AI Is Already Resolving 90% of Customer Service Tickets - and It's Getting Smarter | Shashi Upadhyay, Zendesk

Zendesk went private two weeks before ChatGPT launched, and the moment it came out, it was obvious that customer service would never be the same again. Shashi Upadhyay, head of product, engineering, and AI at Zendesk, joins Craig Smith to explain what the company has built since: a self-improving AI system that doesn't just resolve tickets but learns from every failure, studies what the human did to fix it, and gets measurably better over time. He calls it the resolution learning loop, and for Zendesk's best customers, it's already resolving 70 to 90% of incoming tickets autonomously, up from the 10 to 20% that chatbots managed just a few years ago. The conversation goes deep on the engineering decisions that actually matter: why hallucination is a feature, not a bug, and why the real challenge is knowing exactly when to switch from creative AI to deterministic code; why Zendesk acquired Forethought and what made their approach to going live in days rather than months so valuable; and why, despite all the momentum, Upadhyay estimates we are only about 5% through the adoption of AI in customer service. The bottleneck isn't the technology, it's the change management required to restructure how human and AI workforces operate together. His vision of the end state is striking: personal AI agents talking directly to enterprise AI agents, resolving 90% of issues instantly, while humans focus exclusively on the complex, high-value interactions that genuinely require them. Subscribe to Eye on A. I. for weekly conversations with the people building and deploying the future of AI.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Craig S. Smith?

  From the public's point of view, this can't happen fast enough. BPO was a revolution, like this drudgery that you can outsource to lower cost markets with higher unemployment. But now is that kind of come back to the enterprise because you don't need those big call centers any longer.

• 0:19Speaker 01

  I really do think that there's a golden age of service ahead of us. I'm very hopeful that will happen. Ninety percent of problems solved by AI and solved instantaneously, anytime of the day. The pace of technology is far, far exceeding the pace of adoption. And the pace of adoption is gated so much more by people working together, how companies run their processes, just how the purchasing cycle works, how the implementation cycle works. Those are like laws of physics, they don't improve by a factor of ten, never.

• 0:48Craig S. Smith?

  Can we start by having you introduce yourself?

• 0:53Speaker 01

  So I'm Shashi Upadhyay. I'm the, I run product and engineering, including AI at Zendesk. Been here for 15 months. Prior to this, I was a part of the Google Ads organization, the Ads Product Organization. I was running all the enterprise ads products there for several years. And prior to that, I started a company called Lattice Engines, which was a pioneer in using machine learning and AI for CRM use cases. So marketing and sales primarily. And that company had been acquired within Broad Street and then taken public in 2019. So I've spent like last 20 years of my career on machine learning and AI

• 1:38Speaker 01

  as applied to CRM problems, different stages of the customer journey. And before that, I was trained as a physicist and I've been a management consultant for a few years.

• 1:48Craig S. Smith?

  Let's talk about the acquisition. You guys, first of all, I think most. Digital knowledge workers who are likely my audience know who Zendesk is. And you've recently acquired an agentic platform or agentic AI company called Forthought. And you've talked about how Forthought brings self-improving AI into the Zendesk portfolio. I'm most interested in what you mean by self-improving, but before we get to that,

• 2:30Craig S. Smith?

  explain what Forthought does, why Zendesk acquired it, how it fits into Zendesk's offerings.

• 2:40Speaker 01

  Yeah, so let's start from like, I'll kind of remind ourselves, situate ourselves a little bit in what's happening in the world of service, right? So Zendesk was founded, you know, almost starting seventeen years ago. Originally as a software as a service company focusing on the customer service problem. And the big innovation that Zendesk made at the time was mainly around usability. So like, how do you, how do you sort of launch a system very easily that's beautiful? You know, very pleasing to the eyes and, and a place where agents, human agents would want to spend their whole day in, right?

