How AI Agents Are Redefining Customer Experience with Adrian McDermott, CTO at Zendesk

(Intro Narrator at 00:00:00) Today, we're talking to Adrian McDermott, CTO at Zendesk, about how agents have changed the world of customer experience. Thank you to DigitalOcean for sponsoring this episode. For simple cloud and powerful AI that's built to scale, visit digitalocean.com or just click the link in the show notes. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:27) I'm a fan. I'm a fan of Zendesk. You guys have been around for quite a while, and I'm excited to hear about what's going on and what's new right now at Zendesk.

(Adrian McDermott at 00:00:36) Yeah, this is going to really shock you, Joel, but I think so much of the excitement right now in the customer service industry and technology otherwise is around the application of generative AI to these problems that have been around for some time. I think when we think about customer service, it's in this sweet spot. Right? It's something where humans are generating responses, so generative AI works really well.

(Adrian McDermott at 00:01:02) Many of the tasks are kind of repetitive and follow a rule-based algorithm or workflow, and so it's just a really great place to be applying that technology. And honestly, you know, I've been at Zendesk for 15 years, but definitely the advent of the ChatGPT moment really reinvigorated, I think, how I felt about the industry and the innovation pace as well in software for customer service.

(Joel Beasley at 00:01:32) And so how is it going to change the human agent's role?

(Adrian McDermott at 00:01:36) Well, that's a great question. Right? And there's a number of ways to think about what we're doing. I think it's easy maybe to think about just basically flat-out automating what humans do in customer service, but I think it's a little more nuanced than that, and this is certainly what we're finding.

(Adrian McDermott at 00:01:58) If you think about being a developer, right? Coding, that has been changed radically by AI too. Right? I think customer service and coding are really the two forefronts of deployment. And with coding, no one's out there saying, "You know what? Now that my engineers can be 2x, 3x, 5x, 10x—you know, pick your number—more productive with Copilot, Cursor, Codex, whatever tools they're using, what I'm going to do is I'm going to fire a whole bunch of developers."

(Adrian McDermott at 00:02:29) That isn't what people are saying, because one of the things that they acknowledge is no one has enough products. No one has enough innovation. We need to go faster. We need to build these differentiating aspects of our company. And so everyone's doubling down, and they're using these tools and trying to do a lot more.

(Adrian McDermott at 00:02:44) And then in customer service, it's weird that we don't apply the same rubric. Right? Because if you talk to customer service leaders or you look at business outcomes, I think none of them are saying, "You know what? I have enough customer service. My customers are good. I don't need to do anymore." In fact, most people will tell you that they have a service debt. And so as we apply tools to automate customer service and apply AI-based technology, what we see is the first place people go is to paying down that debt. It's, "Okay, now I can offer chat support."

(Adrian McDermott at 00:03:21) "I can offer 24/7 support. I could cover all human languages that are basically available reasonably." You know, there's all these things that people want to do to reach out to their customers more. And so the first big change is in just the bandwidth that you can offer.

(Joel Beasley at 00:03:37) So you're saying people are in this CX world, which I don't spend a ton of time in, rather than laying off the developers or laying off the CX people, they're instead improving the quality that they can offer with those individuals?

(Adrian McDermott at 00:03:52) I think so. Yeah. I think we're applying AI in kind of three main ways. Right? There's classical automation. Right? So, you know, think about the chatbot. And those have gotten much more sophisticated, and we can talk about how they've been applied. There's human-in-the-loop flows. Right?

(Adrian McDermott at 00:04:10) You know, co-piloting, if you will. Incredibly important in customer service—high-turnover business, often a lower skill profile for complex problems. That ability to tee up the correct answer for someone and drive consistency and rapidity of generation of answers is super important. And then, you know, there is this other ability that you get with generative AI. The idea that you can look at every single interaction that comes into your company and begin to draw insights from it and begin to understand how to implement continuous improvement in your company, which is, you know, I think incredibly powerful.

(Joel Beasley at 00:04:53) So you've got a lot of low-level work that's being automated. And it was being automated before the GPTs—it was just standard code. Right? We were all working towards this anyways.

(Adrian McDermott at 00:05:03) Yeah. I think, you know, customer service has gone through a bunch of transformations, you know, the phone, which is when we invented the rituals of customer service in some ways. "Is there anything I can help you with?" "How would you rate your score out of 10?" "Please hold." You know, those kinds of things. Those rituals then gave way to self-service, which I think is something that we've been driving for, you know, basically two human generations. I think two human generations have been trained to put text into a little box and get 10 blue links and process them. Right? And then we also had chatbots.

(Adrian McDermott at 00:05:38) They were mostly rule-based. And, you know, we had one that we were selling that was implemented, and you could kind of get to maybe 20%, 30% automation on a good day depending on your use case. Right? I think the difference with generative AI, you know, if you think about our advanced AI agents at this point, we're using a frontier model so that when we first meet Joel, you know, the task identification agent, which understands everything that I can do in this customer support environment, is having that conversation with you, and it's looking at—it's representing the brand, but it's listening to your tone and your state of mind and your emotion. It's trying to tease from you what it is you want.

(Adrian McDermott at 00:06:19) And if you want to do a return, which product is it? It's asking you about your product problem and then delving into information. And so, you know, using frontier models, you can do things that would in no way be possible before, even with code.

(Joel Beasley at 00:06:35) Yeah. And I'm looking forward to that as a consumer because I can tell right now when I'm interacting with these customer service bots or systems that they are not all the same. Some are way better than others. Some, it almost seems like it's a person I'm working with that can pretty much do anything on my account. The others feel like it's very rigid, like they have access to a knowledge base and they can spit back to me the same stuff I could Google.

(Joel Beasley at 00:06:59) And so I'm looking forward to as they become more autonomous and they can do stuff. And I think even the ones that can do some things, they filtered down what they can do versus what a human will do. So sometimes you have to go to a human, and I'm looking forward to that day when I could fully, completely self-serve with the robot.

