based voice AI deployments? The answer determines whether your callers experience natural, responsive conversations or frustrating delays and dropouts. SIP and PSTN operate on fundamentally different network architectures, each with distinct performance characteristics that directly impact how an AI agent sounds and responds during a live call.
Choosing the wrong infrastructure means investing in AI voice technology that alienates the people it was meant to serve. This article explains exactly what happens in each deployment type, where problems emerge, and which architecture fits your specific operational needs.
Understanding The Network Difference Between SIP And PSTN
PSTN, the Public Switched Telephone Network, is the traditional circuit-switched system that has carried voice calls for decades. It creates a dedicated connection between two points for the duration of the call, using fixed bandwidth and routing. This means your call path is reserved the moment you dial, regardless of network congestion. PSTN still handles roughly 60% of business voice traffic globally, though that proportion shrinks annually as businesses migrate to IP-based systems.
SIP, Session Initiation Protocol, is a packet-switched alternative that treats voice calls like data packets traveling across the internet. Instead of a dedicated circuit, SIP breaks your conversation into small digital chunks, sends them independently across whatever route is most efficient, and reassembles them at the destination. This approach is dramatically more cost-effective, which is why cloud PBX providers, hosted phone systems, and most modern voice AI platforms use SIP as their foundation. Your AI agent typically runs on SIP infrastructure because carriers charge less to provision and maintain it.
The practical implication: PSTN guarantees your call path but costs more and doesn't scale easily for high-volume operations. SIP scales cheaply but depends entirely on your internet quality and the quality of the network between you and your provider. When a dental practice using SIP-based AI voice agents experiences a congested Wi-Fi network at 9 AM, callers hear stuttering and delays. The same practice on PSTN would hear nothing wrong, because the call doesn't compete with email downloads or video streaming.
How Latency Shapes Caller Perception And AI Response Time
Latency, the time between when you speak and when the other party hears it, is where SIP and PSTN diverge most obviously. PSTN typically delivers latency of 50 to 150 milliseconds one-way, because the call follows a fixed path optimized by the telephone company. End-to-end latency below 150 milliseconds feels natural to human conversation. Above 300 milliseconds, people start speaking over each other and conversations feel awkward.
SIP latency depends on your internet connection, your VoIP provider's infrastructure, and the physical distance between the AI agent's server and your location. A local SIP provider with good infrastructure might deliver 80 to 120 milliseconds latency. The same call routed through a global provider or across poor internet connection can hit 400 to 600 milliseconds. When latency exceeds 250 milliseconds, the AI agent and caller begin to interrupt each other. The caller asks a question, the AI starts answering before the caller has finished speaking, creating the sensation that the agent isn't really listening.
Real-world scenario: A healthcare scheduling practice using a cloud AI agent experienced caller frustration because their provider routed calls through data centers 2,000 miles away. Moving to a regional SIP provider dropped latency from 380 milliseconds to 120 milliseconds, and caller feedback on agent responsiveness improved dramatically. Switching to PSTN would have cost three times more per minute but would have guaranteed sub-150-millisecond latency regardless of network load. The practice chose the regional SIP option because cost mattered more than guaranteed performance.
Codec Selection And How It Affects Audio Clarity On Each Network
A codec is the algorithm that compresses voice into data for transmission, then decompresses it at the receiving end. PSTN traditionally uses the G.711 codec, which captures the full range of human speech at high fidelity. It's bandwidth-heavy but sounds pristine, which is why PSTN calls often sound clearer than cellular. SIP providers can use any codec, and most choose between G.711 (high quality, higher bandwidth), G.729 (compressed, lower bandwidth), or modern alternatives like Opus (adaptive quality depending on network conditions).
When your AI agent uses SIP with a low-bandwidth codec like G.729, callers might hear robotic or hollow qualities in the agent's voice. This is particularly noticeable for AI-generated speech, where codec artifacts combine with synthesis artifacts to produce an obviously artificial sound. Operators typically report that G.711 over SIP produces acceptable quality for AI agents, while G.729 produces obviously cheap-sounding calls that undermine trust. PSTN's G.711 codec, delivered over a guaranteed circuit, sounds consistently natural.
The trade-off is stark: G.711 consumes 64 kilobits per second per call. A small business running 20 concurrent AI calls on SIP using G.711 needs a 1.3 megabit per second internet connection just for voice, plus headroom for other traffic. PSTN delivers the same quality with no bandwidth concern for the business side, though the carrier charges per-minute rates that scale quickly. A contact center handling 100 concurrent calls would find SIP vastly more economical, accepting slightly lower audio fidelity in exchange for manageable costs.
