The Agentic Communication Stack

Julia Nimchinski:
Thanks, and up next, we have Calvin, the Lead Product Manager for Agentech AI at Dialpad. Great to have you here, Calvin.

Calvin Hohener:
Yes, great to be here, Yulia, thanks for having me.

Julia Nimchinski:
What’s in your egentico agentic stack?

Calvin Hohener:
Yeah, so I’m going to talk about, AI-powered customer experience, and pull up a deck right here, and then I’m going to give you guys a demo of Dialpad’s Agentic platform, which we rolled out at Enterprise Connect a couple weeks ago. So let me share my screen, and we can get this going. And it’s interesting that I’m following someone from Gainsight, because you’ll see a lot of, sort of, overlaps with this idea of your customer interactions being the nexus of data, right? the most reliable signal that you really have to use in other workflows. And so, you’ll see a lot of that sort of mirrored, and I can maybe skip through a bit of my introductory slides. But yeah, customer data is fragment… fragmented, and I think the big here… big thing here about it being underutilized Is the fact that it relies so much on user input. It, you know, your CRM requires sales folks to consistently update it for it to be reliable. And that’s something that you constantly hear in organizations is, oh, we can’t rely on our CRM data because, you know, a human needs to update it. Oh, we can’t rely on our ticketing system because we, you know, don’t think our CX folks are going to be able to update all of the fields properly, yada yada yada, et cetera, et cetera. With AI, you don’t need that organizational layer, right? You don’t need the data to be structured to be valuable to you. You can get to what those underlying fields should be populated with directly from the conversations you’re having with your customers, or even internal conversations as well. And so, that’s what we’re saying, you know, the most honest, high-signal customer data that you have is from those interactions. It’s from voice meetings, it’s from your messaging, it’s from your meetings, it’s from your digital and social channels, it’s from your knowledge base, it’s all of the sort of, you know, interactions that you’re having with your customers on a day-to-day basis. And with Dialpad, because we are this end-to-end communications and customer support platform, we’re capturing all of those interactions, and so you have that rich signal in one data lake that you can utilize. In various ways. And so, with Dialpad, we have this understand layer that’s built on top of it, that includes our agentic system, it includes the small language models that we’ve been building since 2018, Dialpad GPT, and it involves our rich NLP and ASR capabilities. And from that. What you get is real-time assistance for your human agents, you get self-improving agentic systems, you get These agentic systems that can actually go out and take action on behalf of callers and solve problems and write to systems, which is new in this agentic game. And so, we kind of consider what’s happening in customer experience, much more holistic than folks may have thought of it before. I think traditionally, folks think of customer support organizations as cost centers. And we are seeing that that’s simply not the case because of the rich data that they provide. So, we built a sort of constellation of AI features, all focused on how interact… how human agents and AI interaction… and AI agents can interact with one another and positively reinforce one another. So, our AI helps human agents in real time with features like AI agent assists and playbooks. Furthermore, the human agent is actually going back and informing all the interactions that the agentic system is having, and so it creates this sort of virtual cycle of the AI benefiting the humans, and the systems benefiting from any feedback that’s given from the humans, i.e. agents taking action, or just what’s happening in the conversation with your humans… with, with your human agents. And so… This sort of slide, I think, tells the story of how Dialpad’s been focused on using AI since we started back in 2018. That’s not when the company started, but that’s when we purchased a company called TalkIQ, which gave us the start of these rich AI conversational analytics products. At the start, it was about providing visibility, right? Here’s the… here’s what’s happening in your transcriptions, here are the analytics on top of the conversations that are happening inside your business. Then, it moved to the assistant stage, right? We have real-time assist coaching, you can get live sentiment based on what’s happening in conversations, AI-generated CSAT scores for every customer interaction, scorecards, etc. And now we’re moving into the agency phase with Agentic, wherein these agents you’re putting out into the world can actually act on behalf of your customers. That’s gonna be the focus of the demo, as well as really this agentic piece that we’re moving into. And so, when you think about building AI agents, there’s sort of four main steps, right? The first is the discovery phase, right? Which use cases am I going to focus on? What is high enough, ROI for me? What is actually taking a lot of time by my human agents currently? There’s the build stage. How do I confidently build an agentic system, right? This is a skill set that not a lot of people have, so there has to be a lot of assistance from the system that you’re using to ensure that that agent is going to be effective in the real world with your customers. There’s a measurement phase, and this happens even before you put the agent live, right? How are we going to prove ARI? How are we going to put this agentic system through its paces before actually putting it live in production with customers? And then, post-production, there’s the governance and optimize phase, right? How do I continually update this agent