People don't typically get excited about networks, because they are viewed as simply pipes connecting applications together -- the applications are where the interesting stuff happens. Email, video, data, and voice all ride on top of the pipes. When the pipes work, they are unappreciated.
The problem is that voice is one application that suffers on a bad network day, whereas an application like email would be tolerant.
But a voice call is fanatical about network performance. The moment your network has a glitch, voice breaks.
Which is why, when we talk about SD-WAN in 2026, we should also be discussing whether the SD-WAN underlaying your voice traffic is actually optimized for voice. In many deployments, it isn't.
The reason is that voice has a different tolerance profile than other applications is defined by three metrics: latency (the time it takes for a data packet to travel from a source to destination across a network, < 150 ms), jitter (the variation in delay for data packets traveling across a network, < 30 ms), and packet loss (the amount of data packets lost traveling across a network, < 1%). All three need to remain inside their narrow tolerance bands throughout the call which requires real time network performance visibility. If any of these tolerance are not met, calls will be subjected to garbled audio, awkward silence, or even a dropped call.
A voice-optimized SD-WAN deployment is one where the controller and the edge routers are meticulously tuned for application policy and then monitored in real time for traffic behavior. Here are the features that do the heavy lifting.
Application-aware routing in milliseconds: the edge routers are directed by a decision-making network controller about where to send each packet. The edge routers react within milliseconds to shifting network conditions which is much quicker than legacy routing protocol reaction times.
Forward error correction: when a voice packet is lost in transit, a voice-optimized network can rebuild it. The audio data is sent with mathematical redundancy that the receiving end can use to reconstruct lost data. On a link that's struggling, forward error correction can make the difference between a low and high quality conversation.
Continuous link probing: the SD-WAN edge sends frequent probes to every available path to measure latency, jitter, and packet loss. When a path degrades below a quality threshold, the controller directs the edge router to send traffic over a healthier path. This failover mechanism keeps calls which would have been dropped in legacy networks, connected.
Quality-of-service enforcement: voice packets are favored over non-real-time traffic and receive preferential handling across the network.
Packet Duplication: is an SD-WAN loss-protection technique in which the sending edge router replicates each packet of a selected flow, typically voice, and transmits the copies simultaneously across two different WAN paths. The receiving edge forwards whichever copy arrives first and discards the duplicate. This prevents call disruption should a path go down during a call.
These SW-WAN features are not new, however they are rarely implemented on most SD-WAN networks.
MPLS networks prioritized voice traffic, but their excessive costs pushed corporations towards less expensive Internet connections. During the migration to Internet, the voice traffic prioritization did not follow.
It is because of this that voice quality starts performing badly within a few months into an SD-WAN deployment. Dropped calls become more common. Contact center agents report that customers are hard to hear. Teams or Zoom calls become flaky for branch office employees. The network team investigates and finds that the SD-WAN is doing exactly what it was configured to do: routing traffic efficiently and properly performing failing over based on path outages instead of path metrics.
Voice AI agents sharpen all of this. When a human is on a call and the network stumbles, the person on the other end asks for a repeat and the conversation continues. However, when a voice AI agent is on the call, there is no one filling in the gaps. A short burst of packet loss can be the difference between the AI understanding the request and returning the wrong answer, which means the network layer underneath voice AI deployments is under more pressure than ever.
The cost of that drift is real. A dropped sales call is lost revenue and a dropped support call is a angry customer. Bad quality internal meeting calls are many expensive minutes of billable time that need to be rescheduled. These issues caused by network performance issues create hidden operational costs.
If you're evaluating a new SD-WAN infrastucture, refreshing an existing one, or inheriting one from a predecessor, voice optimization should be a specific request of the vendor. Ask the following questions during vetting:
If the answers come back as “we treat voice like any other application” or “failover takes a few seconds,” look for another vendor. If the answers come back as “sub-second failover with voice-prioritized path selection and forward error correction on by default,” you're on the right path.
Networks are supposed to be boring. The applications that ride on top of them are where the user experience lives. But voice is the one application your organization can't fake its way through, and the SD-WAN underneath it has to be designed for that. Treating voice as a first-class concern on the network has moved from edge case to mandatory requirements.