How Does Call Quality Differ Between SIP-Based And PSTN-Based Voice AI Deployments In Practical Operation
The difference becomes tangible when you map real call scenarios. A restaurant using AI to handle reservation calls during dinner service needs reliable performance during traffic spikes. If the restaurant uses SIP and their internet connection hits capacity at 6 PM, calls degrade immediately. Callers hear audio breakup, requests get missed, the AI agent sounds delayed and confused. The restaurant burns reputation because callers experience poor service from a machine. PSTN would deliver flawless call quality regardless of internet load, because voice traffic never touches the business internet.
Alternatively, a growing B2B appointment-setting company using AI outbound calling needs to run 50 concurrent calls economically. PSTN would cost 50 to 80 pounds per day in per-minute charges, depending on call length. SIP costs 5 to 15 pounds per day for the same volume, because you're not paying per-minute. The SIP calls sound good enough for business conversations, though not quite as pristine as PSTN. The company accepts this trade because cost efficiency scales their operation.
Reliability differences matter less than operators assume. Modern SIP providers maintain 99.5% to 99.9% uptime. PSTN maintains similar uptime. The real difference is failure mode. When SIP fails, all your calls fail simultaneously because they depend on one internet connection. When PSTN fails, your calls continue because the carrier maintains redundant switching infrastructure. Businesses in industries where a single hour of downtime costs thousands typically prioritize PSTN for this reason.
Memory And Context Across Different Network Architectures
Call quality isn't just audio quality. It includes whether the AI agent can access caller history and context. A persistent voice AI memory system allows an agent to greet a repeat caller by name and remember their previous conversation, creating the impression of a competent human agent. This matters equally on SIP or PSTN, but the implementation differs. SIP-based AI agents typically integrate with a cloud database where caller memory is stored. PSTN-based deployments require separate integration to a CRM or memory system.
When implemented well, caller memory technology transforms perceived call quality by eliminating repetition. A service business receives a follow-up call from a customer, and the AI agent immediately says, "Hi Sarah, calling about the invoice we discussed on Tuesday." This takes five seconds instead of two minutes, and the caller perceives the service as dramatically more professional. The network type, SIP or PSTN, doesn't affect this aspect of quality. What matters is whether the AI platform supports persistent voice AI memory and whether your memory system integrates properly with the deployment.
If you choose SIP, verify that your provider's caller memory system has sub-second lookup speed. If latency between the AI agent and the memory database exceeds one second, the agent will sound hesitant or confused as it waits for caller history to load. Reputable SIP providers cache frequently-called numbers locally to avoid this delay. PSTN deployments face the same requirement but rarely suffer from it because call volume is typically lower, meaning memory lookups are less congested.
When Network Architecture And AI Platform Integration Create Quality Problems
Most call quality problems don't stem from SIP versus PSTN. They stem from mismatched expectations or poor implementation. A business deploys a SIP-based AI agent on unreliable 10 megabit broadband and blames the technology when calls drop. The technology works fine. The network doesn't. Similarly, a business chooses PSTN for quality, connects the AI agent through a third-rate provider, and experiences frequent call drops. Here again, the failure is provider selection, not the network protocol.
AI agents running on SIP struggle most in two scenarios. First, when business bandwidth is constrained. If your office uses 15 megabits per second for 50 employees, adding 10 concurrent AI voice calls will degrade internet performance for everyone. Second, when your SIP provider hosts servers far from your location. A call center in Manchester using a provider whose nearest data center is in California will experience 40 to 60 milliseconds of extra latency simply from the geographic distance. These are solvable problems, but they require foresight.
PSTN-based AI agents fail most often when integration is poor. If the AI platform doesn't integrate smoothly with your telephony setup, callers experience unexpected disconnects or transfers. PSTN doesn't solve integration problems, it just removes network latency as a variable. Some AI vendors position PSTN as premium option for this reason, not because of network quality but because they're solving for operational reliability. If a vendor offers PSTN at four times the cost of SIP, ask what problem PSTN solves that SIP doesn't. Honest answers are rare.
Choosing The Right Architecture For Your Operational Needs
Start by mapping your actual constraints. If you're running fewer than 20 concurrent AI calls, cost becomes secondary. Reliability and audio quality matter more, which favors PSTN. If you're running 50+ concurrent calls, SIP becomes economically essential. If your business cannot tolerate any downtime, PSTN's carrier-grade infrastructure is worth the cost. If downtime is annoying but survivable, SIP with a good provider works fine.
