Episode Transcript
[00:00:21] Speaker A: Welcome everybody, to our latest Metro site and podcast. Vidcast. Vidcast episode. I'm Robin Garris. I'm CEO of Metrogy, and I'm going to talk to you a little bit about who we're talking to today and what we're going to be talking about. So first, the company itself. Cresta is an AI platform for contact centers. It supports both human and AI agents.
And its core products are really in the area of AI agent, agent assist and conversation intelligence, which I think is such a hot area right now in such a. Such a good topic for discussion. But all of these products for Cresta rely heavily on generative AI and increasingly agentic AI as we move forward. So the pace of innovation with these AI platforms and applications, as we all know, it's moving so quickly and many IT and CX leaders simply can't keep up. I hear that all the time. Like, the pace of innovation so fast, I can't keep up with this. You know, there's. Every time I think I've got something, something new comes along. So I think if you ask people even today, how they differentiate between, like, conversational AI, generative AI, agentic AI, the answers are really all over the map. So I was talking recently with Devin about some other things and we got to talking about AI and I'm like, devin, I need you to come on a Metro site and join us and provide some of the pretty, you know, some of the basics and beyond Magentix. So with that, Devin, I welcome you to the stage.
[00:01:44] Speaker B: Awesome. Thank you so much for having me. Robin. This is going to be, I think, hopefully a fun discussion. It's a hot topic like you mentioned, and hopefully we can bring some clarity to it. I can't promise that my definition won't be somewhere on that wide map of what you're hearing, but we'll use examples and we'll get into the thick of it.
[00:02:04] Speaker A: But you guys have been doing this for a long time, so let's start with, first of all, just a little bit of background on yourself and then also your position across. I know you've changed a bit, you know, got promoted, all that stuff, because you're doing a great job. So just tell us a little bit about yourself and background and all that.
[00:02:18] Speaker B: No, I appreciate that. Thanks for. Thanks, Robyn. Yeah, I'm happy to. So I'm the VP of Product Marketing at Cresta, and in that capacity I lead about half our marketing organization. So that includes core product marketing, customer and partner marketing, analyst, relations, content and social strategy, as well And I've spent maybe the last seven or so years of my career directly in the Contact center space. I've been focused on helping to build and to launch and to make successful products that are focused on Contact center and customer experience. AI first at talkdesk and now at Cresta. So this is a space that's near to dear near and dear to my heart and space that's evolved a lot even in my time in the space.
[00:02:59] Speaker A: Yeah.
[00:02:59] Speaker B: But I've always been deeply interested in that intersection of business and the evolution of technology. And you know, with the pace that things are changing and evolving in this CX AI space, it's honestly never been a more exciting time to be working on this.
I've really, really enjoyed the work that I've been able to do and the products that we've been able to build over these last few years. Excited to talk more about it with you.
[00:03:24] Speaker A: Yeah. So I guess that's a good launch to the next question, which is obviously Cresta's been in the world of AI and even generative AI much longer than people even realize. I think I know it was longer than I originally realized when we first started talking like Generative a couple years back.
[00:03:37] Speaker B: Right.
[00:03:39] Speaker A: Can you talk a little bit about Cresta's background specific to AI and some of the products related to AI, even some of the core capabilities that a lot of your companies, your customers are using right now?
[00:03:51] Speaker B: Yeah, no, I'm happy to, I promise I'm not taking shots at anybody with this, but one interesting litmus test I like to use, at least for companies that existed before ChatGPT, which is a lot of the big companies in this space is were they building with large language models, were they working with generative AI? Sort of pre hype. And I think the answer is no for a lot of companies. And that's one of the things that is, that's a little bit unique about Crestor. Right. Because we were born right at the advent of this big paradigm shift in technology. We were founded by PhD students out of Stanford's AI lab in 2017.
And in 2018 is when attention is all you need. The, the transformer paper that led to all of this was actually published. And so we had folks that were getting their PhDs in this space really directly exposed to how this technology evolved. Right when it sort of became a reality. And so our technical co founder, Tim Shih was in early research at OpenAI in 2017. Our current CEO, Peng Wu was the founding engineer and really co founder behind the department Google CCAI and the Vertex AI platform within Google. So he was within Google when that paper was originally written and published.
