The Reality of Voice AI in the Contact Centre: From Pilot to Production - AudioCodes

In this CX Today interview, Marcus Law speaks with Ilan Avner, Director of Product Management at AudioCodes, about the practical reality of moving Voice AI from proof of concept into production.

The conversation explores why so many Voice AI projects stall after promising pilots, what enterprises often underestimate about voice infrastructure, and why connectivity, latency, noise handling, scalability, redundancy, and human-agent handoff matter just as much as the AI model itself.

Avner also discusses how platforms such as AudioCodes Live Hub and AI Agents can help enterprises build a more resilient production environment for Voice AI use cases, including voice bots, agent assistance, and real-time translation.

Watch the full interview to learn what CX and contact center leaders should ask before putting Voice AI into production.

Transcript

SPEAKER_00 0:08

Hello there and welcome back to CX Today. There's no shortage of enthusiasm around Voice AI right now, but enthusiasm and production deployment are two very different things. For a lot of organisations, the pilot is where the journey stalls. Today I'm joined by Elan Avner, Director of Product Management at Audio Code to dig into what separates a successful deployment from one that never leaves the test environment. We'll cover the infrastructure questions, the human handoff problem, how to build an internal business case, and what the next 18 months look like for the contact center. Elan, thank you so much for joining us today.

SPEAKER_01 0:44

Yeah, thank you. I'm glad to be here.

SPEAKER_00 0:47

Well, as we touched on, there's so much enthusiasm around voice AI right now. It's been one of the biggest topics in enterprise tech for the last couple of years. From where you're sitting at Audio Codes, how much of that investment is actually making it into production and how much of that is still stuck at the pilot stage?

SPEAKER_01 1:05

Yes. Okay, so uh Audio Codes, uh we at Audio Code started the uh the engagement with Voice AI about six or seven years ago. Uh and we have all sorts of customers, and uh the experience is not exactly the same among all. Okay, so maybe I will speak about that. There are multiple uh types of the customers. Uh the the first of them are the ones that started together with us about six years ago. Uh and they started with uh the the legacy technology, NLU, flow editors. Uh and when they started, it was not that uh good. Um but since the introduction of the LLMs, uh they see much better uh performance of the bot. It's easier for them to develop the bots, uh and they have a lot of experience, they know their use cases. And this type of customers already went through all the struggles and the challenges of a uh deployment, and they're currently in a massive uh production environment. They automate a lot of use cases. We are surprised to see how much, okay? But this is one thing, okay? Most of the other customers, okay, and this is the most of our customers, and I assume most of the organizations as of today didn't really start production for Voice AI at the moment. And there, and with the new technology of LLMs, uh, they can do POCs very quickly. And the POCs uh sound great, okay, but then they hit a lot of challenges that the other customers hit. Uh and uh these are related to uh the connectivity uh to multiple other things. So usually we will speak about the challenges, uh uh, but uh we see them hitting the challenges. The performance of the bots are not as good as the POCs, and then they halt uh production and they they train again, they they optimized the performance, and then it takes them a few months until they get into production.

SPEAKER_00 3:02

So given the gap that you've just described, there's definitely a gap. Yeah, when when a voice AI deployment stalls or fails, what are the most common reasons you see for that? And what's the thing that enterprises consistently underestimate?

SPEAKER_01 3:18

Okay, so you know that there are multiple components for the voice AI stack. Okay, so you have uh the text-to-speech, the speech to text, you have the bot framework. Um, a lot of things are where you develop your bots or the LLMs, okay? Uh a lot of things are related. So when you think about issues, usually you think about the conversational AI platforms and the LLMs. And yes, there are challenges there related to hallucinations or following the business logic uh uh and uh uh I don't know, uh edge cases uh that may happen that we don't think about in advance. So these are things related to uh the the conversational AI platforms. I I think in this uh uh in this uh session we will not speak about, we will focus more on the voice channels and the voice AI, uh the voice infrastructure. And when you speak about the voice infrastructure, this is the one that is underestimated. And this is the speech to text, the text-to-speech, and also the connectivity to the voice channels, uh which are the contact centers. Usually, again, when you do a POC with allocate the phone number and call your bot, but when you go into production, you need to integrate with your voice channels. And this is usually underestimated because you have challenges with integration with the voice channels. Um, how do you escalate calls to human agent? When you escalate, how do you share context with the human agent? Um for the speech to text, what about noisy environments? How do you perform in noisy environments? Text-to-spee should be native. Um so these are the type of challenges. And I think we'll also elaborate more uh in the next questions, I assume.

