NiCE’s Russell on AI Agents for Contact Centers

(upbeat music) - Hello, and welcome to the Bloomberg Tech Disruptors podcast. My name is Mandip, and with me today is Scott Russell, CEO of Nice Software. Scott, welcome to the podcast. - Great to be with you, Mandip. - Great, so look, I'm gonna jump right into the questions because 2025, as we know, was a great year for coding agent as a use case, and there was a lot talked about in terms of customer service. So you guys play into that customer service use case, and you've obviously been around for a very long time. Tell us how you are transforming the business, and what is it that the LLM companies are doing for you right now. - I think the reality is customer service is a wonderful use case of AI that can be delivered at scale. And what we've seen through 25, and even before that, but through 25 is companies that are using AI, particularly in augmentation, so assisting their agents, next best action, coming up with real-time transcripts to be able to guide and shape an engagement with a consumer in a more proactive way. That's now gone full scale into automation, to be able to orchestrate through AI agents, to be able to deliver activities in an orchestrated way. So really what we're now seeing in customer service is the ability to be able to give outstanding customer service with a combination of an AI platform and with the LLM's now having a even more capable platform with which we use combined with our data, the guard rails and the understanding and the 40 years of history in customer service that we have, you combine that together, you get the best of both worlds, human engagement where it is needed, but AI really paying a forefront role in the way customer service is delivered. So it's exciting to be right at the center of that transition. - And so when I think about contact centers, it's obviously you are handling the calls, then there is quality management, and there is a big legacy process that's already established where you've got SLAs. Like how are you handling that change and what are the tasks that you feel within the contact center where you feel very confident about LLM's kind of driving that change and really doing a better job than what was being done before? - Yeah, I think if you look at the way that the AI and in particularly generative AI has been used in the contact center in 24 and early 25, it was very much around how do we help the agents who are often resource constrained as you rightly point out have they got the right training, have they got the right quality of high turnover of people, can they handle peak volumes and that are not able to be resourced effectively. So businesses were really struggling to be able to deliver high quality customer service in a really complex contact center. So the first move was augmenting or supporting and engaging the agents to be more productive, to be more efficient, to be able to give a better quality answer, nest best action, prescribed outcomes, preempting those outcomes, but very quickly over the last 18 months and I would argue in the last nine months we've seen an incredible acceleration of true automation where AI agents with nice cognitive, for example, were able to fulfill a significant amount of engagement with a consumer, whether it be voice or chat or whatever the medium that they want, and be able to give a really comprehensive and real time response to their needs. But then secondly, with agente AI, being able to perform tasks that otherwise would require humans in the mid and the back office to be able to give an approval to or, or that would take time and a call and it can be happening instantaneously. So what we're now seeing is two to 3% of the volume of interactions coming into the contact center was handled by AI. We expect over the coming years, that can be up to 30, 50, even up to 80% of the volume can be handled by AI either exclusively or in combination with human agents. And that's the exciting transition that companies are going through and what AI capabilities and the LLMs give to us, obviously combined with the expertise and the our ability to take unstructured interaction data and then putting it into a structured form that businesses use to give high quality answers to their consumer needs. - Wow, that's a big step of the 10X increase that you have a timeline in terms of what are we talking about, that change 30 to 50%. - Yeah, look, we've modeled it. I mean, we looked at the billions of interactions that we have on our platform, the tens of billions of interactions we have on our platform and we've modeled what are the interactions that can be delivered either exclusively or predominantly with AI. So their own data suggests that we could move up to that 80%. Now, the reality is businesses then need to manage the transformation because it isn't as simple as just simply moving from a human to an AI, quality guard rails, use of data, accuracy of information, often the willingness of businesses to be able to move to an AI only versus, 'cause they like the humans to be in the loop. They give them upsell and cross sell opportunities, revenue generation opportunities. So there are complexities that businesses think about, but if you just think about the raw needs of that interaction and the needs of consumers, we believe the large majority can deliver by our AI platform. The beauty is the combination of human and AI together gives us the most comprehensive stack to be able to handle our customers needs. - And you mentioned different modalities that are being used when it comes to contact center. What is it that you are using out of the box from LLMs versus where you had to spend time with customizing or fine tuning the performance and then deploying it in your production setting? - Yeah, there's no doubt that the