So they were sort of both deeply exposed to the emergence of transformer based models right when it happened and already, already really sort of laser focused on the business applications of IT and ping even more so directly to the contact center space. He had already been working sort of more in traditional ML focused on context centers right when this paradigm shift happened. So they were, they jumped all over this, as you can imagine, having that exposure. And so we've been working with this technology for a long time in actual enterprise settings. Our first customer was Intuit. We deployed a transformer based language model in production powering chat agent assist, add, intuit in 2018.
So really that same year that attention is all you need was published. And I can't guarantee it, Robin, but I would probably place like a pretty hefty bet that that was one of the first ever deployments in an enterprise context center of generative AI.
So a little bit of a fun fact there.
And it's cool because I don't know, Robin, have you seen all this press this last week or so about the MIT report of, you know, 90% of Gen AI pilots failing? What are your thoughts there?
[00:06:15] Speaker A: Well, I will, I will say that I guess it depends on how you're defining failing. You know, and I certainly when in our research we're not seeing a lot of these pilots failing, we're seeing, I guess it depends on also how you're using it. You know, like AI and cx, there's all different ways you can use AI and generative AI in the company. AI and cx, we find companies are having a lot of success with them. In fact, like, I don't know, it seemed like about 53% of companies right now already showing an ROI. So our research shows something different. I mean that's not to say that there's certainly not generative AI projects and pilots that are failing.
[00:06:49] Speaker B: Yeah, I think a lot of the.
[00:06:50] Speaker A: Reason they fail is because too many companies try and boil the ocean. They try and do too much is my opinion.
[00:06:55] Speaker B: No, I agree with you there. It's just been interesting to see this week, especially leading into this conversation. But you know, thinking back to, and we've been seeing hard ROI on these deployments again back to since 2018 in pretty large complex enterprises. That that original Intuit deployment was actually written about in the National Bureau of Economic Research and later covered in Harvard Business Review where they talked about how successful it was. Right. Like it drove a 14% overall productivity increase, a 35% productivity increase for the bottom quartile of performers. So like even in the early days when the models were a lot less powerful than they are today with the right product built around them and sort of the right approach to top adoption and success within an account and I in from the organization to, to operationalize it, Gen AI has been seeing success in ROI and context for probably longer than most people actually realize.
[00:07:51] Speaker A: Yeah, for sure.
[00:07:52] Speaker B: But so I want to make sure I answer your question. You know, I talked a little bit about the history of Cresta. You know we've spent years, you know, really sort of from an R and D perspective focused on this technology.
You know, Even in like 2021, 2022 we, we were awarded 13 patents that were focused on behavioral modeling and outcome based steering of AI models, audio analysis of conversations.
So there's sort of a long list of research driven breakthroughs that we've had all prior to ChatGPT even entering, you know, the, the general psyche. Like people didn't know what GPT. Most people didn't know what GPT was until ChatG hit the scene.
[00:08:32] Speaker A: That's right.
[00:08:33] Speaker B: And hype cycle. Right. So I think that's, that's a story I like to tell about Cresta is that we're really a Gen AI native company because we started right when gen became possible.
[00:08:45] Speaker A: We look at Gen AI now, we, you know, now all of a sudden everything's about agentic. First of all, Gen AI by the way, the fastest adopted technology I've ever seen in my career as an analyst. It went so quickly that companies were actually using Gen AI in some way, shape or form. You know, some in a more rudimentary way, some, some more, more sophisticated. But now everyone's talking about agentics. You know, in February we asked companies, we asked you know, CX leaders basically if they were familiar with the term agentix. About 52% of them were. Now we're up to our latest research. It's not out yet and this is still preliminary but it's around the low 70% is, is actually familiar. We still got 30% of people out there that don't even aren't even familiar with the term agentix.
You know, when you think about AI like any type of AI, it's going to automate a manual function. I mean at a very high level. Right.