SPEAKER_00 4:57

Yeah, absolutely. Um yeah, to stay with that sort of the technical side, uh, you know, a lot of a lot of public conversation around voice AI is focused on that model layer, you know, as you've just touched on, you know, which LLM and how the how the agent is designed and how the conversation flows. So yeah, how much does the voice infrastructure beneath this actually matter? And what goes wrong when that's not quite right?

SPEAKER_01 5:21

Okay, so the the bot the voice infrastructure is the basics. If the voice is infrastructure is not good enough, uh it it doesn't matter how good your bot is, the bot would not function well. If you garbage uh what we see, you say garbage in, garbage out. Uh so if you uh um the speech to text, for instance, is not accurate, the bot the voice bot cannot uh cannot be accurate as well, cannot uh perform well as well. Uh so uh and what happens in case you hit a noisy environment? Usually that we you would not test it in uh in it uh uh before you go into production. So you need to be able to uh uh to handle noisy environments, to handle load, to handle the capacity, to be able to connect to the uh voice infrastructure, all the contact centers. Um what happens in case uh, for instance, uh the speech text is not responding, what happens with latency, what happens with multilingual, um these are the type of things that uh usually are challenges.

SPEAKER_00 6:28

Uh I'd love to look at this idea of what good looks like. And before we get into you know the approach that audio codes take specifically, I'd love to get your sort of views on this. And you know, uh from your side, can you uh describe what a production ready voice AI contact center actually looks like in practice? Yeah, what has to be in place before a deployment is genuinely ready to go live?

SPEAKER_01 6:56

Okay, so one thing that you need uh first uh contact center automation, let's see what we're speaking about. So that's voice bot, but not only voice bots, also agent assistance. Okay, so uh when we say agent assistance, we we mean that not all of the traffic will go to a voice bot. We'll still have live agents handling the most complex uh environments, uh uh calls, and then uh you need to help them. You can uh have a bot, so a bot that will help them to uh address queries faster and more accurate. So these are the types of uh automations you do on the contact center, and what do you need on the infrastructure? So um first you need uh uh related to the infrastructure, you need to build an infrastructure that can uh that the technology constantly changes. Okay, and um uh we spoke about latency, noise reduction, things like that, constant how native the voice is, to express emotions and so forth. Uh so that means that uh you need to adapt to the technology as it emerges. Because so first thing I would say to anyone who wants to build a contact center voice AI environment is to choose a platform that can adapt to the technology. That means that you can adapt uh new providers as they come. And when we say providers, that means speech to text, text-to-spee, and also conversational AI platforms. Of course, you need to also um be able to protect your investment. So, what happens in case your voice channels change? Okay, so in case you migrate your contact center to the cloud, you need to think of that as well. So you would not need to reinvent your voice AI stack when you move your voice channels. So you need um to decouple that from your contact center. Um more things that you must think about before you go into production of things related to scale, to redundancy. Uh, when we say redundancy, what happens in case the speech text is not responding, or the text-to-speech, or if you have latency issues due to uh late uh due to scalability, uh and also monitoring and debugging and testing before and testing real-world use cases.

SPEAKER_00 9:12

I think what we're talking about now sort of brings us hopefully neatly onto the you know where audio codes fits into this picture. Uh you know, specifically with with Live Hub, you know, could you tell us a little bit about Live Hub? What problems do it does it solve and and why does it matter specifically in the context of getting voice AI into production?

SPEAKER_01 9:31

Yeah, uh the main idea of Live Hub uh is to give you a single platform that you can develop all your voice AI use cases with, with the ability also to choose your providers. Okay, so it doesn't mean that all the voice AI stack, of course, not the all the voice AI stack is in Live Hub. It can you can integrate with external third parties, text-to-speech, speech-to-text, conversational platform. So um that and also connect to the voice channels. So the main idea of Live Hub is a single platform that you can develop all your voice AI use cases from VoiceBot to agency system to return translations, all on a single platform, and uh to address all the challenges related to going into production. And these are first and foremost the connectivity to the voice channels, uh, the ability to change voice channels in case you need, the ability to choose providers, the ability to uh change your providers in case you see a provider that is more advanced. Uh and uh the ease of use is another thing related to Lava because it is a self-service uh platform that gives everyone the ability to uh develop their voice, both connect them to their platforms and connect them to their voice channels.