LLMs, as, and we give our customers choice on the LLMs. So we work with all of the major LLMs and have great partnerships in place, but many times, especially at the enterprise segment, our customers have preferences for certain LLMs that they've got according to their corporate guard rails. But when you think about what we deliver, there is the combination of deterministic. So companies are very clear. These are the answers or these are the questions that according to their guard rails. Maybe it's because it's a highly regulated client where they're looking to be able to provide certain answers under certain scenarios. And that is obviously very well trained by our models. Then on the probabilistic, where they're using the data to inform and guide and have an expiratory, those models that the LLMs are providing, those foundational models that, look, I would argue they're becoming a little bit more commoditized as a utility. So our domain expertise, combining with their utility model, it means that we are able to use our expertise to be able to do more complex use case, rather than the simple out of our platform combined with what the general purpose LLMs are providing. So their advancements helps us to become more specialized, more capable to do the full spectrum of customer service, rather than can I call it the easy, low-hanging scenarios that can be easily done by any LLM and any service provider. - And I mean, there are so many different techniques right now with these LLMs reinforcement learning is something that gets mentioned quite a bit. But at the end of the day, you have to find a way to connect with the internal knowledge of an enterprise that sits probably in your system of record or things that LLMs don't have access to. So how are you bridging that gap? And more importantly, how long does it take to fine tune the performance in a way where you are 100% sure that the LLMs won't hallucinate? - Yeah, there's a few things to unpack there. The first is as the platform that has got the data repository and the history and the knowledge as that system of record, or what I would argue, the system of engagement with the first platform of engagement, voice chat, digital, AI comes into our platform. A lot of that data that companies have that you've referred to, the knowledge platform, their share points, their repository of insights that goes into the learning that previously went into training of humans, now needs to be built and deployed for the AI usage in an accurate way. Much of the data initially is not very useful. Maybe 10, 20% is, so you've got to be able to take unstructured interactions, contacts, metadata and putting it into a form that is usable that drives, you know, that 8, 95, 99% that drives the better outcomes. So making that data useful is an art. And it's the art that we've perfected over many years in enabling humans in the context. And now we're doing it in enabling AI to be that high-performing non-hallucinating platform, but you really need to make sure you fine tune that unstructured data to be able to deliver on the outcome because the one thing I know about customer service, Mendy, no company is prepared to have an inferior customer service engagement than what they're currently doing today. So the guard rails, the expectations, the quality standards where the customer engagement is incredibly high, rightly so. And it means that that accuracy and that use of that data becomes a pretty important aspect. You'd be interested to know that our AI deployments can be very quick. The long pole in the tent is the cleansing, the usefulness, the structuring of that data, so it can be delivered in a contact center environment, whether it be AI or human. Is that the reason why we haven't seen big success stories in customer service the same way we did in coding agents last year because the bar was high. Whereas in the case of coding agents, there wasn't any existing market to benchmark again. So whatever was produced was something incremental. Is that be a good way to frame it? I think it is a good way. I mean, there's clearly a benchmark of reference. And the thing that I always try to explain is the volume of interactions so consumers continue to increase their interactions with brands. You know, ballpark between 5% to 10% a year. So you've got a higher volume coming in. You've got a high expectation and a standard that is already at a benchmark that companies won't reduce. So the AI platform needs to improve upon that. The good news is this. I think in 25 second half we saw the pivot point. We saw the market now customers realize that they can deliver its superior customer service leveraging the AI platform such as nice cognitive. And so I do think in 26 you will see the real expansion, the wide-scale use cases, the success stories, not in a complete replacement of humans in the context. And because remember, they've still a role that they played is a lot of complexity in enterprise platforms that is hard to unlock. But also they want consumers to be interacting with people for other purposes other than just pure customer service. But I think you'll see 26 being a real breakout year. There is no doubt that we're not only going to see growth that we're forecasting, but real customer use case stories that are remarkable and it's fun to be a part of. And when everyone talks about agent take AI, there is that inbuilt expectation that these AI agents can work for a lot longer than do just one task. So is that something that you expect in terms of the length of tasks that the agents would perform and if that's the case, what is it that you will have to do in terms of, again, that model layer, which, as we all know, there are four or five frontier models now that everyone is using. So how does that change your dependence on that LLL