• 3:23Speaker 01

  Hence the name Zendesk and, and it built on that or a very successful run through a public and through COVID. We went private in 2022. And incidentally, and this is why it's important, all this buildup is important. Incidentally, it was two weeks before ChatGPT came out. So.

• 3:49Speaker 01

  Essentially, the company went private and two weeks later, ChatGPT came out. And the moment it came out, it was obvious that the world of service has basically turned over. It's never going to be the same again. You know, it was not going to be. It's almost like a new, a new, um, a new form of labor, right? Had entered the market at that time, right? Because once you can have conversations, you can at least start to answer questions and do some simple things. And that was kind of the state of the business for a while, which is, uh, companies started deploying these. They were still called chatbots at the time, but they were really more like, you know, chat search, uh, systems, right?

• 4:31Speaker 01

  They allowed the end consumer to get answers to questions like, you know, my router is broken. How do I fix it? And these chatbots were pretty good at figuring out how to present that information to you. You know, basically using rags and, and, and a, you know, a bunch of documents you could throw into a rag and then use that pipeline to answer questions. Then in late 2024, when reasoning models started to come out with the O series, it became clear that now these, um, uh, these, these models, essentially these could, could not just answer questions, but could actually think on the fly and therefore take action, at least in the digital world.

• 5:13Speaker 01

  And that was, you know, further enabled by MCP and A two A sort of standards emerging. So that's when the, uh, the, the kind of attention shifted away from simple Q and A to agentic products. And agent, the agent is by definition, it has agency, which means it can do things. Yep. And I think at that point, it really started to mimic like what human labor can do. And the way we, um, uh, we kind of, uh, understood that change was, and the way customers think about it is, they look at a number called automation rates. Automation rates like of, you know, if a hundred queries came in, how many of them can be resolved by, um, by an AI of, of, by AI agent?

• 5:59Speaker 01

  And that number had been in the. Ten to twenty percent for a long time. And the moment true agents came in, which is they could reason and they could take action. That number jumped up to about fifty percent. And then for our best customers today, when we implement, we see between seventy and ninety percent automation rates pretty regularly. Hello. So with seventy or ninety percent automation rate, basically the work, if you think about the workload that a service center has to go through. It has shifted from being very human centric, which is where it has started, to actually being very agent centric. Right? Because agents are doing most of the work.

• 6:39Speaker 01

  Now it turns out that, look, most of that work is still, um, in some sense, I would call it low value work relative to high value work. And I'll define what that means. So we looked at across our, all our customers, we looked at about a billion tickets and said, okay. How many of these are like truly, truly value add as in like you did something that was helpful to the customer? And were able to get them to a better place than they were before versus something the customer could have done themselves. E. G. Like just searched on the web or found a piece of information, but they're still calling a, you know, calling a call center to get those questions answered.

• 7:20Speaker 01

  Right? Or something, um, uh, maybe it's something minor broke and they could have, you know, figured out the fix themselves. It turns out that across a very wide range of industry, the answer is about seventy thirty. So seventy percent is low value, thirty percent is high value. So A I agents are going to take over that seventy percent. That's a given. And the next frontier, of course, is the next thirty percent because the more A I agents can take over those. It frees up the human workers to work on a lot of unserved needs, right? That are out there even today. Like the classic example is it's very rare that when you call a call center, you can get some alive, right? Yeah.

• 8:01Speaker 01

  But if if the if the AI agents can take over that strategy, then we can free up humans to take those live calls and then give customers very delightful experiences. You know, they become more loyal, they become happier, and the whole system works better. So that's what we've been working towards. And we have done seven acquisitions because, you know, Zendesk started as a SaaS company. We recognize that we didn't always have the talent in-house. We didn't always have and and, you know, startups can often move a lot faster than than than we we had been able to, certainly last several years. And where we found good, good teams, good talent and good technology, we have gone out and

• 8:46Speaker 01

  and bought them. So for our agentic journey itself, we had purchased a company called Ultimate, which was based out of Germany. And that gave us the foundation for the first, the the first set of AI agents. We bought a company called Klaus, which was based out of Estonia. And that is an agent that actually monitors other agents. So it sort of keeps an eye on how well is the AI agent actually doing. It's like an independent referee. And then recently, we bought Forthought because Forthought had been really innovating on a number of dimensions. And one of them is what we call the resolution learning loop or self-improving agents, which I'm happy to talk about more.

