The ongoing challenges of Voice AI are by design - How do we fix it?

The gap between voice AI investment and customer reality has never been wider. However, this problem isn't technical, it’s a design choice in which these systems deflect customer demands rather than address them directly.

With smaller call centres handling between 50 and 100 calls per day and larger, automated operation handling even more, the difference between what is promised and what is delivered is apparent. Across healthcare, insurance and financial services, the same pattern keeps surfacing - a business invests in AI, the demo goes well, positive press follows, and yet the customer on the phone is still being asked to repeat themselves six months later.

Conversational AI is forecast to eliminate $80 billion in contact centre labour costs by 2026, and while this evolution is significant, most of these developments are stalling or failing before reaching production. Understanding why is the key to overcoming these barriers and delivering more positive customer outcomes.

Resolution is the wrong goal

Customer service was built around resolution, tickets, queues, case closures, and handle times. It is a system designed to manage problems after they have already happened. Most customer service-focused AI has been dropped into the same model without question. The AI answers the first question, routes the call, and then summarises the issue. The moment the case becomes complicated or commercially meaningful, it hands it off to a human and calls it a resolution. The businesses getting this right are not asking how to resolve customer problems faster, but are asking how to remove friction before a problem occurs. The answer to this question starts at the foundation, with infrastructure built for real-time conversation rather than treated as an aesthetic afterthought.

What the technology needs to do

When it comes to this technology, the biggest challenge is not the AI model but orchestrating the experience underneath it. Voice is less forgiving than text, as people notice a half-second of lag, an awkward pause, a handoff that loses the thread of what was said five minutes earlier. The best models account for latency, interruption handling, and context continuity, aspects that help them hold up in a real business environment.

That experience shapes how we approach what this technology needs to accomplish in practice. Not answer questions, most systems can do that. Not a route call, that has been automated for twenty years. The harder requirement is continuity: a system that carries the full context of a customer's situation from the first second of contact through to the moment their problem is actually gone. Not logged or escalated, simply gone.

The shift that is actually happening

The technology conversation tends to dominate the coverage — latency benchmarks, language counts, model comparisons. But the more consequential shift happening right now is not technical. It is architectural.

For most of the past decade, serious contact centre AI was the preserve of organisations large enough to absorb a bespoke implementation, an extended rollout, and the occasional expensive failure. Everyone else used what was available, which usually meant rigid phone trees and a chatbot that handled three question types and gave up on the fourth. The access gap was real, and most businesses accepted it.

That gap is beginning to close. Integrating AI-native voice infrastructure directly into platforms means that a business that could not have justified a serious AI deployment six months ago may find the same capability available without a long-winded procurement process. That matters beyond the technology story. It changes who gets to compete.

The questions worth asking now

For businesses that depend on customer conversations, the question is not whether AI is making a meaningful presence in contact centres, but whether this technology is removing friction from the customer experience, or simply automating the same queue with better language.

The best customer service interaction is the one that never had to happen because the information was already there, the process had already run, and the problem was already solved. Businesses that orient themselves around that goal now, rather than optimising the ticketing system they already have, will find themselves in different positions in the future.

‍

Hakob Astabatsyan is CEO and co-founder of Synthflow AI, an enterprise voice AI platform that has handled over 65 million customer calls across healthcare, insurance and financial services. Synthflow recently partnered with 8x8, Inc. to bring AI-native voice infrastructure to enterprise contact centres globally.

‍