to give it new skills, give it new capabilities, and ensure that it’s actually serving my customers? So what I’m gonna jump into now is, a demo of what we rolled out at Enterprise Connect, which is the end-to-end build experience to kind of ensure that throughout those different phases of putting an agent out there in front of your customers, you’re building with confidence, and you’re ensuring that you’re actually tackling the right use cases. And so, the feature that I’ll start out with is, one called skill mining. Now, thinking back to your customer support conversations sort of being that nexus for that rich customer data, that’s exactly what this feature is built for. So, if you’re unsure of what to start automating, the best place to find out is actually in your real conversations. So. What this does is the AI will go back and look at a corpus of transcription data from, you know, I’ll show you right here, you know, particular call centers, that you want to look at, a date range, we’re gonna expand this, it doesn’t necessarily need to be just a week or 30 days. But when I click mine for skills, what’s gonna happen is the AI’s gonna look at this transcript data, and it’s gonna suss out use cases that, A, are frequent and actually taking a significant amount of my human agent’s time, right? So, these are things that are ripe for automation. The flip side of that coin is, is that actual process something that can be readily automated using AI, right? And we can tell that mostly by how successful human agents are executing that task, right? Is it something that’s repetitive, mundane, that they’re doing very, very frequently? That’s something that’s ripe for automation. And so, mining for skills will give me that feedback. While that mining’s happened, you can kind of see some overview, statistics that we provide for particular agents on this landing page here. But what’s particularly relevant here is these are the skills that were sussed out of the actual conversations with customers, right? So, 27% of conversations are involving this mundane task of generating a returns label, right? And so you may have had some sort of sense of that, hey, my human agents are spending a lot of time doing this mundane task, but this actually bears it out, with AI. Now, what’s great about this is it lays out some steps of what needs to happen in order to properly generate a returns label. But if I click Create Skill, what this is gonna do is it’s gonna take me into what we call the agent studio, which is where you’re actually constructing the agent, giving it skills, and building out its capabilities. And the AI does all that for you. So if I click Create Skill. it takes me to our canvas, and you can see that it’s created this agent with this return label scale, right? You can actually see that the steps that go into that are fairly complex for such a simple operation, right? You need to verify the customer. get the order details, validate the address, check returns policy, etc, etc, right? So, what our AI does is it looks at a simple process like generating a returns label, and ensures that the end-to-end, you know, completeness of that process, as far as a standard operating procedure would be concerned. is actually fulfilled. And so… When I click Create, and update, that will… generate this skill. Now one thing that you’ll notice is that we have this little spinner running, and what that is, is a feature called Compass. And so what Compass does is it looks at every edit that you make to your agent during the build phase, and it’ll essentially audit those edits, right? It’ll ensure that what you’ve just entered is logically complete, right? If you are creating a scheduling agent, you actually have it hooked up to your calendar system, things like that. It’ll suggest things like guardrails to ensure that, that conversation actually, you know, can be completed end-to-end. It will suggest things like updates to the actual language that you’ve learned, that you’ve used in your prompt if you’re entering it yourself, rather than using the AI-generated, prompts. So… If I click up here on the one skill, I can see what those suggestions are, right? I can view details. I can ignore or accept individual skills, or individual suggestions, but I also have this assistant that kind of lives with me along the side here as I’m building my agent, and I can actually, like, just ask it to explain itself, right? Click this explain button, I could have said, here, explain the updates that you’ve just given me, and it’s gonna go through one by one, and go through the recommendations, talk about the problem, and then how it could be fixed, right? And this translates to these suggestions that you see over here. I don’t have to update then manually. Again, I can just click Improve Skill. And you’ll see that it’s taking the prior prompt. And giving me a nice diff between what it had before and the new suggestions, which include, some additional guardrails. Again, I can go in and accept or reject individual suggestions that it’s making, or I can just click update, and in one step, it will take those suggestions and actually update that agent. You can see that it’s added this, this workflow here, so this is a deterministic workflow to get a validated address when you’re generating a returns label, an important thing. But Compass knew to add that guardrail, and that was one of the suggestions that it gives us… that it gives us. Just real quick, I’ll show you the assistant, like, I can just… I can ask it anything, like, I need a skill to find the best cheeseburger… In the location specified. And it’ll… create a cheeseburger-finding skill for me, right? Which is kind of silly, but, you can just see how powerful the AI is, that, you know, this is a very logical way to go about this, right? It’s collecting the location input from the caller. hitting a Google Places