The network layer has traditionally been the piece of the CX stack that nobody wanted to own.
Like the ugly sweater that your auntie bought you for Christmas, you’ve got to keep it, but you stuff it into the back of the wardrobe and try to ignore it.
The network layer equivalent is pushing it off on the telco, moving on, and worrying about the things you can actually control.
However, that attitude is becoming harder to sustain.
AI voice agents are now deployed at scale across platforms like Amazon Connect, Webex, and Genesys – and they have latency requirements that expose weaknesses in legacy infrastructure that contact center teams were never designed to manage.
Simultaneously, BT Openreach’s permanent shutdown of the UK’s PSTN network in January 2027 is creating a hard deadline that will force enterprises to confront infrastructure decisions they have been quietly deferring.
“This part of the transformation is going to fly under the radar for a lot of companies,” predicts Zeus Kerravala, Principal Analyst at ZK Research.
“They’ll probably go through the migration of PSTN to SIP, but not fully understand the implications of what they’re getting themselves into.”
The implications, as it turns out, are pretty significant.
Zeus Kerravala shares more of his insights and opinions on the network layer debate in this exclusive interview with CX Today.
Why 800ms Is the Number CX Teams Need to Know
The technical case for why the network layer matters now comes down to latency; specifically, what happens to it when you introduce an LLM into a voice interaction.
Traditional SIP networks were designed for human-to-human calls. A delay of around 200 milliseconds on a human-to-human call is barely noticeable, and both SIP and PSTN networks were built with that tolerance in mind.
The problem starts when an LLM enters the picture. Kerravala explains that the model introduces an additional 200 to 300 milliseconds of processing time on top of existing network latency.
Combined, those numbers start to add up – and when it gets above 800ms, the point at which industry benchmarking suggests AI voice conversations begin to break down, the experience deteriorates quickly enough that customers notice.
“The traditional SIP configurations were designed for human speech,” Kerravala says. “When you start to mix in an LLM, that’s when it becomes noticeable.”
From a CX perspective, it can be simplified to this: when a customer says the bot feels slow, somebody has to be able to find out why.
In most organizations right now, nobody owns that problem clearly enough to answer it.
The PSTN Reset: Opportunity or Like-for-Like Swap?
The UK’s PSTN shutdown gives this conversation something of a hard deadline.
BT Openreach’s permanent switch-off of the traditional copper telephone network lands on January 31, 2027. Every enterprise that is still routing contact center calls over PSTN has to migrate to IP/SIP before that date.
The question is whether enterprises treat that migration as a genuine infrastructure reset or simply replicate what they already have on a newer network.
Kerravala’s advice is that “you shouldn’t be taking this new technology and making it look like the old technology.”
The PSTN worked the way it did precisely because of its inefficiency. It managed quality through constrained bandwidth – a model that guaranteed a consistent experience but also masked the underlying network’s limitations.
Moving to SIP means joining a shared, high-traffic IP network that operates on entirely different principles. The reliability assumptions that CX teams have built their operations around no longer automatically apply.
“If we’re going to be running on a network that carries other traffic, more traffic, shared traffic, when a customer says the bot is slow, the CX team has to have the right tools in place to go see why,” Kerravala says.
Kerravala’s thoughts are echoed by Forrester’s 2026 network infrastructure predictions, which anticipated that business demands from AI systems and conversational AI-powered service desks are now “driving substantial organizational changes” in how enterprises approach network operations – pushing what was previously an IT concern squarely into cross-functional territory.
Kerravala also believes the volume implications could surprise people:
“I’ve always said that the only interface that we’re all born with is voice. If it works well, we’re going to use that as the default.”
His prediction is that voice traffic could actually increase in the era of AI agents — which makes getting the infrastructure right more consequential, not less.
The Visibility Gap – and Why CX and IT Need to Work Together
A tool often cited in this conversation is Cisco ThousandEyes, which has been repositioning network observability as a CX capability rather than a pure IT operations concern.