• 9:27Craig S. Smith?

  Yeah, well, let's talk about that. You say that they've been focused on reinforcement learning?

• 9:33Speaker 01

  They've been focused on innovating in the market with with a number of different things. For example, they were the first to introduce computer use in in in the service space. So a computer uses the the basic idea that like, there are many cases where you don't have APIs or access to systems, right? So an agent has to be able to take action on different systems. But often, you know, they have to go take action in a purchasing system or a return system that is not it with there's no API exposed, right? So you basically have to log into the screen and read what's on the screen, identify what's a button, be able to take action on that.

• 10:15Speaker 01

  So in the AI agent space, Forthought was the first one to to to build a computer use system for for for the service use case, right? My point overall was they've been very innovative, usually ahead of the rest of the market in terms of coming out with new ideas. And one of the things that we Zendesk have been working on is is improving the learning loop. So we call it the resolution learning loop, right?

• 10:47Speaker 01

  Ourselves and in parallel, Forthought had made a number of innovations around that too, and we felt combining forces could accelerate us down that path.

• 10:58Craig S. Smith?

  What is resolution learning?

• 11:01Speaker 01

  So resolution learning loop is a it encapsulates a number of ideas. So let me kind of talk through a few of them. So one great thing about services, services problem solving, right? Like, somebody comes to you with a problem, you either solve it or you don't solve it, okay? And if you solve it, once in a while, customers give us feedback as in, yes, you solved it for me or you didn't solve it. Not always, like many times they don't respond, but you have that signal, like the one zero signal. We also get signals in other ways. So you can you can actually see the trace of a problem as it flows through the system because first

• 11:45Speaker 01

  the autonomous agent tries to solve the problem. So we have the entire conversation log. And then it hands it over. Let's say it's not able to, if it's able to solve it, great, you know, you solved it. If you're not able to solve it, it hands it over to a human. But now the human is doing a bunch of things and there's a trace of all of that. What have they done? What actions did they take? So all that information is fodder for getting the system to do better the next time, right? So let's just take the case of the AI agent itself. So the AI agent can operate in multiple modes. If it solves a problem, that's a great signal. This is what worked, okay?

• 12:26Speaker 01

  Because this worked, I can do what I took and I can automate it. So the next time somebody asks the same question, I don't have to go through the whole thinking process again. I can just automate it, it becomes automation. So the next time somebody asks a question, I just, I just hit the automation and boom, the problem is solved, right? So like, the system learns that way. The second way, it can go back and in cases where it failed to solve the problem, it can go and study, right, in a different run. What did the human actually do? Like, what happened since then to solve the problem? Right? And then we are able to actually run simulations on top. So,

• 13:06Speaker 01

  like, you throw a bunch of different problems at the agent and see if it's now able to solve those problems or not. So that becomes the second thing. And then the third thing is, we actually have a, in addition to a, uh, the AI agent, autonomous AI agent, we also have a co-pilot agent. So every human agent gets a co-pilot. Right? And if the, the human agent will interact with the co-pilot, solve the problem together, the co-pilot can observe that, hey, like, this is what worked. And then push it back to next time there's the same problem comes up, the automation agent can run with it.