API, comparing the ranked results, etc, and then providing the details. So, this is a way that I can just sort of speak skills into being for my agent, and keep them, you know, rooted in the best practices that Dialpad knows, and on the rails. So I’ve created this agent, and now I want to test it before I put it live. And this is where we built a feature called Proving Ground. And so, from Proven Ground, basically what this does is it’s a scenario generator, so the AI will look at this agent that you’ve just created, all of its configurations, all of the skills that you’ve given it. Again, those skills can be mined from your historical data, so this is very much so pulling through what’s actually happening with your real customer conversations all the way through to this, you know, scenario generation suite. And it’ll give me scenarios to test. And there’ll be happy path scenarios to ensure that it’s actually doing what it’s set up to do. It will have, like, adversarial scenarios, frustrated callers, things like that. And then it’ll also test, like, sort of, just sort of the edges of the scope that the agent’s been set up for, right? If I’m… trying to book an appointment outside of open hours, or something like that. How is the agent gonna handle that? If I’m talking about something that’s relevant to the company, but not something that the agent’s been set up to handle, how does the agent, you know, take that and direct that to a human agent, pass that off to a human agent? That’s part of Dialpad, right? So pretty straightforward. I can generate new tests. That will give me this suite of scenarios. I can test them on either voice or digital or both. And when I run this test, it’s actually simulating those live scenarios, right? So… this is pulling from stock data, but if this were in real life, it would actually take time to run those scenarios, right? Actual voice conversations are simulated voice conversations. Chat sessions are actual chat sessions, and they have all the same traceability that a regular conversation on Dialpad has. So I can go look at that transcript. And see exactly what happened for that particular scenario run. And then the results that I get, you know, look something like this, right? A description of what happened in each scenario, the channel all involved, the particular skills that are involved in the execution of that scenario, so I know right where to go back to and look at any potential problems that could arise, right back to that build process. And then difficulty. Complexity is really what this should be. We’re probably gonna update this to complexity. To match. the initial sort of outlay that it gives you. So that’s Proving Ground. So this is, you know, a chance to sort of pressure test the agent before you actually take it live. And I want to touch really quickly, I know we’re running up on time, on this sort of govern and optimize phase. Because it’s often not thought about, and this is more than just analytics, right? Analytics, yes, that’s kind of a table stakes thing. You need to be able to go in and ensure that The handle times and the CSAT scores for my digital agents are just as, you know, on par as my human agents. But, on Dialpad, we’ve also created what we call Guardian, which is a real-time conversation analyzer that happens in every single one of these agentic conversations, both voice and digital. And so, what that does is it analyzes every turn that’s happening in the conversation to keep it on the rails, essentially. There are 3 main analyzers that run currently, and we’re looking to expand these, but there’s a safety and compliance analyzer, that’s built on Google Model Armor. It will ensure things like PII doesn’t, you know, get printed to that chat conversation, or it protects against things like prompt injection. hate speech, things like that. So that’s the safety and compliance later. There’s a conversation alignment layer, which essentially ensures that the conversation is something that the agent can handle. We have really rich intent detection, so, you know, understanding if the conversation is in pursuit of that intent. And that can be updated as the conversation progresses as well. That intent can. But, you know, if it starts to… the user, the caller starts asking about something the agent can’t handle, it will pass that on to a human agent. And then the all-important frustration analyzer built on our sentiment model, that we’ve had for quite some time and have been iterating on. And then it also detects on if that frustration is actually directed at the agent, or just in general. Or not, right? So if I call in and I’m really frustrated about the product, but I’m readily receiving help from the agent, it’ll keep that session with the agent. And so that’s just a little bit about that sort of guardian layer, that runtime, safety layer that we have running. And so, Yeah, just kind of hitting on the points, your customer conversations are actually the root of all of your data. It’s the source of truth for what would end up being entered manually into a CRM, right? AI makes every one of those actions actionable, right? Even if it’s a customer support. you know, conversation, based on what happens in that conversation with Agentic, I can maybe drop that caller into an email drip campaign to promote, you know, additional services to them, for example, right? All of that can be automated. So because of that, customer success becomes a growth engine. One platform, it’s very important to have all the data centralized in one platform. And then also, when you’re thinking about human and digital agent handoff. having them on one platform is key, and the winners are going to be the folks who can really sort of operationalize all of this unstructured data, pull it into workflows, and execute on it. And that’s what Dialpad is there to help you do. I think I’m just at time.