At MWC 2026, Cisco unveiled AgenticOps and broadband assurance innovations specifically aimed at the visibility gap for AI workloads: the problem of understanding performance across the parts of the network an enterprise doesn’t directly own or control.
ThousandEyes, Kerravala explains, is one of the few observability tools that can span not just a local network but across the internet – identifying where slowness is occurring, whether it sits in AWS, a carrier network, or somewhere in between.
“We’re not troubleshooting CX by looking at a red-green dashboard anymore,” he says.
“It’s more the fact that things aren’t working properly; brownout situations where things are, as they say, too wrong for too long. That’s where observability becomes important.”
The question of who owns that conversation (CX or IT) doesn’t have a clean answer yet.
Kerravala’s view is that the best organizations won’t try to assign it to one team:
“I think it’ll be the CX team and the operations team working together, because if we’re using these things for customer interactions, that has to become our priority.”
The business case for that can be summed up by picturing a scenario where one or two bad interactions is enough to damage brand loyalty, which means poor AI voice quality moves from being a technical inconvenience to a revenue problem.
Kerravala believes that “the only way to do that is to bring those teams together, and it needs to be driven from the CIO down.”
Gartner’s Infrastructure and Operations Trends Report for 2026 identifies agentic AI and hybrid computing as the top two I&O priorities for the year.
Both depend on network performance standards that most organizations are only beginning to design around.
Edge Computing: Experimental Today, Mainstream by Next Year
One emerging response to the latency challenge is edge deployment, which moves AI inference processing closer to the user rather than routing everything through a centralized cloud and back.
Kerravala describes the concept through a practical lens:
“If I’m issuing voice commands to a self-driving car, I really want the car to process that data locally.”
The same logic applies to in-store retail environments, smart stadium technology, and any high-volume CX interaction where round-trip cloud latency is a constraint.
For contact centers specifically, he sees a tiered model emerging: edge processing for local or in-person transactions, regional hubs for companies with geographic footprints, and centralized cloud infrastructure for global or less latency-sensitive workloads.
Forrester’s 2026 predictions lend some weight to that trajectory, noting that IoT and 5G growth are already pushing microservers to the network edge across industries – with factories, hospitals, and autonomous vehicles among the early adopters.
Kerravala describes the current state of contact centers as “somewhat experimental,” but his timeline is short.
“I think if we have this conversation in a year from now, it’ll be mainstream in the CX world.”
Given the pace at which AI infrastructure has moved over the past 18 months, that prediction is less contentious than it might appear on the surface, as Kerravala explains:
“It’s the fastest-moving technology I’ve seen, and I’ve been covering tech a long time.”
What Good Looks Like in 2026
For enterprises trying to get ahead of this, Kerravala’s practical recommendations follow a clear logic.
The first is a shift in how organizations conceptualize their infrastructure. “You need to stop thinking about your network as separate networks,” he says.
“CX transactions don’t care that you’ve got four different WAN network managers. You need to think about that transaction end-to-end – from the mobile device to the cloud – as one network.”
The second is moving away from traditional uptime monitoring toward observability tools capable of identifying performance degradation before it affects customers.
“All networks are built today with resiliency in mind. You could probably go turn something off, and nobody would notice. It’s the performance issues that are harder to manage.”
Those observability tools need to cover the entire end-to-end network; not just the parts the enterprise directly controls.
The third is organizational: siloed teams managing a unified network is a setup for finger-pointing rather than resolution.
And the fourth is using AI to stay ahead of the problem. “AI can help you move from reactive troubleshooting – where the user is complaining about the transaction being slow – to being more proactive and predictive,” Kerravala says.
“Let’s say over time your call volume continues to go up. AI can predict the moment at which your call quality will degrade to the point where users will complain. Stay ahead of it.”
The network layer is no longer an infrastructure concern that CX teams can outsource to someone else and revisit when things break.
In an AI voice era, it is a CX concern – and the enterprises that recognize that before their PSTN migration deadline will be in a very different position to those that don’t.