layer if the length of the task is increasing? Yeah, I think there's two angles to it. One is the length of the task. The second is the complexity of the engagement. So again, in the customer service world, consumers aren't as predictable as what we think they are. They might come in and have one intent, but then they will quickly, if they get a faster resolution, there might be two or three or four or five. And so you're in a conversation that requires a continuity what we already see. And I'll give you a good example of this is, with nice cognitive, we have a customer, Lufthansa in Europe, one of the biggest airlines. When there is a weather outage or a major impact, let's call it at the Frankfurt Airport, that impacts a lot of consumers in a very short period of time that can be in you think about rescheduling, rerouting, reorganizing the flights that is sound simple, but actually has a lot of system activity. And they've used the agentic platform of cognitive to be able to perform those proactively reach out, handling tens of thousands of concurrent calls at the same time on those critical events, something that humans in a contact, you just can't scale to that. And what we then experience is long wait times as we're trying to get our urgent task solved. So you can see the convergence of the agentic ability to be able to perform complex tasks that otherwise would have been time consuming, being able to handle real time in that customer service scenario. Last but the thing that I will say about the agentic side and what excites me the most is in customer service, you are often limited to the expertise of the human at the other end of the line. And a lot of the tasks that needed to be done required tasks or approvals, you think about a credit card transaction reversal or a billing charge that you wanted to remove or a scenario's like that. That often required approval or task beyond the person in the context and with the agentic, we're able to perform those tasks in real time using on the enterprise platform without leaving the customer engagement. So your customer engagement platform becomes a whole lot more powerful than what it ever has before, which means you've got a more efficient organization in delivering customer service as well. - So in that example, were there any any decisions that were made by AI in terms of processing credit card transactions? - Not certainly not on our platform, man deep, but I think you can obviously see why companies go into customer service with AI with very clear expectations around guard roles. I mean, some of our competitors have had incidents around, you know, security and others. So organizations are very well accustomed, especially when you're dealing with financial data or private or you know, personal sensitive data that the automation is fantastic, but it needs to be combined with the knowledge base, the guard roles the companies have around the way that they're going to deliver that, which means they use a combination of deterministic out time, outcomes and probabilistic with generative AI and they're very thoughtful in the way they deploy that. It's why platforms like ours become really critical 'cause they can do both and even make the decisions about when to bring humans into the mix or not so that they're delivering a great customer service and they're not having poor hand or for experiences. - And I want to quickly go back to that point about the capabilities of the frontier models. Like, do you think these models are going more towards specialization and what I mean by that is, are they like specific models better for let's say, voice type of interaction than for image generation or for tech space chatbot functionality? - I think it will inevitably work that way, Mandy, you are right. You think about the specialization of this model that recently had a trip to Japan and we've got a great business there and you think about just voice digital in the language context but also in the societal context and how you're delivering that on a behavioral basis and a context basis. So I think these frontier models being more specialized and in our relationship with those LLMs, we really do partner very successfully in leveraging their advancements on the models that we can embed naturally on our platform, bring our insights of the structured and unstructured data and the CX models and we've got thousands of small models, CX models that we apply over the top to be able to then deliver great customer experience. But there's no doubt that they're bringing a capability that makes it easier for us to bring that to light, whether that, and especially in our world, voice chat, tech space platforms, the ability to be able to do that real time is quite remarkable in the last 12 or 18 months. - And do you have a view on proprietary versus open source models? I mean, it feels like open source had some momentum then things were quiet and suddenly it started to hear open source a lot more lately, so. - Yeah, it's interesting, isn't it? It ebbs and flows, I don't think that. I think you will see proprietary models that have got really smart, embedded IP that will create a differentiator that comes to the fore and then I think the open source will bring scale and opportunity to what. So I don't believe it'll be one of the other. I think it'll even flow and we've got to be adaptable to leverage both. - And from your perspective, I'm guessing you will be keen to use open source more or would that be more work in some cases? - No, no, of course, I think we obviously see the benefits. I mean, our business is geared towards delivering great customer experience and engagements. If we're able to leverage open source models combined with our expertise, our data, our insights, we're able to do that real time more effectively than clearly