Julia Nimchinski:
Thank you so much, Calvin. One of the questions that often come up here, and I see it in the chat again, so what’s the hardest part in actually making these agents reliable in production, and actually making them work?

Calvin Hohener:
Yeah, so I think that the key… To doing that is understanding the… The parts of the flow that truly need to be agentic, right? Where you need that sort of, flexibility to adapt to the, you know, different way people are speaking, the different ways that a conversation can flow to achieve an intent. it’s balancing that agentic sort of action with determinism, right? If I’m collecting a customer, you know, ID number or something like that, I can’t rely on an agentic system to pass on that exact string, you know. all the way down the line, right? So having those, those guardrails that you can dictate into the system, where you have very deterministic flows. Coupled with agentic flows, that are flexible and can handle real speech, and then also the agentic part of it allows actually writing to systems. I know a previous speaker was talking about MCP, that’s what we’ve standardized on as well. So that’s what you also get with Agentic, but it’s this balance of determinism and agentic flows that you really need to get right to ensure that the business outcomes are what you need. And, that’s something that we’re really focused on.

Julia Nimchinski:
comes to, you know, some of the… I don’t know, it’s too early to talk about ROI, obviously, but there are some results already that obviously you’re seeing in Dialpad. Could you share some case studies, anything more tangible.

Calvin Hohener:
Yeah. Yeah, sure. So one particular example comes to mind. So this customer gets a lot of, actually inbound leads. They’re a company that does, like, language immersion trips, so if I were a… Spanish professor here in the US, and I wanted to take my advanced Spanish class to Spain, I would… I would contract through this company, right? So they get a lot of inbound leads. And, they’re a global company, so their inbound call centers are only staffed during certain periods of the day. So one of the things that we’ve offloaded to Agentic is an after-hours lead qualifier. In the previous system. When agents got in the middle… in the morning, they had to listen to all of the voicemails that were left from all of the inbounds that they had from the previous night. And, in most cases, call those people back to understand if there would even be a qualified lead, you know, based on the number of learners that they have, the type of institutions they work for, etc, etc. There’s a number of qualifying conditions that that needs to be met. And their agents were spending, you know, up to lunch, essentially, just calling back all of the voicemails that they had collected during the previous day. So we created an after-hours agent for… Agentic agent for them, which is really simple. It says. you know, we’re closed right now, but I’d love to collect some information from you, and we can have someone call you back tomorrow, right? So it would actually collect all of those… those things needed to qualify a lead, right? The number of learners, the type of institution, etc, etc, and would actually write to their CRM if this lead is qualified or not. So now, every agent coming in in the morning is going right to their list of qualified leads and calling them back from the night before, and leaving all of the other voicemails that they got. So it literally took down that time that was spending sorting through voicemails from half of a day Down to, okay, give me the 5 calls I need to make right now, 2 qualified leads, so I can actually get to sales quickly. So that’s a use case that, again, is a little bit unique, kind of out of the box, but, like, ripe for an agentic solution to solve.