(Adrian McDermott at 00:07:20) Why do you hate people so much, Joel? No, it's a good goal. I think one of the things I would say is that I think the future is here. It's just unevenly distributed.

(Adrian McDermott at 00:07:33) Because, yes, if we look at our data, you could get to 40% with RAG knowledge, you know, capturing standard operating procedure and capturing policy. And that's within reach of most people. I think one of the things from a CIO, CTO perspective that has to happen, though, is if you've run a sort of semi-complete or incomplete digital transformation in your organization—right? So, basically, you know, if I am a—we currently don't deploy computer use in chatbots or agents in customer service.

(Adrian McDermott at 00:08:10) I don't think it's a little early with computer use to do that and take those risks. So what we have to do is basically ask a customer, "If you want to solve this next 5% of tickets, I need access to the order management system. So do you have a REST API that performs this, that lets me retrieve this data and write back this data and do these things?"

(Adrian McDermott at 00:08:32) And so actually, it's not the reasoning engines. You know, we're not waiting for a new generation of model or something that can also solve the Riemann hypothesis here. We're actually just looking for a REST API and just general digitization of the company's processes such that, you know, the only way to achieve a task isn't what we call swivel chair integration, where an agent is logging from one system to another and flicking between tabs in their browser.

(Joel Beasley at 00:09:01) That is so cool. And I don't hate humans. I love the humans. I married one and made three. So I do like them. But what I also like—because I like a high quality of experience.

(Joel Beasley at 00:09:14) You know, if you walk into a Taco Bell because it's the only thing open and you have to eat, you love the fact that there's a kiosk and the food just comes out. I like the convenience factor. I think the humans will find other better things to do. You know, I don't think they need to be sitting there eight hours a day typing into a little box.

(Joel Beasley at 00:09:35) They need to go out and live some life, enjoy some sunshine, you know, and figure out what else they can contribute to humanity.

(Adrian McDermott at 00:09:42) Yeah. I think you want to have complete agency and control over your support or ordering experience. Right? And we all, to a certain extent, feel like that. Recently, we sampled, you know, 15 million ticket conversations across a whole bunch of different Zendesk customers, kind of looking at some of these things, and we were categorizing what is the nature of the inquiries that they were making to these companies.

(Adrian McDermott at 00:10:11) Like, what is the nature of customer support that they have? And you can say that about half, approximately, was people dealing with a failure. Right? Something went wrong with the product, the service, the delivery. They can't understand the technology.

(Adrian McDermott at 00:10:27) They're asking questions, and they're reaching out for remediation or help. It's about half. There's another 25% which are these moments to engage with your customers where they want information, they want to be informed, they want their experience with your brand to be enriched. And actually, in many cases, they're actually looking for a person. Then the final quarter of tickets that come in are people who, you know, what they need to do is right there on the website.

(Adrian McDermott at 00:10:56) It's right there in the app. The delivery date is December 10th, and it says it right up at the top. And they're still calling to say, "When's the delivery date?"

(Joel Beasley at 00:11:05) Or they're still calling to say...

(Adrian McDermott at 00:11:06) "Can I get it on the 9th? It would be great if I could get it on the 9th because that's my son's birthday or whatever." And those things, I think you and I might have this preference, you know, given our slightly different but similar demographic in terms of technology. We have this preference sometimes for that control and that self-service, and we really enjoy it. But the 25% of people who just kind of want to do it verbally and have a human help them, you know, maybe you can automate about half of those.

(Adrian McDermott at 00:11:38) Maybe you can automate about half of the failures. But, you know, I think it has to be the choice of the customer, not the choice of the brand. Right?

(Joel Beasley at 00:11:49) Yeah. But the brand offering me all available options would be great.

(Adrian McDermott at 00:11:53) Yes. We definitely want to automate—or I think you automate 100%, but you have humans. You're going to need humans to answer some percentage of those depending on what you do, unless you want to go to the other extreme. Right? There are businesses that famously provide no customer support email or phone number. Everything is in the app, and it's your challenge, right, to get out of the escape room of their giant website and figure out how it is that you do a product substitution or which new product is...

(Joel Beasley at 00:12:22) You need support, so you start a podcast. You grow it, make it huge, you get the executives on your show, and then you can finally ask for support. I have done that before.

(Adrian McDermott at 00:12:34) I will not lie. Right? If I'm reaching out for support for a company—like, send them, you know, hit the email address or something—and I get a response back and I recognize it as a Zendesk thing, I never don't respond by telling them that I'm the CTO of Zendesk, just in case it gets me better service. Often, it just gets me feature requests. But there you go.

(Joel Beasley at 00:12:55) Right? Yeah. Yes. I always tell myself when I'm talking to all these great people from the podcast, "Alright, this is not the time for feature requests."

(Joel Beasley at 00:13:07) Okay. So, curious—are your customers, they're experiencing the shift, right? So they've got their teams, they're using your product, but you're rolling out new features. Are you impacting the structure of their teams constantly? How is that working out?

(Adrian McDermott at 00:13:20) Yeah. I think if you think about how a support department is classically structured, right, you have tiers of support. And so, you know, tier zero, tier one is kind of triage, simple break-fix, simple question and answer, I think, most impacted. Right? These are the people in your words you think should get out and go into the sunshine.

(Adrian McDermott at 00:13:43) I think that so many of them get upskilled and they move to tier three, tier four where there's complex problem-solving steps or complex relationship steps or things that happen. Right? And so how can we upskill so many people so rapidly? Well, I think that's where the copilot experience comes in and the insights come in, right, where if the greatest agent that you have is sitting right next to me, teeing up the answer and drafting it and going into my back-office systems and pulling the data out about Joel's account and drawing conclusions, it's really going to help enable more people to do that.

(Adrian McDermott at 00:14:21) I think above that, right, I think it's a huge change for, you know, what is often described as the most important role in customer service, which is the team lead who's managing both how support is given and what to do with new types of issues, but also motivating and training the humans and making sure—you know, staff retention is a huge problem in these jobs. Right? And so making sure that people are motivated, showing up for work, and logging on to the right channels to give support. I think there, the job changes where they now have two classes of agent. Right?