• 13:47Speaker 01

  So a lot of this is just closing the loop because we, we were a system of record, right? That's what SaaS companies grew up as. Because we're a system of record, we have the trace of everything, starting from when the customer came in until the problem was solved. And that trace, all those, that, that event trace basically becomes the ground truths, the, the fundamental information, or, I mean, the term that's been used today is context graphs. The context graph on top of which these agents can learn after the fact. So the next time they encounter the problem, they're going to do better. Right? So that's the idea of resolution learning loop. The last thing I'd say is oddly like one more we found, which has been very interesting, is AI agents are great at complaining. So like, if they're not able to solve a problem,

• 14:32Speaker 01

  they'll tell you, hey, I was not able to solve a problem because I couldn't get access to this data. So it's starting to surface, like, for example, you may have, you know, information about a product in two different places that are conflicting with each other. Maybe there have been, you know, there's some version control problem or something like that, right? And it's like, hey, I couldn't solve the problem that I got confused between these two, right? Like, like, I saw this conflicting information, I got stuck. So that can become an alert to a human who is managing a knowledge base to go say, okay, let's go resolve this conflict, let's have a common set of facts. And so that widens up the improvement loop for the entire organization, not just for the AI agent.

• 15:14Craig S. Smith?

  Yeah, yeah, that's interesting. And so Forethought has has built this resolution learning loop. And you guys acquired it or or I mean, did they build the agents from scratch in this system? Or are they using open source or off the shelf agents? And that's one question. The other is, what are the underlying models that you're using? Because certainly, you know, hallucination continues to be an issue.

• 15:52Speaker 01

  Yes.

• 15:52Craig S. Smith?

  And and, yeah, you you you need to be careful in customer service that's that the underlying model is reliable.

• 16:04Speaker 01

  Yeah, so coming back to Forethought for a second, right? Forethought was actually one of the first AI agents company. They started even before the chat GPT came out, right? So they're they're, I think, like founded in twenty sixteen or twenty eighteen. And they had they had won won the TechCrunch Award the year they were introduced because of all the promise that they held.

• 16:29Speaker 01

  Their approach, one of the things we really liked about their approach was

• 16:36Speaker 01

  they really mined whatever information was available at the customer very deeply. So, for example,

• 16:44Speaker 01

  when they connected, when they landed at a customer and let's say connected in a Zendesk instance, they would go and read the last year's worth of tickets automatically and classify them and come up with suggested procedures, basically. Procedure is like how you solve the problem, right? It's like step-by-step procedure. So the auto-created, the beauty of auto-creating it is that you can go live really fast, right? So instead of the customer going in and entering a bunch of prompts and saying, you know, when somebody asks for a payment, do this, then do that, then do, you know, like, it doesn't do any, it just goes and reads a bunch of tickets and says, what did the human do? And then auto-creates a bunch of procedures based on that.

• 17:28Speaker 01

  And that allowed them to go live very fast, right? So, so often these implementations, when you look at the AI agent world, right, like often these implementations can take months, right? They take months. And that's why when customers talk about, oh, it took forever for us to implement, that's what they're talking about because it took them months. You know, people don't often, like a lot of these things run very organically. They don't have their own processes documented anywhere, right? But they sort of exist in the tickets. They exist in because there's a his, it's a historical record of what actually happened. So that was like their, a pretty major innovation that attracted us to them because they could go live very fast. And they were often taking advantage of data that we had better than we were doing it ourselves, right?

• 18:14Speaker 01

  So, so that was, that was very, very powerful. On the resolution learning loop itself, it's really a combination. Zendesk has a number of approaches in place and they're bringing new approaches. I think the key thing is there's, you know, while we call it the resolution learning loop, it's not one loop. It's actually multiple loops and there's loops within loops. And their approach, which was, I'll just talk a little bit about it, is it was slightly different from, it was complementary, meaning the way they have approached it is to say, we'll auto-generate the procedures. And then we'll tweak the procedures as we go and see if it improves the results or not. And I will do it as an A B test, right?

