
From enhancing customer experience to capturing the B2B market, voice AI offers big opportunities to telcos. But as operators bring AI into voice services, they soon find that many AI solutions weren't built with strict carrier-grade standards in mind.
Additionally, much of the AI now reaching phone calls arrives over the top (OTT) – including device makers, apps and AI receptionists that businesses connect through simple call forwarding – and some providers simply resell such services. Either way, call audio leaves the network and is handed to a third-party service. The OTT integration looks faster to stand up, costs less upfront and carries less operational responsibility. But the convenience comes with a loss of control and ownership. The OTT provider ends up owning not just maintenance and updates, but the value-generating layer itself.
Author: Tomasz Człapski, Software Delivery Director at Software Mind (photo)
There is an alternative solution that operators can choose to regain control over service quality, business case and revenue. Integrating AI directly into the voice path means you can turn your AI-enabled voice services into an asset and use it to create new revenue opportunities. Here's how it works.
In-path AI integration puts control back on the operators' side
First, let's look at what operators gain by keeping AI in the voice path. One of the clearest benefits is lower latency. A pause might easily go unnoticed in a chat conversation, but it becomes an audible problem on a live call. Audio routed to a distant data center adds round-trip time that makes transcription, translation or a bot response feel slow. Processing media inside your own network keeps any delay within the tolerable limits customers expect. The in-path integration also lets you decide where each AI engine runs – on-premises for the most latency-sensitive services or in the cloud where a longer delay is acceptable.
The same thing happens with call reliability. Consumer AI tools are built for best-effort availability. But voice networks have to meet carrier-grade service-level agreements – typically five nines, or roughly five minutes of downtime a year. Route a call through an external service and the uptime slips out of your hands. An AI capability that lives in the voice path is managed like any other network element – with high availability, latency-budget enforcement, traffic prioritization and automatic rerouting to alternate engines – so it meets the standard the rest of the network already holds.
This high level of control enables you to generate new services from your voice network. You can offer live translation or call transcripts to consumers, packaged AI services to business customers and – because data residency is under your control – voice services to regulated sectors like financial services and healthcare, where a bank can connect its own model to the in-path infrastructure without exposing sensitive information to an outside network. In-path integration turns voice into a platform you can build on, rather than an external capability you rent.
What an in-path voice AI looks like in practice
The AI voice framework our team at Software Mind uses when helping operators expand their services is one way to achieve carrier-grade AI in voice. It runs on the same proven media engine behind our other Amplitiv voice solutions, deployed with more than ten operators across Europe, most of them Tier 1, and constantly developed to address new market demands.
When a call connects, it is routed through our customizable Amplitiv AI Media Server, which adds an AI layer to telco-grade media processing. A built-in AI gateway combines live call audio with session metadata and triggers the relevant features, such as real-time or post-call transcription, call summaries, live translation, sentiment analysis, fraud detection, dynamic voice prompts, voice bots, noise reduction and speech enhancement.
Recordings and generated content are stored securely inside your environment. The framework is pluggable: it works with cloud services, self-hosted models or a client’s own AI, so operators are never locked to a single vendor as models change.
The success of your voice AI is in how you run it
A common issue emerges from our discussions with telcos. Operators are reluctant to layer AI services onto the general-purpose media servers that already sit in the voice core, such as the IMS Media Resource Function. Those nodes are built for stable, high-volume base voice services, not for frequently changing AI-driven VAS.
Adding a dedicated, AI-enabled media server alongside the core offers a simpler solution that doesn't alter base voice services. Instead, it allows you to run AI workloads on a flexible component that can be scaled to meet dynamic demands and that becomes the foundation for developing future services.
To make the most out of your AI-powered voice services, keep the AI layer pluggable so you can switch providers or run several in parallel as needs change. As for the best place to start – look for use cases where value is highest and effort lowest. Real-time transcription enriches calls that already happen with minimal change, while voice bots, which need dialogue design and domain knowledge, are better added once you establish a strong foundation.
Bringing AI into voice is less about the model and more about where it runs. Run it in the path, and voice becomes an asset you own and can grow.