(Adrian McDermott at 00:14:55) They have an automated agent that needs training, that needs care and feeding. You need to look at the data. You need to report on them. You need to look at the QA tool, which is inspecting the conversations and understand how is this tone working, how is the nuances of this conversation. Are we adhering to our actual policy or hallucinating?

(Adrian McDermott at 00:15:15) These things can be problems. And so that role changes dramatically. And then I think in terms of back-office and support folks, we talk about the emerging skill set of the AI service architect. Right? Everyone in service knows how to write a knowledge base article and document standard operating procedure for agents and take them through a flow—human agents.

(Adrian McDermott at 00:15:38) I think balancing that out so that you can cover 100% of cases and give autonomy and do that for an AI agent, which is a brand-new skill and non-obvious, is really the most disruptive change in contact center work. And it applies actually broadly to most industries where folks woke up one day and found out that they were supervising AI agents, and they were responsible for care, feeding, training, and management of those things, which, full disclosure, I think in most cases, not all of the tools that you would use to manage an AI agent exist yet or certainly have been fully built out.

(Joel Beasley at 00:16:20) I think when you mentioned turnover and the customer support, I could totally see that. I often, people will comment on how calm I am, which leads me to be like, oh, you're getting crazy calls. Like, people are yelling at you. I can see how that would happen. I can see that that would create turnover.

(Joel Beasley at 00:16:40) It will be funny in the future when they're just yelling at the robots, although I do yell at some of the robots today. But my question is this. We look at support as human contacting a company and that company having a human hybrid technology support stack interface. Is there anyone that is building, like, another tool where it's like, I just go to that tool and say, hey, I want support, and I just give it access to everything I have, and then it goes and gets the support for me? And then eventually, it'll just be, like, my robot talking to their robots.

(Adrian McDermott at 00:17:14) Yeah. I mean, the MCP, the Model Context Protocol, and agent-to-agent standards exist. And I think one of the challenges with LLMs, right, is their lack of memory. You know, Claude Code wakes up today and is a significantly better coder than me on day zero. It's actually always gonna be a significantly better coder than me, actually. There's a reason I don't code anymore other than vibe code. But it wakes up, and it's like a new employee that comes in. Tomorrow, it's the same new employee, and it's forgotten everything that it learned, basically. Like, it's not growing out that context window. It's not learning the way a human does in that way.

(Adrian McDermott at 00:17:55) And so I think you have to adjust kind of the way you think about deployment to really understand that. And I think that will change. Right? But I think, what would we like to see and use in a model? Well, I'd love to see really, you know, really long context windows being used so that everything Joel's ever said to me, all my relationship as a company to you as a customer of that, whether it's B2B or B2C, or you're an employee of my company. Right? I'd love for my model to know all about that. And to me, that's the ultimate CRM where some photographic infinite photographic memory knows everything that we've done together and could probably immediately remove friction. Right? If you think about what is the biggest source of friction in customer service, it's you being forced to repeat yourself.

(Joel Beasley at 00:18:48) Yeah. Have you looked at the Claude projects or the Grok projects yet?

(Adrian McDermott at 00:18:57) No.

(Joel Beasley at 00:18:59) Yeah. So these are newer features that have come out pretty recently where you can actually give it this back. You could train it more like an employee. And every time that there's something new it needs to learn, you can put it into it. And then when you go start a conversation inside that project, it's trained on everything that's previously known from that project.

(Adrian McDermott at 00:19:17) Yeah. I mean, it's a way, an elegant way to stuff the context window.

(Joel Beasley at 00:19:23) It's an elegant way to stuff the context window. Yes.

(Adrian McDermott at 00:19:27) I think, back to your original question. Right? That memory will make those personal assistants really personal assistants. I think, yeah. As that skill develops, we would all, right, we would all wanna send our personal support agent with a memory of us who understands our cell phone contract off into the cloud to go figure that out and kind of deal with it for us. Like, who wouldn't really? That would be just this is gonna be a joyful moment for us all.

(Joel Beasley at 00:19:58) Right? What AI are you using the most in your personal life outside of work?

(Adrian McDermott at 00:20:03) Well, I am promiscuous. I try to use a lot of AI and move from thing to thing. So Gemini 3 was released yesterday. Oh. And so for me, I just started, I switched. I was using ChatGPT. So I drive to work, and I'll drive around and I talk to my, I talk to my phone or I talk to my car, which also has a model built into it. And I have conversations and I do love that, and so I switch out whichever model it is. So this week, as of today, I'll go home and I'll talk to Gemini, the Gemini app. And it'll be great. It'll be great.

(Joel Beasley at 00:20:44) That's, you know, the first person I ran into that did that, Mohak, the CTO of LinkedIn.

(Adrian McDermott at 00:20:49) Oh, who?

(Joel Beasley at 00:20:50) Yeah. And when he shared that with me, I had never thought about it. Like, I'd never thought about doing that. I think it was right when the voice feature first came out, and he shared that he did that with me. And then I did it, like, the next day just to experiment. Now I do it. In my car rides, I just talk to the AI. It's great.

(Adrian McDermott at 00:21:08) It's one of the reasons I have such a big preference for Waymo over Uber. Is because I feel weird talking to an AI in an Uber with a human in the front seat. It just seems awkward. But when you're in a robot car, talking to a robot is, like, supernatural. Right? I love that.

(Joel Beasley at 00:21:31) I do love it. So I actually have become a fan of Gemini pretty recently, mostly because it was the only context window that could accept video. Like, audio and video. So I could ask it questions about audio and video. Like, Grok wasn't letting it happen. Claude, GPT was like, no. And then I thought it got a lot better. When it first came out, I was like, but then over the past two years, they've really stepped up their game. I haven't tried 3. I just learned that it's out from you. So I'm gonna give that a shot after. But I find myself navigating and trying different things and usually, almost always coming back to Grok.