• 18:59Speaker 01

  And we'll do it as an A B test and see if like a small tweak to a procedure at least to a better result. And if at least a better result, then I'm going to just switch to that. It's a, it's a, it's a slightly different approach than the, some of the others that we've been talking about, right? Just like combination of these. And I really believe, look, the, the, the true answer is that over time, every, we have to think about AI agents, the way we think about, you know, humans, right? You put a very smart person, you bring them into a new job. They are smart, but they lack context. There is no, you know, they have to be trained, right? They have to be washed over. And then over time, if they have the self-improving mindset, then they get better and better.

• 19:44Speaker 01

  And that's exactly what we're trying to emulate with this.

• 19:48Craig S. Smith?

  Yeah. And, and the automations, you said they build automations that then the next time they encounter that problem, they just kick off the automation. Is that an agentic automation? Is that something more basic? Is it all built into the system? Or is it, I don't know, using Zapier or, or an external service like that to build the automation?

• 20:22Speaker 01

  Yeah. So, so first of all, you've, you've hit the nail on the head, right? The, the, the, the biggest challenge with AI agents has all along been that they are inherently unpredictable systems, if not built correctly, right? I mean, L L M's are unpredictable, right? I mean, that's like probabilistic by very nature. So hallucination is, is, is a, is not a bug. It's a feature really, right? That's why they're so good at creative tasks, for example.

• 20:56Speaker 01

  We have to, in the enterprise setting, find ways to rein that behavior in, right? Because enterprises are on the other end of things. They actually don't want creativity. They want standardization. They want consistent behavior. They want, you know, consistency with the brand of the company. You have to use the same tone of voice. You have to use the same way of talking, right? That's what, that's what, that's what they're going for. So the number of ways to come at this, so coming back to your automation question, the best AI agent is one which does a great job of understanding intent. Like, what are you trying to do, right? Like, do you have, are you, are you lacking some information that I need to get to you? Or do you have a problem that you need solved?

• 21:41Speaker 01

  And if you have a problem, what problem is it? Is it that you're trying to return a product? You want a refund? You want a discount the next time? And they're very good at going back and forth conversationally to figure it out, right? But the moment I know you are in one of those branches, the moment I know you just want your product returned, right? I don't, I don't, we don't want the, the, the creative side of those agents to express themselves anymore. Because the return process is a very well standardized process at a company, right?

• 22:15Craig S. Smith?

  Yeah.

• 22:15Speaker 01

  It should follow a certain path. It should go through a certain set of steps and it should happen reliably, right? Hundred percent reliability. So to your question about how do we implement it? The answer is we support both. Zendesk has something called an action builder, which allows us to touch other systems and take action on them. But if a customer decides that they want to use their own, like you mentioned Zapier or something like that, they can also just integrate with it and we can take actions in those places. So we want to, we, what, we like working with customers wherever they are. If they have a, you know, a repository of these automations sitting somewhere and they just want to trigger it, great. If they wanted to build it on our system, we'll support that too.

• 22:57Craig S. Smith?

  That's interesting. Uh, so, um.

• 23:02Craig S. Smith?

  That the the underlying L L Ms are are primarily OpenAI or Claude, is that right? But but then the automations they build, they're they're not calling an external service, they're they're building a deterministic automation.

• 23:26Speaker 01

  Yeah, absolutely. So the for the modeling part, the L L M part, right? Uh, we use, uh, we use OpenAI, uh, we use Anthropic, uh, for some use cases. Uh, we also use Google Gemini for some use cases. Um, and then for certain, and and especially where where performance matters, like you have to get the um, the system to behave very well consistently, we use the, you know, frontier models.

• 23:54Speaker 01

  For cases where um, we have like a large volume of stuff going through. Like we have, we have part of our generative search where you can go on a knowledge base and just, you know, like, talk to it instead of like typing cryptic keywords. Um, uh, we run those on open source models that we have fine tuned ourselves. And our general strategy here is multi-model. And we will also provide our customers, over time, what we see happening is customers may just say, I only want to work with, you know, this particular model. And we want to provide them that flexibility on whichever model they want to use on our system.