the benefit ultimately is the consumer and the enterprises we serve. So it's not an exclusivity statement, but it is beneficial to our company for sure. - And you mentioned Lufthansa as a customer that has deployed the agent AI. I mean, you guys have customers across different industries. So what are the other industries that seems to be early adopters of this agentic functionality? - Yeah, so industries that are in high volumes of customer engagement and service, so Disney is a good example. Hundreds of millions of subscriber with their Disney streaming and Disney plus and services. So being able to perform both helping their agents, which they call cast members to be able to serve their consumers fast, more agile, more ably, but then perform to us. So whether it be bookings, upgrades, renewals or additions to their trends and being able to do that through a pure AI platform versus involving. So you can see consumer product companies, customer service-orientated industries clearly being participants, whether it be in financial services or associated. So they'd be the industries that I'd call out. What I think is coming very quickly is retailers. And the reason why I like that is retailers have got a significant presence of customer engagement that is not bound by the bricks and mortar. They have a lot of interaction and a lot of ability to be able to not only serve, but upsell and crosssell. So customers of ours like Walmart and others are clearly making significant inroads in using agentic AI and leveraging platforms like ours to do it. - Yeah, and have there been any reasons for why things have slowed down even though we have technology that keeps improving by the week. But things slow down from your standpoint because customer wasn't ready with data or APIs or something else. - Look, I think there's a couple of things that's probably worth highlighting in the world of customer service. I think a couple of years ago, maybe companies jumped a little bit too quickly into thinking that they could replace all of the agents and just do it purely with AI. And I think they learned very quickly that delivering great customer service isn't just about automation. It's about fulfilling the needs of their consumers in a comprehensive way. That's not to say that the AI can't play a more comprehensive, you know, a significant role as I mentioned earlier. But it needs to be in light of giving high quality response, high fulfillment rates, high containment rates, reducing handling times and doing it in a measured way. So I think those early, let me call difficulties, maybe a few years ago, maybe paused or slowed a little bit about the potential benefits. But we've seen that change. Honestly, Monday been in through 2025, the acceleration of the AI volume of interactions was extrapolating triple digit every month. It really we saw their significant ramp up because I think once companies saw the one use case, then they turned to you to two, to five, to 10, to 15. They're able to then build it out on the same platform and doing it quite easily given the deployment capabilities. - Just to play a devil's advocate on that. Like, why do you think the MIT study then said that 95% of these don't work out? - I think again, maybe a little bit too ambitious on the breadth rather than being targeted and specific, prove it out and then build upon it. So certainly for the nice customers, we see a lot of our customers build out both automation scenarios where they're doing pure self-service, augmentation scenarios where they're assisting their human agents on next best action and orchestration scenarios where they're using agentic to perform tasks through the enterprise. Very purposeful, very clear about their scope, deploying those, deploying successfully. And then it's about scale. And I think they're looking for that at proof points. And there's no doubt. Enterprise is not interested in science experiments. They're interested in being able to give real return and benefits and we certainly stand up to that. So I think also the industry has become much more adept to being able to deliver really significant proofing use cases that are repeatable, that they've delivered, we've delivered for others and give it into their environment as well. I think you'll find in the world of customer service, 2026 will be a significant year of expansion and growth where we will be able to prove that you can deliver AI in customer service at scale. It will be able to deliver real return on investment for the enterprise, but most importantly, consumers will be happy. They'll have a better experience, real-time response, better service delivery and so they're going to be improving their relationship or their NPS with their brand. And last thing since we keep hearing about the memory shortages and shortages for all kinds of supply components, do you guys see any shortage or supply constraints when it comes to running your business? - Not really, not for us, I mean, I guess there's always, we hear and talk about GPU availability and we supply constrained. Our partnership with the AI providers let us innovate without many of those constraints. So I guess we don't see that as a limiting factor. I think as we go further into the agentic capabilities and we go deeper into complex resolution and customer service and how that gets built out into an enterprise and broader data sets and broader systems of record. I think we might start hitting those, but we've got a lot of runway before that becomes a constraint in customer service, not for the foreseeable future. - Great, Scott, it's been great having you on the podcast and I know if we had to keep it tight today but come back soon and I wanna wish you the very best for 2026. - Thanks for your time and thanks for you to the listeners as well. Thanks, Mandy.