Gidi Adlersberg: I can tell you that beyond the obvious criterias for, for this kind of experience, for like things like latency and accuracy of information, all the things that you would expect from a bot, part of the criteria was actually also level of empathy, and is the bot able to maintain a consistent level of empathy and things like that. So, so it's very well thought of actually.

Tom Arbuthnot: And welcome back to the show. This week, we are talking Teams Phone Agent, one of the hottest new capabilities in Microsoft Teams currently in the Frontier Program. This is Microsoft's Teams voice AI capability, kind of a voice attendant or voice receptionist. We talk about what's in the box, and we also talk about the extensible capabilities and the new Unify integration model for Teams Phone Agent. AudioCodes are the first certified solution for Teams Phone Agent extensibility, and Gidi takes us through what that means, what capabilities are in the box versus where partners bring value. And also, we generally talk about some voice agent projects and where the pitfall are and where some of the benefits are as well. Many thanks to Gidi for jumping on the show, and also many thanks to AudioCodes for being a supporter of the community. Really appreciate all their support. On with the show. Hey, everybody. Welcome back to the show. Really excited for this one. We're gonna be talking Teams Phone Agents and Teams Phone Agent extensibility as well. Voice AI is, like, the hottest topic in our space at the moment. As this goes out, I just recorded a podcast with Chris at Maersk talking about their voice AI journey, and everybody's is talking about it, experimenting, and now we have Microsoft's option in Frontier. And Gidi's here to help explain some of that and also some of the extensibility where the certified partners can come in. And AudioCodes are actually the first certified partner for the Teams Phone Agent extensibility. Gidi, you can help me with the proper, proper program name now it's settled. But first of all, Gidi, welcome back to the show.

Gidi Adlersberg: Thank you, Tom. Always happy to be here. Excited to talk about this one.

Tom Arbuthnot: Awesome. So let's just give a little everybody a little bit of context you've been on before. Obviously, a big history in our space, but give us a little bit of background to you and your role at AudioCodes.

Gidi Adlersberg: Yeah. So I'm Gidi Adlersberg. I manage the Voca CIC business line here at AudioCodes, basically our Microsoft Teams Contact Center solution, but obviously with plenty of AI in it, and I think today we're going to zoom into exactly that, so.

Tom Arbuthnot: Awesome. So I guess to set the scene for everybody listening, Teams Phone Agent is the name Microsoft settled on for their kind of a voice AI auto attendant. So it's kind of an extension of the native auto attendant capabilities. Ilya demoed it on stage at Infocomm, which was the kind of official public launch. And this is Microsoft essentially voice AI attendant receptionist, and it has some capabilities in the box. So you can add a knowledge base, you can extend it into Copilot Studio but it seems like it's been designed to be extensible from, from day one. Gidi, just a native Teams Phone Agent, like, what, what, what capabilities have you seen in that versus what you've then done with AudioCodes?

Gidi Adlersberg: Right. So I think clearly a-as, as you said, Tom, the, the story is evolving the, you know, the classic auto attendant augmented with, with AI and large language models to just, just do a lot more generally. Maybe more specifically, especially when we think about the enterprise space, there's a whole range of complexity when we think about voice agents or voice bots. Everything from highly complex, highly customized multiple integrations with knowledge bases and CRMs and whatnot, and all the way down to the just more basic experiences. Many of them, many of these basic experiences are positioned in the market just terminology by terminology-wise, they're referred to as AI receptionist. There are multiple vendors that

Tom Arbuthnot: Yeah, that's what, that's what we've heard from the other UCaaSS, haven't we? They typically say AI receptionist when they mean their kind of inbuilt capabilities.

Gidi Adlersberg: Exactly, exactly. And, and by the way, quite frankly, we adopted that terminology in Voca CIC as well. And I guess the Microsoft flavor of it is, is the Microsoft Teams voice agents, but really what they mean is, in simple words, kind of entry-level AI experience that covers the basics. But, you know when we say basics, it doesn't mean it's— it, it solves for, I would quantify it as at least, you know, a good sixty, seventy percent of, of reasons why people are calling in. So an entry level use case, more turnkey, less, less customization, more turnkey but also a lot more scalable, a lot more easy to, to work with, a lot more easy to adopt, and it solves for the vast majority of why people would call a main line.