Network quality is the deciding factor for most businesses. If your internet connection is business-grade fiber with 10 milliseconds latency to your ISP and 99.9% uptime, SIP will deliver call quality equivalent to PSTN at a fraction of the cost. If your internet connection is cable or DSL shared with 30 other households, SIP will disappoint you. Run a speed test and a latency test from your office. If latency to major internet backbones exceeds 30 milliseconds or your available bandwidth is marginal, PSTN is the safer choice.
Consider a hybrid approach if you can. Many businesses run outbound AI campaigns on SIP, where cost matters and one-way latency tolerance is higher, and inbound customer calls on PSTN, where quality and first impression matter. This requires two separate phone systems, so it's typically viable only for larger operations. For smaller teams, choose SIP or PSTN, optimize the infrastructure around your choice, and implement features like built-in CRM integration to maximize perceived quality through context and caller memory.
Frequently Asked Questions
Does SIP always have worse call quality than PSTN?
No. SIP with good internet, a regional provider, and G.711 codec produces call quality indistinguishable from PSTN. The difference emerges only when infrastructure fails: congested internet, distant data centers, or low-bandwidth codecs. PSTN's advantage is consistency, not absolute quality.
Can I switch from SIP to PSTN without changing my AI voice platform?
It depends on the platform. Some AI voice vendors support both SIP and PSTN backends. Others specialize in one. Ask directly whether your vendor's infrastructure can be provisioned on PSTN lines. Many cannot without rebuilding core components, meaning switching often requires platform migration.
What latency should I expect from SIP in my location?
Latency depends on your provider, your internet quality, and geography. UK-based businesses using a UK-based SIP provider typically see 80 to 150 milliseconds latency. Latency above 250 milliseconds becomes noticeable in AI conversations. Request a trial from your provider and measure real latency before committing.
Does caller memory technology work differently on SIP versus PSTN?
No. Persistent voice AI memory and caller memory technology function identically regardless of network protocol. What matters is whether your AI platform supports memory features and whether your database integrates with acceptable latency. Choose your platform first, then optimize network architecture around it.
Why would I ever choose PSTN if SIP is cheaper?
Cost isn't your only constraint. PSTN guarantees call quality during network congestion, eliminates bandwidth concerns, and provides carrier-grade reliability when downtime is expensive. For customer-facing inbound calls where first impression matters, PSTN's consistency justifies the cost. For outbound campaigns or internal communications, SIP usually makes financial sense.
Can I test call quality before deploying an AI voice system?
Yes. Request a pilot from your prospective AI platform. Ask them to provision calls on your preferred network type (SIP or PSTN). Run 20 to 50 test calls, measure latency, assess audio quality, and evaluate how the agent handles your actual use cases. Most reputable vendors offer this. Any vendor unwilling to pilot is worth investigating further.
Evaluating Your Network Before Committing to an Architecture
Before a business can compare SIP and PSTN meaningfully, it needs a baseline picture of its own network environment. The protocol choice matters far less than whether the supporting infrastructure can carry it reliably. A practical starting point is a structured network audit conducted during peak operating hours, not at quiet periods when results flatter. The audit should capture round-trip latency from your premises to several major internet exchange points, jitter readings across at least a thirty-minute window, and any packet loss that appears under normal load. Jitter, the variation in delay between successive packets, is worth examining separately from average latency because AI voice synthesis is particularly sensitive to irregular delivery even when average latency looks acceptable. A call with steady latency sounds natural; a call with erratic delivery sounds broken even if the average figure appears fine.
When reviewing audit results, treat any packet loss above a small threshold as a red flag for SIP deployment, not merely a note. Packet loss forces retransmission or concealment algorithms that introduce audible artefacts, and AI-synthesised speech tends to expose these artefacts more noticeably than a human voice would. If your audit reveals packet loss, investigate the cause before choosing your network architecture. A simple router replacement or VLAN separation between voice and data traffic may resolve the issue without requiring a shift to PSTN.
A Worked Scenario: Tracing a Single Inbound Enquiry Through Both Architectures
Consider a fictional specialist equipment hire company, Harlow Plant Services, evaluating AI voice for inbound enquiries. A customer, referred to here as the caller, dials to ask about availability for a specific piece of equipment. The request involves a name, a hire period and a site postcode.
Under a SIP deployment, the call enters through the company's internet connection, is routed to the AI platform's nearest data centre, and the agent begins processing speech. Ownership of the enquiry sits with the AI agent until it either resolves the request or escalates to a human. Every second of processing requires the AI to pull availability data from a back-end system while simultaneously managing the audio stream. If the internet connection experiences jitter at this moment, the caller may hear a half-second gap in the agent's response, perceive hesitation and begin talking over the agent. The agent, hearing the interruption, may restart its response. From the caller's perspective, the conversation feels unstable.