Explain to the audience like what really makes a transaction agentic. Like where's that line between okay, this is non agentic, but this is agentic. Maybe if you have an example or something, you know, like, I get that question and it's like, you know, a lot of people are really trying to understand, well, wait, is what I'm doing agentic or not? You know, what should I be doing something different? You know?
[00:09:55] Speaker B: Yeah, I think so. It's a good place to start, right? There's definitely a lot of sort of mislabeling out in the market as well. Just given that these hype cycles move quickly and agentic is the hot thing right now. So everybody wants to call themselves agentic. Right?
[00:10:09] Speaker A: Right.
[00:10:10] Speaker B: And in some cases they are, in some cases they're not. In some cases it may not make sense to be right. And we can talk a little bit more about that and, and sort of some of those distinctions later in the conversation, I'm sure. But just starting with agentic and the way I would define it, just to try and make it a little simpler to, to, to understand, I think like I define it as AI that isn't sitting around waiting for you to ask it something. Right. It can actually take initiative. It's not reacting to prompts. It's, it's AI based systems that can set goals, that can make plans on how best to accomplish those goals and are actually given agency, you know, hence the name, to carry out tasks in service of those goals on their own or with very minimal human guidance. So you know, if you think about a calculator or even just chatgpt in its basic form, right, you're, you're asking it for an answer and it's giving you the answer. AI is a much more sort of capable and context aware assistant that understands more broadly what you're trying to accomplish and it can notice what needs to be done, it can figure out the steps to use what tools it needs to interact with along the way and then actually take those steps, use those tools. It can adapt its path based on what's working or not working.
[00:11:21] Speaker A: So let's say I'm trying to plan a vacation, right? And I go into a generative like a ChatGPT or something like that, and I ask for some, it's still going to know, okay, I need to take these steps to figure out which hotels are going to match her criteria that she's given me.
But that's not necessarily agentic. How does that example then become agenda?
[00:11:39] Speaker B: Yeah, well, you're prompting it, right? You're, you're basically asking it for help and saying, here's the situation, let me describe it to you.
And then it's going and it's providing you with information and guidance. Now I think like with, with the more common use cases of ChatGPT that are getting closer to maybe be feeling agentic is like if you use deep research or something like that where it's going out and it's interacting with the web or if you're giving it permission to use tools. But in those cases those are still sort of very tightly defined, like user defined actions that you're allowing it to take or telling it to take. Where you know, an agentix system can sort of understand broader context and goals and determine what to do, actually proactively decide what to do, build a plan, take action, use tools to solve the problem in service of the goal.
[00:12:27] Speaker A: Yeah, so basically I don't know, I've gone, I went on a vacation and I really like this particular, I'm just doing this silly example, this particular hotel brand now Gentex. Now AI May, my AI assistant may know that and see on my calendar I've got a vacation block. Well, I'm just going to proactively book this, this hotel brand that she likes.
[00:12:47] Speaker B: Because maybe I don't think that would be a great system necessarily that you'd want it to do that for you.
[00:12:52] Speaker A: You know, because it's a good rate. Oh look, Robin's busy, she hasn't planned her vacation yet even though she's got the time marked off and they have a great rate right now. So I'm going to book it and alert her and know that there is a cancellation that can a week. So just, you know, I'm saying like.
[00:13:07] Speaker B: That would be great.
[00:13:08] Speaker A: I would love that.
[00:13:09] Speaker B: Well may, maybe you would. I think like where, where I anchor obviously is in sort of the, the customer service and sales domain where like you do need many more guardrails and sort of frameworks around what it is and isn't allowed to do. But like maybe let's anchor an example there.
You know, like something as simple as a customer calling in, wanting to know where their package is. Like in sort of a non agentic system. A system that still could be using generative AI though, right. The customer could say where's my package? The AI, you know, looks up the tracking info, it actually generates the dialogue back to explain the current state of the order to the customer.
[00:13:48] Speaker A: Right.