Tom Arbuthnot: Yeah. Very, very much the Microsoft way as well, like build the, the, the fundamentals in and then leave it to the customer and the ecosystem to do the more complex stuff, the more line of business stuff.

Gidi Adlersberg: Exactly that. Exactly that. And also the— another important aspect of that maybe is just who implements it, who looks after it, right? So, you know, in the Microsoft native world, it's more of a, more of a, I would call it a license activation level, right? You get the experience enabled, but it's on you as the, as the Teams admin to, to, to configure it, to get it to work. Something goes wrong, it's you who, who needs to go and fix it with, with, with us or, or any id— any other third party that will enter that program later on. The opportunity is to get all that as a fully managed service, right? So, you know a vendor like AudioCodes looking after the deployment from, from planning through implementation and all the way to just ongoing the to— day-to-day support, optimization and so on. There's a lot going on in these voice agents, so it's good to have someone to look after those.

Tom Arbuthnot: Definitely. So l- let's take it back a step, 'cause this is kind of building on the Your Team Certified contact center and the Unify program in the Microsoft have for certification, which is kind of their top tier program for certification for contact center. Take us through kind of the Voca journey and what, what is the Unify co- certification versus this, which is something different?

Gidi Adlersberg: Right, so that's, I, I think it's, it's a bit of a funny one for us, right? Because, I mean, those in the market who, who know us, AudioCodes, and, and namely Voca CIC know that actually we started, I mean, back in, I don't know, 2020, 2021. By the way we're talking pre-LLM days, pre-Ge- GenAI days. We already had a pretty robust conversational IVR solution that has a, that had a Teams application, et cetera. Then came the whole contact center opportunity and, you know, getting to do that from scratch, which is actually a privilege, allowed us to, to jump straight on the newest, you know, technology stack available to achieve that integration. And, and of course, I mean the Azure Communication Services technology stack what later evolved to the Unify integration model for the Microsoft Teams Contact Center. What's interesting is that with this program for the Teams Voice Agents, it looks like Microsoft is positio- positioning that under Unify as well which tells an interesting story in terms of Microsoft kind of I guess, kind of telling the world, "Here's the latest technology stack. You wanna do contact center on, on top of Teams, this is what you probably wanna use. You wanna do a voice agent on top of Teams, you wanna use the, exactly the same stack." and we can get more geeky about it if you'd like, like the underlying elements and stuff, but yeah.

Tom Arbuthnot: Yeah, let's definitely dig into that. So this is the, the certification program is called the Microsoft Teams Voice Agent Certification Program, and currently the docs, this may change, but it's Teams Unify Integration Model for Voice Agents. So as you say, they're kind of tying the Unify model and that platform into the, the voice agent story.

Gidi Adlersberg: Exactly that. And, and why it makes sense from a, from a technology standpoint is simply by, like, if we take a look at the, at, at the underlying foundation or com- Azure kind of elements that make up for the model when we think about Unify we're talking about two key two key elements. One is the calling SDK, which for example, in Voca CIC, we use that SDK extensively in our agent and supervisor application to achieve the integration with Microsoft Teams Phone. In a s- in one sentence, it means that our agent application from a voice handling standpoint, acts exactly like a Teams client. So there's no, like, hairpinning, there's no API-ing. It's just a native Microsoft SDK that runs the call. So that's one leg of u- of Unify or Teams Phone extensibility is the calling SDK. The second leg is call automation. And again, keeping it high level here Microsoft call automation is a API SDK that basically does the entire call orches- orchestration and call control. And in the Teams voice agent, a, a program, basically kind of the, the, the standard there is you need to use call automation basically to manage the call from the bot all the way to the live agent. If, if there's any live agent escalation happening, you wanna get the call from the bot to a Teams user, all that orchestration, transferring, audio streaming of the bot, et cetera, is all orchestrated by, by call automation. So basically, it's the same foundational elements that make up for the Teams Contact Center Unify integration, and now also for, for the voice agents.

Tom Arbuthnot: Nice. That really helps. And so i- is this a separate product, a separate thing? Do I have to have Voca Contact Center to do this? How does that look from an AudioCodes perspective?