(Adrian McDermott at 00:22:07) Oh, interesting. I'm a non-returner because by the time you're sort of, you're tired of your latest AI best friend, there's another one that just, you know, there's another, like, oh, there's another one. And so then you're like, oh, I'm gonna try that. And then there are some tools that I continue to have just, I take great joy from, like, I love Notebook LM. Oh. Google Notebook LM for summaries of things. You know? I receive a lot of written or verbal content, and my capacity to consume it is far, you know, is far outstripped by my desire to consume, and so that really helped me with that. And, you know, you can kind of be lazy. So even if I don't wanna talk to a model on transportation, I'm listening to a model that was doing the reading that I should be doing, but I'm sat there with my eyes closed. Unless I'm driving, obviously, and do that. So I really like that. And then I've been trying to vibe code more and, you know, use different models for that. I started with GPT and built an iPhone app, which I've never done before. And now I'm, like, trying to get Claude Code to, I was inspired by one of my coworkers who started writing his weekly team reports and people reports by plugging Claude into a bunch of different, you know, a bunch of Confluence and Slack and a bunch of other tools that we use and then asking it questions about all of those conversations and having it figured out. I'm like, oh, I should, I wanna, I really yearned to be able to do that for myself.

(Joel Beasley at 00:23:48) Are you doing, like, Cursor vibe coding, or are you doing, like, Lovable Dev vibe coding type of?

(Adrian McDermott at 00:23:53) I work with someone here who's, like, totally Cursor-filled. He has Claude Code. He's doing this massive refactoring project we're talking about the other day. He's called Code Run ahead of him and generate an enormous amount of changes, and then he loads it into Cursor and resolves all of the issues and inspects it and gets the tests working again and everything else. Right?

(Joel Beasley at 00:24:18) That's awesome.

(Adrian McDermott at 00:24:18) So that's probably next on my list, but I'm a bit of a brute force text editor guy. So, yeah. I haven't gone all the way in yet.

(Joel Beasley at 00:24:29) There's a couple startups that we've talked to out there that are doing these, I think two of them in the past year that they specialize in the assistive, you know, Copilot stuff, but large context code, codebases. So they'll go into enterprise that have massive codebases and then make it easier to use the assistive models with them because they can get kinda lost with the Copilots and stuff. They can lose context. You know?

(Adrian McDermott at 00:24:54) For a 17-year-old code base like ours with, you know, 14 acquisitions, that is going to be invaluable. Invaluable.

(Joel Beasley at 00:25:03) Now, for a quick shout out to DigitalOcean for sponsoring the episode. DigitalOcean is a cloud infrastructure with integrated AI tooling. Combining reliability and affordability, there's nothing you can't build on DigitalOcean. Simple cloud, powerful AI, built to scale, that's DigitalOcean.

(Intro Narrator at 00:25:20) Now, back to the episode.

(Joel Beasley at 00:25:22) Now let's talk about the resolution platform. What is that at Zendesk?

(Adrian McDermott at 00:25:27) There's this movement in SaaS. Right? And SaaS is moving from seat-based to consumption-based. And so then, you know, we're all kind of thinking about what is the consumption metric that works for our users. If we take inspiration from AWS, those things are usually driven by value or cost. And we were sort of, interactions was the normal model, but I think we sort of felt like we could leap into the future a little bit and go with resolutions. So, you know, a resolution, right, is AI being applied. And so Zendesk is a platform, you know, probably, I think it's over 4 billion resolutions mostly human-powered, but we probably have 800 million AI interactions in '25, and what we want to get to is a platform where we basically have the same incentive as our users, and then they're paying us for that. Right? Which is, you know, I'm here to help my users resolve their problems. Right? I'm here to help my customers resolve their problems. And so tie, build a platform for that. Tie your product to it. Tie your pricing to it.

(Adrian McDermott at 00:26:38) And, you know, really shift from, in some ways, seat-based economics to resolution-based economics. So the resolution platform is all of the things that you need to be able to do that. You need Copilot technology. You need QA technology. You need agents. You need insights technology like an admin Copilot, like helping you optimize and tune the system. You need knowledge agents. You need RAG agents. You need generative search, and all of those things come together to solve the big problem of I want to generate support resolutions for my users as efficiently and joyfully for them as possible.

(Joel Beasley at 00:27:08) That's pretty cool. Are you at the point where if there's a startup that's, like, super modern SaaS type startup, and they wanted to just give you full access to, like, almost all of their data and then have your AIs just digitally serve the people. Are we there yet? Are we still a little bit of ways?

(Adrian McDermott at 00:27:24) Pretty close. Yeah.

(Joel Beasley at 00:27:25) Pretty close.

(Adrian McDermott at 00:27:26) If we take a company doing, you know, human-powered support right now on Zendesk, you know, we can just go in and read, you know, tens of thousands of past interactions, past tickets, and then generate a fully functioning knowledge base. So then we'd turn on generative search immediately. And so you'd be up and running at the front door of your support site with an answer system that's probably gonna hit, if it's covered your ticket database, it's probably gonna hit 40% of your inquiries. Just, you know, that's sort of, like, one, two click kind of work. And then adding in sort of the Copilot experience and the agent experience, that comes, you know, that requires a little bit of human judgment because what you're talking about there is policy. What you're talking about there is how you as a company wanna treat the people that reach out to help you. Or, you know, back to the original numbers. Like, how many of those people asking to get that delivery on December 9th are you gonna connect with them? Because, you know, bad experiences in customer support drive customer churn.

(Joel Beasley at 00:28:30) Right?

(Adrian McDermott at 00:28:31) And we also know that automation drives escalation. Right? The more you automate, the more people are gonna start hitting zero because they don't feel like they're getting what they wanted.

(Joel Beasley at 00:28:39) That's me.

(Adrian McDermott at 00:28:41) There you go.

(Joel Beasley at 00:28:42) That's me. Every time.

(Adrian McDermott at 00:28:44) You're the canonical customer support guy, obviously.

(Joel Beasley at 00:28:47) I just wanna talk to a person when I do it. Like, unless if their systems are good enough to where they can serve me. Like, if an all-powerful AI answered the phone and was, like, I will do your bidding and I could get whatever I need done done, I'm happy with that. But often, it's, like, it's very low intelligence AI in my personal interactions.