Julia Nimchinski:
Definitely amazing. As a product leader, last question, what excites you the most? What’s in your roadmap for 2026? What are you allowed to share?

Calvin Hohener:
So, I think… well, what am I allowed to share, and what am I excited about? I’m excited about these systems to get more robust and more sophisticated. What I’m really excited to happen, and I don’t think is really caught up yet, is… The consumer in this. I think as the general consumer becomes more comfortable interacting with AI, even using AI tools themselves, the more readily they’re going to want to talk to an AI to solve that simple problem, because they know it’s going to get done more quickly, and will be resolved with greater accuracy, and I don’t have to, you know, wait in a long call queue just to look up my, you know, the status of my order, and then give a bunch of information again, passing off context to different folks when I talk to them, right? So, I think, I’m excited for the customer sentiment and flexibility and understanding of these tools to catch up with where we are on the technology side, because I see a mismatch there. With that, I think another aspect that I’m really excited about is, verticalizing these tools that we’ve built. So, right now, we have a generic tool that, you know, it’s super flexible, can solve for a bunch of use cases. But I want to get to a point where we can go to a healthcare company and say, hey, we have these flows nailed for healthcare specifically, right? We understand the verification regime that has to be in place for healthcare, and so we have this complete end-to-end solution for you. Which, again, we do Already, from a product perspective, it’s all about, sort of, the packaging. in positioning of the tool that we have to really nail, you know, retail, really nail financial services. Healthcare is a big one that we’re focusing on. So that sort of verticalization and tailoring this generic solution to those flows is fun and exciting for me.

Julia Nimchinski:
And what’s next for DellPad for 2026?

Calvin Hohener:
Well, so a big focus is on our Gentic system. Again, we are a full end-to-end UCAS and CCaaS provider, so, people come to Dialpad for many different pieces of their communication stack. We really excel as that sort of end-to-end platform for your communications. And so, I’m excited to see AI sort of bleed into more of those use cases that we’re not traditionally thinking of. You know, maybe internal collaboration use cases that aren’t really being focused on now, for example. I see a lot of that. Again, taking just a conversation as the sort of basis for any data that gets written to any system, we don’t have to worry about that writing part, we can just, you know, mine conversations for the data that’s out there, and I think that that’s really powerful, in a lot of different contexts.

Julia Nimchinski:
Calvin, for everyone who’s watching, what’s the best next step to just, you know, engage here? Should they contact you on LinkedIn? Where should our people go?

Calvin Hohener:
Yeah, so, dialpad.com slash agentic labs. is a great place to start. You can find me on LinkedIn, Calvin Hohner, that’s C-A-L-V-N, H-O-H-E-N-E-R. I know my last name is spelled weird. But yeah, like, just talk to a vendor. I mean, the good vendors, like Dialpad, will have a lot of use cases in mind, and if you don’t know where to start. just go talk to somebody, and they’ll… they’ll have ideas for you. And so, come talk to us, reach out via that. If you’re already a Dialpad customer, talk to your, you know, support… your success manager, and we can get you folded in and started on this agentic journey.

Julia Nimchinski:
Thank you again. Thanks. Amazing.

Calvin Hohener:
Thanks so much.

Julia Nimchinski:
And… thanks. And that’s a wrap on Day 1. Phenomenal lineup of speakers and platforms. Join us tomorrow for Day 2 with James Carrier from NFX, Udi Lederg, Chief Evangelist at Gong, Manny Medina from Paid, and Chief AI architect from NYU Stern and Parce Pathil, who was the architect and is the architect of Reed Hoffman’s AI. So, very exciting sessions tomorrow and the next couple of days. Join us. And yeah, chat us in our Slack as well. See you tomorrow.