SPEAKER_00 10:47

Now, one of the things that comes up a lot when we are speaking to CX leaders is this handoff moment and the point where the AI reaches its limit and a human steps in. And you know, if organizations get that wrong, you've undone a lot of the goodwill that this automated interaction's built up. So, how does that idea work in a well-designed voice AI contact center? And yeah, how how does audio codes uh approach come in there?

SPEAKER_01 11:13

Yeah, uh you touch a very complex uh point because uh this relates to the integration with the contact center. There is no one answer to that, okay? Uh because that depends on the contact center you use. Uh now we have in mind connectivity. This is what we have in mind always. So we allow the connectivity to any contact center, and this includes also the escalations to human agents, although the options to share contacts with the human agent. And then the integration is done differently per contact center. This means that we did an effort per contact center. And once you uh use live harm, we will either guide you on how to do that, or we have a managed service so we can do the integration for you. And that would be different for Nice in Contact, for uh Amazon Connect, for Five9, for Cisco, for Genesis, for Avaya. Each one of them works differently, the integration is done differently. And this is uh our brand and butter that the things related to integration and connectivity.

SPEAKER_00 12:16

So let's say that an organization has run a successful pilot. Uh, you know, often after a successful pilot, the next challenge isn't external, but it's actually internal, whether it's finance, risk, and operations all needing to be brought along. So, what does a compelling business case for a full voice AI production deployment actually look like? And what's the risk to businesses for not moving forward?

SPEAKER_01 12:40

Okay, let's start with uh the risk. And I I will put aside uh and I will speak also afterwards about the common things, the most common things that the ones that anyone thinks about. But the first one is um the your enterprise uh perception, okay, because enterprise, each enterprise has a competition. Uh and uh you will find out that your competition adopted voice AI and you will be left behind. So that means that you will be perceived as an old organization. And your competition will be perceived as a modern one that uh uses voice AI. So your perception is very important. You don't want to be perceived as an old organization that is not moving forward with the technology. And then you can go to the most basic things. Uh, what is the reason what are the reasons to use the AI in general, okay, or uh voice AI for the contact center. So that would be first the user experience. Okay, that also really relates to perception. Uh, because usually we go when you go to contact centers, you would uh you would call a contact center, you would uh have the whole time. You will wait a long time until someone will answer. Okay, so this is one of the reasons to adopt voice AI. And you will also be surprised that uh voice bot can give you a better user experience also when they speak to the customer. And the other thing is that the cost uh the cost saving for your contact center. So these are the most common ones, okay.

SPEAKER_00 14:08

Well, Ilan, it's been a great conversation. I'd just love to wrap up with the view from here and a bit of a look to the future. So, where do you see voice AI in the contact center heading in the next 12 to 18 months? And what should enterprise CX and IT leaders be preparing for now?

SPEAKER_01 14:25

Okay, so that that's that's a hard question. If you would have asked me uh one year ago what would happen in one in today, uh, I don't think I would anticipate uh what the technology would look like. Uh so we we have so many changes in the technology, but uh there there are some things that are clear. Okay, so one thing that is clear is that the technology constantly changes. Okay, so what I recommend is uh to choose the platform that uh that can adapt to the new technologies. It's a must today. It's a must. You must always constantly uh look for the technologies, the best ones, and adopt them as well. Uh otherwise you will be left behind. If you ask me what I think would happen in 12 months or 18 months from today, I I think it will be very common. You would not be surprised if you would call a contact center and you will uh uh and you will speak with a bot. That will be no surprise. That will be very common. I am sure about that. Uh uh already today I'm speaking with bots and you will see it more and more. Uh again, also for the agent assistants. And uh yeah, we also see today bots speaking with bots. So maybe you will have a bot representing you speaking with the contact center. So uh I that may also happen. I assume that will happen as well. Um and yeah.

SPEAKER_00 15:50

Fantastic. Ila, that's a that's a brilliant note to end on. It's been great speaking to you. It's been a really you know you know great perspective on where these friction points are and what production grade of voice AI actually requires to work at scale. So thank you so much for joining us.

SPEAKER_01 16:02

Thank you so much.

SPEAKER_00 16:04

So if you want to go deeper on anything we've covered today, we'll have written content linked below, including a look at where why enterprise voice AI projects stall before they reach production and a closer examination of how Live Hub fits into existing content centre infrastructure. But until then, I've been Marcus Law.