(Adrian McDermott at 00:29:09) Yeah. I think, as I said, the future's here. It's unevenly distributed. Right? Yeah. And some of what you want actually is not, I think, is not necessarily the AI. Right? You know, you're doing that bean burrito run at 2:00 in the morning, and you're standing in Taco Bell. Why are you using the kiosk? Because you like to self-serve. There's no AI. I mean, I don't think there's a lot of AI in the, I have no lived experience, but I don't believe there's any AI in the Taco Bell kiosk experience. I'm sure it's lovely, though. And, you know, it is fast. But I think we all want that experience in many ways from the brands and tools that we interact with. Right? You know, we want agency to be able to do it ourselves. And that isn't necessarily, that's how well did you do on your digital transformation. Like, are you, were you born in the cloud? Were you born mobile? Are you getting there? Are you all in? Like, how much control have I given Joel? So I think it varies in that way.

(Joel Beasley at 00:30:13) Are you guys the first to do the, so resolutions are outcomes. Right? And then you're changing your pricing model to be, like, outcome-based pricing. Is that correct?

(Adrian McDermott at 00:30:21) That is correct.

(Joel Beasley at 00:30:22) Are you one of the first people to do this, or is everyone in the industry switching? What's going on there?

(Adrian McDermott at 00:30:27) I think we are one of the first. It's certainly the first scale player to be playing around with these metrics and kind of align with our customers' outcomes. Right? I think, you know, we have 20,000 customers now using our AI in some way in generating responses. And so I think we've been trying to move as quickly as possible. It's clear it's gonna disrupt customer service. Right? And, you know, the classic texts of Silicon Valley, Crossing the Chasm, The Innovator's Dilemma, they talk about how you have to go disrupt yourself first. And so we're kind of all in on that, trying to innovate as fast as possible.

(Adrian McDermott at 00:31:08) Now some of these experiments may not work. Right? Because if you are, if you've been, you know, support is a factory powered by humans. Right? And so you know that if you have someone in the seat, it costs you this for the human and this much for the software, and you'll get this much throughput out of them. 60 tickets a day, whatever it is. Right? Whatever your metrics are. I think it's a new world to then have it be variable. And you wanna know, you know, what our customers wanna know is they can turn the spigot on and off, that we're gonna shoulder some of the risk with them.

(Adrian McDermott at 00:31:43) And at the same time, they're making these commitments, and they want to know that we can get them to those higher automation rates so the economics work out for them. And so, to me, it's all about shared destiny and trust to bring our customers along. And we're inventing the rule book as we go, and that's great, right? Because that means that we're innovating.

(Joel Beasley at 00:32:01) Yeah. Because what happens if you make the customer support so good that people start using it so much, and you're delivering so many outcomes that it's just wild? And you don't have that natural limiter of the human time in a day. It's this weird efficiency increases usage thing.

(Adrian McDermott at 00:32:18) Jevons paradox, baby, as we like to say. So yeah. And actually, we've seen that. Right? When all service was in person, you know, if you didn't want to go talk to that person, you didn't really seek service. Right? Then the phone came along. It was a little more impersonal. You could do it. The Internet came along. Website was on 24 hours a day. You could just go do a search. I think AI experiences and ultimately, you know, we are examples talking to our telephones as we transport our atoms. Right? We're examples of people who—great voice experiences are very low end. Well, you know, they don't take a lot out of you when it's automated the way it does when it's a person.

(Adrian McDermott at 00:32:59) And we absolutely think two things are going to happen. Support volumes are going to go up as your experience gets better, and that's a measure of success and not failure, because you're building connection with your customers. Right? You're actually growing, in some ways, your brand or the minuscule or massive amount of importance you have in their life. That's fantastic. But the other thing is, back to the principle, automation drives escalation. And so you will still be needing some humans sitting around to help you with those escalations and to deal with Joel when he feels like he isn't getting the answer he's looking for.

(Joel Beasley at 00:33:34) You'll always need it at some level, right? Because even—yeah, because you have big clients. You also have a portfolio of clients. You have some customers that are lower lifetime value and some customers that are higher lifetime value.

(Adrian McDermott at 00:33:46) Yeah. There's no one size fits all. Right? And, you know, Zendesk support is B2E, business to employee. It's B2C, business to consumer. It's B2B, business to business. All very different dynamics, different within verticals. And so I think this is where, you know, chief customer officers and support leaders really think about what are the customer journeys I want to enable, and who, you know, from an economics point of view and also from just a brand point of view, who do we want to be and how do we want to show up?

(Joel Beasley at 00:34:19) What do you think this teaches us about the future of work as a whole? Like, where do you think we're going in the next 10 years?

(Adrian McDermott at 00:34:29) I think, you know, you can spot two not very hard to spot trends in customer support that apply broadly to work. Right? One is that you can automate some of the repetitive work, but not all of it. And it can be really efficient and actually is a good experience for customers. Another is that the more that we apply AI assistance to tasks, the more productive we get, the more consistent we get, and just the better results that we can get from that.

(Adrian McDermott at 00:35:06) And so much of—I think so much of the—so many of the tokens produced in the world, probably the majority, are not full automation. They're human-in-the-loop assistance. And I think everyone—you know, this company is famously measuring their employees on how much AI they use. It's not so that they can replace them. It's so that they make sure that they're, you know, innovating, thinking in a modern way and driving towards better outcomes and better productivity. I think customer support, you know, and software engineering—but we're already living in the future where most people are already doing so much of their work assisted by AI. There is kind of a third category, right, which is something that we're really seeing in customer support, which I would say is, you know, if you're a scaled business and you do 100,000 customer conversations a month, right, 100,000 tickets a month, when you were only doing a hundred, it was pretty easy for the team lead to read all of those and go back to the product team and say, "Yeah, I think you need to do this. This is what we're seeing." You know, these things are coming up, and you could spot trends and think about how to change the company or the product or the service or service itself. I think at 100,000, you lose that ability, or at a few thousand. I think we can add that back in. Right? We have these systems of intelligence that can read every single inquiry.