[00:13:48] Speaker B: So that would be non agentic but using generative AI in a hypothetical agentic scenario, the customer's asking where is my package? The agent is actually conditioned on solving the fundamental problem for the customer or driving CSAT or resolution rather than just Answering that exact query. Right. And so it checks the tracking info, it notices that this package is delayed beyond the sla. Like let's say it's prime, you know, the two day delivery window or something like that. It notices that. Right. And then based on its goal and history of this customer, let's say it can make a decision to offer compensation or you know, expedite the shipping for free on the next order.
It can log the issue, it can update the CRM or it can determine maybe that's an edge case where we need approval for something from a human. Right. So it's much more adaptive and can be built around sort of more second order goals related to the customer or the scenario at hand and decide how to use context in service of accomplishing those goals rather than just, you know, the simple input, where's the order?
Let's use AI to go look it up, let's generate a response, it's all done.
[00:15:08] Speaker A: That's may not even need that prompt of me calling and saying where's my order? Or typing and saying where's my order? Could proactively see, oh wait, there's a, there's a problem here. So I'm going to reach out and.
[00:15:19] Speaker B: Say you could certainly design a system. And that's, I mean I think that's where some of the really exciting, the next evolution of where this technology is going is in proactive service, in hyper personalization and those types of certainly where we see a ton of untapped potential when it comes to agentic AI use cases.
[00:15:39] Speaker A: Yeah, for sure. And you start thinking of some of the vertical industries where it could be so profound, you know, like healthcare for example. I know obviously a lot of guardrails there and you know, you have to definitely rely on.
[00:15:51] Speaker B: There's give and take. There's give and take for sure. You know, anytime you're talking about a probabilistic system that is then given agency to do things and access data, you know, I think you definitely want to take advantage of those tools in situations where it can add a lot of value and be dynamic and sort of redefine what service can be. But at the same time there are limitations in terms of, you know, reliability and reproducibility and things like that. So the higher the stakes you need to be absolutely sure that you've designed something well and that you have really sort of robust oversight and testing and simulation systems and guardrails in place as well as deterministic flow fallbacks for Agentix systems in certain scenarios where they lose their Agency and it just does what you define it should. In that context, what would you say.
[00:16:40] Speaker A: Like based on your customers who are doing things now, like, what would you say are the biggest use cases? Like, what's the business value that it's really to companies right now?
[00:16:51] Speaker B: I think there's a couple key categories. I think the most obvious and common answer in the contact center domain and the customer experience domain is conversational AI agents. So they're actually having conversations with customers and they're taking action across other systems to solve more complex problems. And again, like you, you know, traditionally, I think we, we've all thought about the key metrics. There is like deflection and containment. Like are, are we able to just like get the thing done without having to escalate to a human? But now with conversation intelligence and predictive outcomes and these types of things, you can start to sort of create like new dimensions of goals and outcomes where the goal may not just be to like simply contain at all costs. It might just be the goal might be as simple as resolving it successfully with a high degree of customer satisfaction. And we now sort of have the data and inputs to use goals like that and deploy agents that one can resolve a wider range of issue without having to involve a human, but also ensure that like the customer is getting a great experience out of that.
[00:17:56] Speaker A: Yeah.
[00:17:57] Speaker B: And so there's numerous benefits. I think the really obvious one and the one that gets talked about the most is related to cost reduction. Right. If you're removing a human from the loop in a higher volume of conversations. Right. You're saving money. But I think that there's numerous benefits to the end customer as well. Right. Like we can now potentially offer 24, 7 service in scenarios that we didn't before. We can eliminate wait times entirely for many issues. We can reduce wait time for the remainder of the things that aren't being automated. Right. Because we're freeing up human agents. And I think the other interesting thing that I like to talk about when it comes to conversational agents is they can actually outperform human agents in certain contexts. Right. Like they don't get tired, they don't get tired, they don't get frustrated, they have endless patience.
Fun story I love to share when I talk about this is we have a customer who uses our voice AI agent for tech troubleshooting. And some of their customers, a significant portion of their customers sort of are in the older demographic. Right. And so this might sound counterintuitive. Like you think older customers, they're less receptive to AI. But their elderly customers love interacting with this voice AI agent because it just, it guides them step by step and it never loses its cool. Like no matter how many times it has to repeat itself or explain a simple concept, it just goes, it treats them exactly in the way that they need to be treated to get through the troubleshooting workflow and leave having solved the problem or giving them a tangible next step. And so it's pretty amazing to hear the reaction. I've actually listened to some of these calls.