Gidi Adlersberg: Yeah. So, so, so what we do is we kind of we make both options available. Actually, it's fairly simple. So in simple terms, you, you, you get, you know agent seats or agent licenses from us. You can then bring these voice agents on top of those as an add-on. That's one way of doing it. Another way of doing it is to just buy it as a standalone solution regardless of any contact center needs 'cause if we

Tom Arbuthnot: So I've just— I just… I have a main number scenario where I want to tie into my line of business application. So I put Teams' phone agent on the main line, and then I use AudioCodes to help me build and platform and manage the voice agent piece. I don't need to have the Voca Contact Center to do that. But if I do, I can use them both together as well

Gidi Adlersberg: That's e- exactly that. Exactly that. And, and, and again, in terms of positioning, we position that as AI receptionist, which by the way I know this is like, here we're focused on Microsoft obviously but, you know, the wide AI receptionist offering that we have under Voca is completely multi-platform. I guess, I guess that what we're, what we're doing is when it comes to Microsoft Teams, we, we are able to go a lot deeper and bring even more capabilities and empower the experience with, with a certified solution, with yeah, maybe, maybe we'll also cover the actual experience and what it looks like. But it's more than just, here's a bot that can transfer a call somewhere if needed.

Tom Arbuthnot: Yeah, that multi-platform thing is interesting. I'm having a lot of conversations with people in contact center at the moment where they're assessing voice AI options and, and they're ending up… So Chris at Maersk was a really good conversation. If you haven't listened to that podcast, check it out. Like, like they looked at their native contact center platform, but they looked at other solutions to build the line of business voice agents independent of the primary contact center platform. So it seems like a bit of a reassessment point for a lot of organizations to see how they meet these new voice AI business solutions.

Gidi Adlersberg: Right. F- for me it makes perfect sense, and the way that I think about it, it's exactly like how we talked, you know, again, before AI, we talked about- You know, the high capacity complex contact center use cases, but we also talked about lightweight users, right? And the more informal use cases. And I guess it's, that's kind of what we're seeing here, but in, in the AI or, or, or kind of voice agent flavor of it, where, yes, the, the, you know, if we think about what's, what's the best use of these voice agents today in an enterprise, for sure it's in customer service, but it doesn't mean that customer service is the only place that can, that can, you know, benefit from it, so.

Tom Arbuthnot: Nice. And so take us through, like, what's a practical use case for this? What have customers been doing so far? What are some examples?

Gidi Adlersberg: Yeah, so, so when we think about use cases of, of, you know, this kind of experience, you're talking about everything from, you know, basic, first of all, caller authentication, which, you know, in a traditional IVR can be quite, quite the headache. So that, that alone, you know, for some brands, you know allowing customers to self-authenticate with a bot is, is, is pretty significant. But then where it gets, obviously where it gets more juicy is when you're able to and again, as part of that program that's what Microsoft is testing and that's what was implemented, being able to train the bot on, on whatever knowledge base, right? To automate just, you know, common questions and answers, troubleshooting scenarios, Everything from, like, basic information basic business information, all the way down to— I can tell you that in the, in the actual testing with, with with Microsoft for this certification, we simulated an internet service provider support desk. Okay? So it's a s- it's a scenario where the bot, you know, literally works with a calling customer to troubleshoot why the internet is slow today. Okay? So obviously that uses a lot of, you know a lot of harvesting o-of, of, of a knowledge base to be able to answer intelligently. By the way, another requirement that that's part of, of, of that, of that bot and, and capabilities is can the bot adapt in real time to the customer's not just sheer dry response, but also the, the underlying sentiment, right? So if a customer is more frustrated, we want the bot to respond a bit differently et cetera.

Tom Arbuthnot: Yeah, pacing's interesting, isn't it? You think about dealing with customers, like there could be someone who's… Like the internet is a great example. Like the customer's less tech-savvy, so you wanna be, like, slower. Like, "Can you find the router?" You know, "Are the lights on?" Like someone who's tech-savvy might want the bot to be like, "Right. This, this, this, this, this." "Yes, I've done those all." So the adaptability is a really interesting one.

Gidi Adlersberg: Absolutely, and I can tell you that, one, you know, beyond the obvious criterias for, for this kind of experience, for, like, things like latency and accuracy of information, all the things that you would expect from a bot, I can tell you that part of the criteria was actually also level of empathy, and is the bot able to maintain a consistent level of empathy and things like that. So, so it's very well thought of actually to support these, these use cases.