(Adrian McDermott at 00:36:27) Right? And so we have a tool right now, a QA tool that our customers use, so quality assurance of every conversation. And initially, that started out in the old era of, you know, Adrian's a customer service agent. You would sample a hundred of my tickets, and then you would score me for grammar, politeness, completeness. Was there a salutation? Did I say goodbye? And I checked that everything had happened. Did I comply with whatever? And then it became AI-enabled so we could read every single ticket, and we could score them on those rubrics, right? And we actually built little models based on BERT that could do that across, like, 16 different things. At this point, we can apply, you know, a fairly sophisticated—you know, I think we're using Llama 8B, if I'm not mistaken, model. And you could say, you know, if you're a gaming company, a gambling company, like, does it sound like Joel has a gambling addiction? Because that's something that needs to be flagged, right? Does it sound like Joel's underage? You know, you can start to ask these meta questions. And so, you know, it's one of those cases, right, where if you have a problem or an opportunity in AI, you can probably solve it with more AI. And so having AIs watching your AIs watching your AIs is actually extremely powerful. And I think for all jobs, I think customer service is early in this, honestly, where we're capturing these insights.

(Adrian McDermott at 00:37:51) And I think it's interesting to think about all back office, front office jobs and how they transform when the AI is monitoring and pulling insights by being able to see the breadth of your experience or the breadth of work across the whole company.

(Joel Beasley at 00:38:06) I like that. And you're going to need the humans to watch the AIs that are watching the AIs that are watching the AIs. That's going to be very important, because you have to almost observe that it's happening correctly, right?

(Adrian McDermott at 00:38:16) Someone has to do something with the result. I mean, recommendations that, you know, are not read—they're like trees that fall in the forest. Do they make a noise? You know, how do we know?

(Joel Beasley at 00:38:27) I don't know. We're going to—I'm excited about the future. I think have you heard Elon mention this, like, universal income thing? I think that's a pretty cool idea. The automation—like, once you have robots that can make robots that can mine for minerals to make more robots, you kind of take the labor concept and you make a whole new type of economy possible with it. And I think it's going to be beautiful.

(Adrian McDermott at 00:38:51) I think it is, and I think it's sort of—when creativity can kind of create things unbounded by effort, by human effort, you're still bounded by resources, right? But you can do so much more. And we sort of—we see that in software engineering. Right? Would my colleague have taken on refactoring the entire top bar with its 17 years of cruft if he didn't think to himself, "Well, I could actually do that now. I could do it on my own," you know, which would have been two scrum teams probably for half a quarter. You know? But, "Yeah, I'll just do that."

(Joel Beasley at 00:39:28) It's Tuesday. I'll just do it. It's the afternoon, and we'll just get it done. You've been at Zendesk 15 years. Is that correct?

(Adrian McDermott at 00:39:37) That is correct.

(Joel Beasley at 00:39:38) How many people were there when you started?

(Adrian McDermott at 00:39:41) 30 to 40.

(Joel Beasley at 00:39:42) Wow. Where are you at now?

(Adrian McDermott at 00:39:45) Over 6,000.

(Joel Beasley at 00:39:47) Whoa. So you've learned a lot. And did you—how did you rank up there? Where did you start?

(Adrian McDermott at 00:39:54) I started—I think there were 10 engineers, a couple of founders coding or designing, a product designer and a product manager or something, and I came in to run product and engineering. Okay. Because I think the title was VP of Engineering. And then sort of grew that team to maybe 1,500, 2,000 over a decade. I think we IPO'd after about four years, and we had a billion in revenue after about 10 or 11, something like that.

(Joel Beasley at 00:40:23) So it's just constantly been a new—I don't ask people anymore why they stay. I've learned that when they stay for 15, 20 years, it's because the job constantly changes. Yeah. Has it been exciting?

(Adrian McDermott at 00:40:34) It is. It is. I think in the AI era, what's been really exciting is I actually don't manage 2,000 humans anymore. I'm not the societal administrator of that small hill—of that medium-sized town anymore, which is great. I just focus kind of on strategy and M&A, which is something we've been very active in, partly because I think we're in a period of massive disruption, right? And in those periods, I think you have to spend a lot more time thinking and experimenting and mostly listening. And there's a lot of—there's a lot of challenges in customer service land where people are getting pushed to massively reduce their workforce. So, you know, "Why aren't we—you know, company X laid off 700 people. Why aren't we?" And these kind of things. And then there's also this rational idea of, "I'm driven by—yes, we're driven by cost, but we're also driven by NPS and CSAT and gross retention rate on our revenue, right?" And that understanding. So I think spending a lot of time with customers on those things and with other technology companies has been refreshing. Let's put it that way. That's been a fun way to kind of pass the time the last couple of years.

(Joel Beasley at 00:41:59) What is just-in-time leadership?

(Adrian McDermott at 00:42:03) For me, just-in-time leadership is something—it's a principle that I definitely use. So I've done three startups, two hits and a miss, and scaled kind of teams in all of them, right? And you begin with—you begin where you can all fit in the standup, and then you can't, and then you kind of need to separate and you need a layer of management, and then maybe you need an engineering director or a product director, and then maybe you do. So as you stair-step up the human complexity continuum until eventually you—what is the—there's that number, right, the point at which you can't understand the relationships between all of the people in a given group. I think for humans, it's about 130. For dogs, it's about 20. Yeah. Dunbar's number. Yeah. And I think as you hit those limits, then you have to think about, "Now I need more leadership." And some people who are on the journey with you are, like, ready to step up and own groups of groups or teams of teams. And sometimes you have to go outside and find someone who's seen the movie before because you're growing so fast that you need that experience. And so—and often you make mistakes, and that's okay. You make a mistake. You acknowledge it. You correct it. You keep going. But I think just-in-time leadership is about knowing when it's time for a structural change as you grow. You know, growth is the most amazing thing to inject into a business. It also covers a multitude of sins, let's be honest, but it is the most amazing thing to inject into a business.