Under a PSTN deployment, the same call travels over a dedicated circuit. The audio stream is isolated from internet congestion, so the agent's response arrives at consistent intervals regardless of what else is happening on the company's data network. The caller experiences smooth turn-taking. The AI agent still needs to retrieve availability data, but the audio path itself introduces no variability. The request moves from receipt through ownership to completion without audio interference.
Neither scenario says anything about whether the AI gives the correct answer. That depends on how well the back-end data system is integrated. This distinction matters when diagnosing problems: a pause alone does not establish whether the cause is the network, speech processing, a model response or a connected business tool. Separating audio quality issues from knowledge or integration issues is an important evaluation discipline when reviewing pilot call recordings.
Criteria for Comparing Providers Within Each Architecture
Once a business has chosen SIP or PSTN, provider selection introduces a second layer of quality variation. The following criteria are worth examining when comparing providers, along with reasoning for why each matters in a voice AI context specifically.
- Data centre geography relative to your premises: Physical distance contributes irreducible latency. A provider whose nearest node is in a different country adds latency regardless of how good their network is internally. For AI voice, where conversational turn-taking depends on low latency, proximity should be a shortlist criterion rather than a secondary consideration.
- Codec flexibility and defaults: Ask what codec the provider uses by default and whether it can be changed per trunk or per call. If a provider defaults to a compressed codec without disclosing this, audio quality may fall below your expectation without any obvious explanation. Request a specification sheet, not just a marketing description.
- Failover behaviour during outages: Ask the provider to describe specifically what happens to active calls during an unplanned outage. Do calls drop, or does a failover route activate automatically? How long does failover take? For a customer-facing AI agent, a call that drops without explanation damages trust more than a brief period of unavailability.
- Traffic prioritisation mechanisms: Some SIP providers offer Quality of Service configuration that prioritises voice packets over other traffic types. Others do not. If you intend to run AI calls across a shared business connection, knowing whether QoS is available changes the risk profile of SIP deployment substantially.
- Transparency about interconnect routing: Some SIP providers route calls through multiple intermediaries to reduce their own costs. Each hop adds latency and a potential failure point. Ask your prospective provider how many network hops a typical call traverses and whether they can show a representative route trace.
Designing a Practical Acceptance Exercise Before Go-Live
A vendor pilot is more useful when it is structured around observable evidence rather than general impressions. Before running test calls, define in advance what acceptable quality looks like for your specific use case. For an inbound customer service scenario, acceptable quality might mean the AI agent completes a standard enquiry without the caller needing to repeat themselves, and without any audible artefacts that a staff member, listening to a recording, would flag as distracting.
Run test calls at varied times of day including your anticipated peak period. If your business receives its heaviest call volume on Monday mornings, run pilot calls on Monday mornings. Record the calls, then have someone unfamiliar with the technology listen and annotate any moments where the audio quality, pacing or apparent hesitation affected the experience. This blind annotation approach removes confirmation bias from the evaluation.
Also test what happens when things go wrong. Simulate a dropped call mid-enquiry and observe whether the system logs the partial interaction and whether a returning caller is handled gracefully. Examine whether caller identification works correctly on a second call, and whether the agent treats the returning individual as a known contact rather than a new one. Ownership of the incomplete enquiry should be traceable: a staff member reviewing the system after the dropped call should be able to see what was requested, what was answered and what remained unresolved.
Handover Considerations When AI Escalates to a Human Agent
Regardless of network architecture, a frequently overlooked quality dimension is what happens at the moment an AI agent transfers a call to a human. If the handover involves audio re-routing, the caller may experience a gap, a click or a change in audio quality that signals the transition. On a SIP system, this re-routing is a software event and can be engineered to be nearly seamless if the platform supports warm transfers. On a PSTN system, transfers involve switching equipment and may introduce a more noticeable pause.
From the caller's perspective, a smooth handover preserves the impression that the organisation is coherent and prepared. A human agent who receives a transferred call without context about the conversation so far must ask the caller to repeat themselves, which undermines the value of the AI interaction that preceded it. The human agent should receive a summary of the AI conversation before the caller speaks again. Whether that summary arrives by screen pop, in-ear briefing or a text notification is a process design decision rather than a network decision, but it depends on the integration between your AI platform and your staff-facing tools being properly tested before go-live.
When evaluating a voice AI deployment for your organisation, the most useful question to ask may not be which protocol is better in the abstract, but which protocol your existing infrastructure can support reliably enough that audio quality stops being a variable and allows caller experience to be shaped entirely by what the AI says and remembers.