Tom Arbuthnot: And how does it work, Gidi, from a voice models perspective? Is this using models on Foundry? Does the, does the customer get to choose the model? Are you providing the model? How does that work?

Gidi Adlersberg: So in our case, you know, as this is provided via a managed service we will, we will you know, choose the, the, the underlying model. We do have an option to allow the customers to bring you know, bring their own. As your you know, foundry model. But yeah, in under a managed service, we like to take those decisions, you know, we know what, what will… What's going to work best for each scenario. A- and, and maybe another important thing to cover as we're talking about a voice agent, in this experience, it's purely you know, voice-to-voice or speech-to-speech a engine, right? So there's no you know speech-to-text, and then text-to-speech, which, which can slow down the experience.

Tom Arbuthnot: Yeah. If people haven't played with some of the latest models, so they, they, they had a jump again a few months ago, and like they're… The, the true voice models are i- incredible now in terms of the like being able to interrupt them and go back and forth and, and less, less latency. They're getting really, really good.

Gidi Adlersberg: Yeah, absolutely. Sounds very natural as well. Just, I guess, a lot closer to a human-like experience, But just far more consistent.

Tom Arbuthnot: And so what is it like if I want to do this as an organisation? What does it look like? I mean, AudioCodes are very partner-led as well. Like, is this… I, I obviously I can do Teams Phone Agent, the first part on my tenant, either myself or with a partner. The AudioCodes bit, do I come to you? Do partners configure it? Do I self-configure it? What does that look like?

Gidi Adlersberg: Yeah. So, so, so again, for us, implementation, and especially when it comes to AI- It's so easy to get wrong, right? So we actually while we do allow the customers to, to control, you know, if anyone wants to self-configure this, they are able to. But especially in these early days of AI and, you know, it's not as trivial to get it to work smoothly we actually, you know like to manage the implementation for, for the customer. So basically, you know typically a process would look like, you know, there will be some interested customer. We will kind of get an understanding of what they're trying to achieve. We will build a POC with them show, show how it works. And then as we move to production, starts the, the, the implementation phase. In a simple project, it can, it can take just a few days. In a more complex project, maybe a few weeks. But yeah, the overall idea is that, you know, we are, we are controlling the, the implementation process, making sure everything actually works.

Tom Arbuthnot: Yeah, so you've got the AudioCodes expertise kind of over the project, making sure it's all aligned and and in the, meeting the customer expectations.

Gidi Adlersberg: Exactly. And just helping that deployment to not be one of those… I remember last summer there was this MIT research that talked about 95% of AI pilots failing. I'm, I, I, I want to believe that a year after we, we are in a slightly better place, but still it's, it's very easy to get wrong, and we like to be there to ensure that, you know, the customer gets an actual working experience.

Tom Arbuthnot: Yeah, and I think a big, a big part of that AI projects failing and the risk in these voice AI projects as well is not a clear scope, a clear understanding, a clear ROI. Like, the, the, there's investment here to get this stuff going, and there's ongoing investment to make it go. So the business has to understand, "Why do I want to do this? What are the upsides? How do we know it's working? How do we measure success?" I think a lot of the early AI projects were just like, "Let's do something with AI and show it to the business." But well- Exactly… Did they, did they want it? Did they see the ROI? So it's I always say with these projects, more than half the work is not the technical making the AI work, it's the scoping the ROI, the, the case, and, and getting the business knowledge into the AI agent. Because if, again, if you don't feed those systems with the proper knowledge, they're not gonna perform.

Gidi Adlersberg: Could not agree more. Could not agree more, and it's just B- just knowing how, also how you're going to measure it. I mean, you mentioned it, but w- I'm seeing too many examples of, okay, you know, there's, there's the obvious matrix but when you think about what's really a resolved case, actually, if you spend a minute to think about it, it's not as trivial to defined, to define what, what's considered a resolved case. And we see many customers struggling with that, and that's where, you know, an experienced vendor like us, but not only just any experienced vendor can help a lot how to actually measure the impact of, of these things. They might sound great, but do they really make an impact business-wise?

Tom Arbuthnot: Yeah. Yeah, I mean, like I, I, I was talking to an organization who were doing it, and they'd, you know, given a vendor the knowledge base, and as soon as they started the, the, the testing, they were like, "Oh, it doesn't know this, it doesn't know that." It's like, well, yeah, it's not, it's not in your knowledge base. There's all this institutional knowledge that the human contact center agents knew that had never got put back into the knowledge base. So actually feeding the right information was the, was the gap.