(Adrian McDermott at 00:43:33) And so as you do that, having a strategy for, you know, breaking what you've been okay with, you know, destroying the structures that you built and building new ones every now and again, not to generate chaos, but, you know, I often say in leadership and in management, you know, you need to be a lot less like a vending machine and a little bit more like a slot machine from time to time. Someone should be able to come up to you, hit B12, and get a Snickers bar. Right? That's just not how it should be working.

(Joel Beasley at 00:44:02) I haven't heard this one before.

(Adrian McDermott at 00:44:04) You haven't? Yeah.

(Joel Beasley at 00:44:05) No, no, no. Can you say that again?

(Adrian McDermott at 00:44:08) Yeah. I think as a leader sometimes, right, you have to not be predictable or be there to take orders for employees. Right? They should come out, put a dollar in, press B12, and a Snickers drops out. They should put a dollar in and the reel spins, and that is, like, two cherries and a lemon because they get three cherries. And they should feel a sense of angst as they get into that. Like, "This isn't just the office that I go to when I want to hire another person. I want more resources or blah blah blah." Right? Yeah. I'm not the universal donor of succor and joy.

(Joel Beasley at 00:44:40) I married a slot machine. I have no idea what I'm going to get sometimes.

(Adrian McDermott at 00:44:46) There's no comment that I could make—

(Joel Beasley at 00:44:49) No comment. On the shelf. Okay. You talk a little bit about early—

(Adrian McDermott at 00:44:53) You could go home and tell, you know, vending machine, slot machine continuum. I'm guessing you're a vending machine at home, by the way.

(Joel Beasley at 00:45:01) Me? Yeah. Yeah. I'm pretty consistent with my kids. I'm pretty consistent. Yeah. Yeah. I think they're eight, six, and three. And they—I want them to see me as the pillar. They don't always like what they get out, but it'll consistently come out. Yeah.

(Adrian McDermott at 00:45:19) Oh, that's awesome. Yeah. Mine are 24 and 21. I just want them to see me.

(Joel Beasley at 00:45:30) Oh, wow. Are they out in the wild?

(Adrian McDermott at 00:45:34) Yeah. They are. One's in college and working, and they're both in—they both live in Southern California.

(Joel Beasley at 00:45:40) So you get to see them a little bit. You're not too far from them.

(Adrian McDermott at 00:45:43) We're all going on vacation for Thanksgiving. It's going to be awesome.

(Joel Beasley at 00:45:47) That's good. Boys, girls, both?

(Adrian McDermott at 00:45:49) Two boys.

(Joel Beasley at 00:45:50) Two boys. Okay. Boy dad. Let's do it. Yeah. I've got one girl and two boys. So yep. Right in there. All right. I wanted to wrap up on some leadership stuff. Early career versus late career leadership, do you have some thoughts on that? Can you walk me through it?

(Adrian McDermott at 00:46:08) One of the things about early career, right, is that we often spend too much time proving ourselves, and so that means that we often, I think, are too defensive in terms of the way that we think about our teams and the way that we work. I think late career, or what am I in now? The late twilight of my career. You can no longer read without some kind of mechanical assistance. But as you get into that later period, right, I think one of the things that you are much more comfortable with, and you need to be comfortable with this pretty early on, is not being the smartest person in the room, not having all the answers, but having some good questions.

**Adrian McDermott at 00:46:49**
And I think those, you know, I think as you go through your career, right, in corporate roles, you endure or occasionally enjoy almost infinite management training and guidance, right? And some of it sticks. And I think a lot of it really comes down to not the tactics of management so much, but of the application of good character. Right? You're parenting with consistency and discipline, but understanding that your role is not to be the friend of your children or the vending machine of joy, sweets, and late bedtimes and device time. Right?

**Adrian McDermott at 00:47:36**
So there's some other role that you are forced to play. I think those are skills that you develop as a leader as you go forward. Right? And ultimately, you get to the point where I think one of the things I recognize is that, you know, some of your role is to be decisive. Right?

**Adrian McDermott at 00:47:53**
I like to say, you know, I'm sometimes wrong but never in doubt. And I think that's sort of an important characteristic in leadership is to take a position and be okay being wrong.

**Joel Beasley at 00:48:10**
I like that. Do you think that being a parent helps you become a better leader?

**Adrian McDermott at 00:48:18**
I'm sure it does. I was going to say that, you know, being a manager and a leader helps, I think it does help me anyway, be a better parent in some ways. I'd seen all kinds of behavior before I started breeding and seeing these independent humans, you know, who were motivated and doing their own thing and, you know, occasionally in frustrating ways, and understanding that they had, they were not non-playable characters in my life, but they had autonomy and goals. That is, you know, kind of a really important way to learn, I think.

**Joel Beasley at 00:48:57**
Yeah. I've always thought of the kids as, like, if I were in the eighties and I could walk into the computers that were the size of a room, and I could see all the little components but much larger. Yeah. That's what I feel like the kids are with their emotions, their thought processing. They're humans, but everything's bigger, so it's kind of easier to see what's going on. You know?

**Adrian McDermott at 00:49:18**
My first internship, I'm old. And my first internship, actually, there was a machine room, you know, kind of raised floor, cold room in the office, and it had actually two mainframes in it. They were British mainframes, so they were colored hot tango and beige, and they were like, you know, made a bunch of noise or whatever. And you'd be like, what's that? And they're like, that's the disk drive. Like, the size of a chest freezer. Yeah. I remember that. But yes, to the amplification, right, I think does that mean that as we grow, we subdue—as we get older, we subdue the sort of honest outlet of those pieces of ourselves that we start to kind of control them and be more staid? Maybe it does.

**Joel Beasley at 00:50:09**
Maybe. I like this. I was reading some of the prep, and you tell people on day one that, quote, "I'll prepare you for the next thing you want to do." I really—that's—I do a lot of these interviews. I haven't heard that one. I really like it. I might steal it and pass it along.