• 24:33Craig S. Smith?

  Yeah. And and is there a human checking these automations that the agentic AI is building? Because it the the agentic AI could be learning from bad habits or or amplifying errors.

• 24:50Speaker 01

  Yeah. So for for the automations themselves, um, you you point to a very interesting trend in the industry, which is, you know, like, uh, with the amount of auto coding that agents can do, they could kind of write an automation on the fly, right? Yeah. Just code it. So this is what I need to do. Um, especially if you Provided the context of what systems are available and like they can absolutely do that.

• 25:17Speaker 01

  The the way around it is to have a a verification step on all of these things, right? And the verification step is it can be another AI agent which have of course has its, you know, own issues, right? Because I mean, a verification agent can also hallucinate and and and behave in unpredictable ways. But today we approach, like if a customer tells us that a certain step is deterministic, we approach it with deterministic code. If you have very large customers that have, you know, literally like billions of dollars flowing through the system, right? Sometimes hundreds of thousands of customers, sometimes millions of customers.

• 25:58Speaker 01

  What's deterministic should stay deterministic, right? And so some point in the system, whether the automations themselves are deterministic or the verification is deterministic, there is some place where it has to be caught, right? And and we believe that that separation will stay, right? Meaning, you know, it's just like, often it helps to think about these things from the perspective of they are just a different form of labor, they are just like people almost by it, right? And like everything you put in place, there's place for creativity, there's a place for, you know, the the difference that we all have as human beings. And then there's a place where everything has to work exactly like a machine.

• 26:38Speaker 01

  And everything that has to work exactly like a machine is all the old principles of software development. You know, build good code, verify it a thousand ways, make sure it works under every situation, every edge case is studied, all those practices. They're not going to go away, right? They're not going to go away.

• 26:55Craig S. Smith?

  Yeah, well, that's fascinating. And and how long have you had this resolution learning loop in production?

• 27:07Speaker 01

  We, if you think about it as like, again, like many loops, right?

• 27:12Craig S. Smith?

  Yeah.

• 27:13Speaker 01

  We have had the most basic version of it in production for at least a year. And, and, and this is why, like, when our customers go live with us, they'll often go live with us as a, you know, say 50% resolution rate. And then it starts to creep up as we see more and more interactions, right? Starts to creep up, we generate more procedures, those procedures do better. We see what procedures don't work, those procedures get retired. So that version has existed for, for a while. But what we want to push this to, like, some of the other examples that you, that we mentioned, where you actually like going and studying every single ticket and what happened to that interaction. That is stuff we're building at the moment, right? That's not in market today, but it will be in market later this year.

• 28:00Craig S. Smith?

  And you say the resolution rates are creeping up. Can you talk about what the resolution rates are now? What the automation rates are? And how you see that by the end of the year?

• 28:14Speaker 01

  Yeah, so, by the way, I use, I'm using those terms sort of interchangeably, right? So just, just let's, for purposes of this discussion, let's use a very simple definition, which is an automation rate or resolution rate is number of issues that were resolved by the AI, divided by the total number of issues that came through the door, right? So it's that simple. And the number that we feel very confident with today is if you look across all our customers that have gone live and have had some time to kind of climb up that curve, that number is 70 to 90% today, right? 70 to 90% today.

• 28:57Speaker 01

  Now, there is absolutely some variance, right? So if a customer, just take the example of an e-commerce customer or a highly transactional customer, where most of the times people come with very simple problems, right? It's like, I want to return the product, or I want to get a refund, or like those kinds of things. You can go into like 80s very quickly, sometimes into the 90s, right?

• 29:21Speaker 01

  If a customer has a more complex setup, right? Like, let's say you're a business and you're reaching out to another business and you're having some problem with the teleconferencing system you bought, like the situation we were in earlier today, right? And you're trying to get something like that resolved. It's not a one turn thing. It's a multi-turn, you know, and the conversation itself can stretch out over many days. There, the resolution rates can be as low as thirty to forty percent, right? Like, you can solve the easy problems, but the hard ones then now need human involvement. So, uh, the, the consumer examples are showing us what the art of the possible is. But there's a lot of work to be done on taking the more complex examples of the same journey, right?