They start treating it like a human within a few minutes, like calling it by name and they're genuinely thanking it and praising it at the end of the calls. It's really quite endearing to hear, especially coming from people who you might assume are less receptive to AI overall. And like you said, the data shows can be less receptive. So that's the obvious use case, but I think there's many others. Right?
There's a lot of ways that we can augment human agents using agentic AI. And I think that the most obvious and sort of catch all category of that is workflow automation, right? Where the human agent can handle the conversation itself. But an agentic system or an agentic assistant, you know, whatever you want to call it can execute more complex workflows and processes for them and it can use context from the conversation to interact with other business systems. You know, in the case of some of our customers, the systems that their agents have to interact with are, are, you know, they don't have open APIs, for instance, they can't be integrated fully. And so now where we're going next, this is not current day product for us, but where we're going next is computer use agents that can actually do these workflows locally for the human agent, even using systems that don't have open APIs. So it sort of expands the scope of what can be automated in terms of workflows and systems that are often closed down. And I think, you know, since you said business cases and value behind all this, like it's similar to the conversational automation side where there's a cost based business case in the right situations, right? It can reduce ht, it can cut down on hold time. But again, I like to think beyond those costs, right? Like both of those benefits, they, they both impact the end customer experience as well, right? You're saving time for them. Nobody likes waiting on hold, right. It can help new agents ramp up faster and navigate complex processes with fewer mistakes. And I think it might sound a little bit Cliche, but like, it frees up the human to do what we do best, which is connect with one another and listen and show empathy. In our scenario before, where the AI agent was identifying opportunities to go above and beyond. Like humans can be really good at that too if they're not like heads down trying to navigate some complex old billing system while also, you know, listening to the customer and, and what they're saying. So that's one that I love. Both of those are sort of real time use cases. But I'll just do one more plug here and then I promise I'll move on. Robin, you mentioned that you are really interested in conversation intelligence. I think that's sort of a, another frontier for agentic AI, right. I think it can play a key role in making conversation intelligence much more impactful and closing the loop between insights and action. I think in a lot of cases, you know, the prior generation of conversation analytics and conversation intelligence, there was a lot of insight there, but insight is only as valuable as what you do with it. Right?
And so I think there's a lot of things where using agentic AI connected to Insights can actually close that loop. And so, you know, today our AI analyst product, it interacts with the end user. It helps clarify their questions, right. They are able to ask natural language questions about their conversations and helps them zero in on relevant conversations so that they can get sort of a more accurate and representative analysis. But it also provides, you know, clear reasoning and evidence for its response. So there's like elements of agency in the current product, but in the future you won't even have to ask questions, Robin, like it will just understand you as an Insights user. It will understand the context of your business. It's going to proactively run analysis behind the scenes, make suggestions and then even take some of those next step actions for you. So one of the things we talk about is like the concept of a self healing knowledge base where it's identifying gaps in the underlying knowledge and pulling things from conversations and validating those and then helping develop knowledge. Right. Or it's generating a prototype for a conversational agent. The first use case that we talked about based on a conversational flow that it identified has high automation potential. So these are areas where, you know, an agentix system can sort of change the way that insights works and close that loop from insights to action.
[00:23:47] Speaker A: All right, well that is fantastic. It looks like we are out of time and I could probably talk to you about this for another hour, but I really appreciate you taking out the time and just giving some of your insights in this whole area and just what promise it's showing for companies. So, again, thank you, Devin, for joining us and for giving our audience some insights from your massive experience in this space.
[00:24:10] Speaker B: Oh, happy to be here. Thanks so much, Robin. And if you want to keep chatting about this, I'm always available. It's always a privilege to have a conversation with you and all the great data that you bring to the table. So thank you.
[00:24:21] Speaker A: Well, thank you, Devin. Thank you.