**Adrian McDermott at 00:50:26**
Like, I joined three early startups. Right? And one of the things I realized is that with your early employees, there's kind of a Faustian bargain that you strike with them. Right? From them, you're looking for their labor. Right? And you want it sort of unfettered, unmitigated. Right? We didn't say nine-nine-six back in bubble one of the Internet, but, you know, we were sort of looking for something similar. That was long before I was at, you know, ten-three-four, which is kind of where I'm at now, but we didn't say that.

**Adrian McDermott at 00:51:03**
And so to get that kind of loyalty and contribution from someone, I think part of what their motivation, what they were looking for is, you know, I might not get to that next role at this company. Because if this company is successful, I'm going to be too junior to be the engineering director at this company. Right? I'm going to have to go down a rung and build something for myself to be that engineering director. And then, ultimately, I think a lot of people in leadership and products and engineering do have that ambition to be the, you know, the head of product or the head of engineering or the head of both.

**Adrian McDermott at 00:51:41**
And so that thing is, like, you know, you might get off the train at some point. My job is to make sure when you get off the train, when you get off the Zendesk train, you're ready to sign up, drive the next train, and you've seen a good movie, and you know what it looks like, you know what good looks like, and you're going to take some things or many things from it. Some things you'll change because it's your preference. But it's my job to get you ready for that moment and to thank you for your service. Right? Whether it was, you know, one, two, three years, you helped—you know, we were going through a period of growth or not, but we were building something together.

**Adrian McDermott at 00:52:20**
And I thank you for your contribution. I hope I got you ready for the next spot, and the door is always open. Right? That's sort of the ideal, I think, employee relationship.

**Joel Beasley at 00:52:30**
And when you are interviewing people, bringing them on to a team, under a leadership position, what's one of the traits that stands out to you most? What are you really looking for? You can't look for everything. You can't look for a hundred things. You have to really boil it down to a couple.

**Adrian McDermott at 00:52:47**
Yeah. I think it's changed over time. But I think at the moment, maybe if I was going to qualify it, for me, it's evidence of grit. Right? Like, the ability to have sort of grit and application, moxie, call it what you will.

**Adrian McDermott at 00:53:05**
That is in some ways the defining characteristic of, I think, really successful people in startups and in companies. Right? Like, you know, the sort of—people talk now about mission-driven tactics. Right? I don't care how you get to the point that we ship this software. Someone's—and I do care. You know, I shouldn't say that, but you don't necessarily want every roadblock to be a problem. You want someone who's going to understand that the mission is the mission, and I keep going, and I have the grit and determination to do it, and I can bring a team along. And if it's not going to happen, I can sound the alarm, and you could come and help me and kind of do it in that way. I think grit drives that.

**Adrian McDermott at 00:53:50**
And that determination and self-reliance and agency is something that I'll take in many cases over sort of raw resume applicability or raw academic applicability of a candidate.

**Joel Beasley at 00:54:10**
I like that, evidence of grit. And I like that you put the asterisk on there about caring more about the outcome than being the hero. Because you can get grit with people who want to be the hero, and you can get grit with people who want the outcome to be achieved. And that's the difference between the one that will sound the alarm and won't. Right?

**Adrian McDermott at 00:54:28**
Yeah. That's true.

**Joel Beasley at 00:54:30**
Because I made this mistake. That's how I know about it. So life is long. Okay. Well, hey.

**Joel Beasley at 00:54:41**
Closing, I'm going to ask for one last piece of leadership. This has been a heavy leadership—you're very—I really enjoy getting to talk with you. Actually, you got a lot of experience. I want one piece of leadership advice, and I'm going to give you some constraints for it. Is that okay?

**Adrian McDermott at 00:54:55**
Yeah.

**Joel Beasley at 00:54:55**
The constraints are someone else told you this. They say, "Hey, Adrian. You need to think about this or do this." You did it. It proved to be effective, and you've kept it in your repertoire for some time now. So tested, true advice that you've actually been using.

**Adrian McDermott at 00:55:14**
I'm going to, you know, I'm going to go back to the first sort of real job I ever had. I was doing software with this company in London, and they transferred me to the Hong Kong office. And this guy met me off the plane. I'd barely left England before. I was probably, like, 23, 24, you know, child, basically.

**Adrian McDermott at 00:55:41**
And we were—he was giving me advice as this grizzled old expat, you know, kind of software manager or country manager he was, I think. And he leans over and he's like, "Adrian," he's like, "in life," he's like, "especially in Asia," he's like, "people actually do stuff." You know? They're like, "Hey. Come over here and do this," or "You should work on this," or "You should do this." He's like, "It rarely hurts you to say yes." He's like, "You will come up with reasons why you shouldn't do that. You will be risk averse or you will kind of, you know, have all these structures in your mind or, you know, in your emotional response that lead you to say, 'No. I don't really want to do that.'"

**Adrian McDermott at 00:56:29**
And he's like, "But do it." He's like, "You know, go through the door, cross that chasm, do whatever it is, or just say yes. You know, take the lunch. And good will come of it. If you're open to those things, the universe will provide."

**Adrian McDermott at 00:56:45**
The experiences will be there, and they'll be good. And I think keeping—I've kept—that was thirty years ago or more. And I've kept, I do think back to that advice in many, many ways where it's how I ended up, you know, working for a startup in San Francisco. It's how I had so many of the other kind of great experiences is someone's going to be like, "Hey. Why don't you do this?" You'd be like, "Alright. I'll try it."

**Joel Beasley at 00:57:15**
That's how you ended up with two kids. I'm sorry. I had to. It was right there.

**Adrian McDermott at 00:57:21**
It was right there. Yeah. Too bad, actually. Said the man with three kids, but whatever. Yeah.

**Joel Beasley at 00:57:28**
Oh my God. Adrian, you're a lot of fun to hang out with, man. You must have a real good team over there. They must enjoy working with you quite a bit.

**Adrian McDermott at 00:57:36**
I think they're secretly relieved when I go off on that one.

**Joel Beasley at 00:57:41**
That's how it should be. Alright.

**Adrian McDermott at 00:57:43**
I think so.

**Joel Beasley at 00:57:44**
I think so. I think so. Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.