Gidi Adlersberg: That's a classic one, and another classic one is, I call that kind of the modern disguise of, of AI, where, You know, i- i- in the traditional IVR experience, we all know, it's no secret that, you know, the, the IVR tree p- of press 1, press 2, press 3 is… Was typically designed in a way that makes it pretty difficult to get to an actual human. And with these AI bots, it gets even worse because you're not stuck in a tree, you're stuck with a bot. And we are hearing about too many cases where, you know, callers just scream, beg to get to, to speak to… With, with an actual human being, and the bot just insists on, "No, I, I can help you." So again, in this in the Teams Voice Agent program, big part of it is, are we able to escalate from the bot to a Teams user based on presence, streamline the context of the interaction so there's no reasking of information that can- Yeah be frustrating. So yeah, there, there's a lot of these corners that, you know if, if you're not experienced, it's very easy to, to, to miss a few along the way.

Tom Arbuthnot: Nice. And we're still waiting on Microsoft to give us kind of a public stance on the Teams Phone Agent commercial model. Like, I have every expectation there'll be some kind of consumption there. We've seen with Cowork and with the other things going towards some kind of consumption. What does it look like from an AudioCodes perspective? We don't have to get into the exact pricing, but what's the model for, like, let's say a bank using this, which will be, you know, thrashing it of hundreds of calls a- an hour to a mom and pop, you know, s- small dentist franchise that does a few calls an hour. What's that look like?

Gidi Adlersberg: Right. So I, I'll start by saying that for us Making commercials easy and, more important, predictable is v- is extremely important. We believe that especially in an era like today where technology is just so commoditized, AI really compresses the, the market to the same, you know, same feature table, same… Ev- everyone has everything, basically, and if they don't, they're a few weeks away from releasing what- whatever it is they're missing. A big part of how we can make a difference for a customer is exactly with, with something like that. So helping a customer to know what their bill is going to look like in the end of the month, it sounds simple, but actually goes a really long way when we talk about AI these days. So I mean, what I, what I can say is that when it comes to our model, you'll not see any token-based charge. We, we, we would like to take that cognitive load on us and remove it from the, the shoulders of the customer, and that's just another way for us to give value, right? So, so our offerings in, in AI receptionist, Teams voice agents, are completely capped per session, per agent, per concurrent session, per agent, et cetera. So you know what the bill is going to look like at the end of the month. That's really important for us.

Tom Arbuthnot: Nice. Yeah, I hear a lot of people asking for that. They're very, very unsure how to go back to the ROI and then the use case. Like, what's the… How do I calculate the, the business ROI if I've got, if I've got uncapped liability on this scenario or that scenario?

Gidi Adlersberg: I saw this cool LinkedIn post I think it was about a week ago, that someone compared this to the, the cellular world in, in the '90s, where when it started, it was all minutes, right? We had to… You, you bought bundles of minutes, you bought bundles of- Yep… Number of SM- text messages. And as it got more and more commoditized, today, we're just paying fixed prices. So I think end of the day, that's, that's where the market wants us vendors to get to, and we're just trying to do our best to get there sooner than later, I'll, I'll say.

Tom Arbuthnot: Nice. That's good to know. Yeah, so thanks for kind of catching up. Maybe we'll do another session where we'll actually walk through some, some configuration, 'cause it'd be great to talk through the Teams Manager bit and how it, how it interconnects as well. And also depending on when this podcast goes out, we're looking to put together a kind of a, a friendly Team certified unified contact center face off which we'll have you, you involved, and Gidi. So we're gonna get a few of the top contact centers and just do a, a back-to-back session where we talk about the, the USPs and the capabilities.

Gidi Adlersberg: That's brilliant. I really look forward to it. I think it's a really interesting pack. There's some really talented people in this space, and that should be, that should be good fun for us.

Tom Arbuthnot: Awesome, Gidi. Well, thanks for catching us up. Congrats on being the first to be certified in this new Teams Phone Agent Certification, and we'll talk again soon.

Gidi Adlersberg: And thank you very much, Tom, for having me, and yeah, see you, see you next time