• 30:06Speaker 01

  Up to similar numbers. Yeah.

• 30:08Craig S. Smith?

  Yeah. Does Zendesk have its own call centers? Or are your products sold to enterprises who then contract a call center, a BPO, to handle the live calls?

• 30:28Speaker 01

  Yeah, we have historically only sold software, uh, so we don't, we don't run call center ourselves. Either enterprises run our software themselves or, uh, or they use a BPO who's also, you know, could be running our software. That's right. Yeah.

• 30:43Craig S. Smith?

  Yeah, and is your, your largest market with the BPOs or, or, uh, with enterprises?

• 30:52Speaker 01

  No, our largest market is with enterprises. So, um, I kind of going back our history, right? The ways Zendesk got traction was by starting small, so not going after the largest businesses, but starting with small businesses and make it very easy, like very typical software as a service. You know, there's a webpage, you log in, you immediately experience the product, you get your first ticket and now you're hooked. And then like as those companies grew, in fact, uh, historically, our biggest customers were often startups, right? As they grew, they became much bigger and we became bigger and we were able to support enterprises as a result. So, you know, today we support enterprises, but we grew into it.

• 31:33Speaker 01

  So most of our customers are still like just buying direct from us and using it for the internal teams. Um. BPOs are an important part of our ecosystem, but they're they're they kind of came later, so to speak. Yeah.

• 31:45Craig S. Smith?

  Yeah, presumably if if um in on the average, if you're solving eighty or ninety percent of the uh of the issues that that uh come in, um you need fewer and fewer uh people in a call center. Uh and and you're saying then those people move on to a higher level, more complex tasks. But do you see the day when call centers will be obsolete and you can have and enterprise can have an a small in-house team to handle uh these more complex tasks, which frankly, to me, sounds

• 32:29Craig S. Smith?

  better because they're sitting in the enterprise. Uh they they know the enterprise. They know the people in the enterprise. I mean, it's just um uh we've all had experiences with call center agents who clearly uh don't know the company that they're answering for and and it gets very frustrating very quickly. So it just seems like BPO was was a revolution, like this drudgery that you can outsource to lower-cost markets with higher unemployment. But now, uh is that kind of

• 33:10Craig S. Smith?

  come back to the enterprise because you don't need those big call centers any longer? And and I'm not saying right now, but if this uh agentic uh resolution loop uh keeps improving.

• 33:27Speaker 01

  Yeah, look, uh the the the general trend in the market, right? If you just stuff, but you you said something very interesting, right? Which is um like the the the service experiences are often painful, right? They're not fun, right? And it's in, you know, no, no, no one like wakes up in the morning and says, I'm really looking forward to today because, you know, I'm I'm gonna I'm gonna be making a bunch of calls to call centers, right? Right? People do wake up in the morning and say, I'm excited, I'm going shopping, right? But no one wakes up in the morning and says, I'm gonna go like have a service experience today. So there is something broken about the system today, right? Um, and what's broken is that exactly what you said, right?

• 34:11Speaker 01

  You you you uh uh on the one hand, you have consumers who have increasingly less and less control over the service experience they're gonna have. Um, you know, they have to go through email and this interface and that interface and like can't talk to a live person, which is what they want to do. And uh the the other end of things, by the way, is also frustrating. It's very, being a call center employee is a very tough job, right? Because you're just like, all day long, you're listening to upset people who've been waiting for you for ten, twenty minutes. Right? And just take that job and extend it to very often these people work on, you know, Christmas and New Year's and all those days, right? Like, it's just not a great job, right, to start with.

• 34:54Speaker 01

  So our mission as a company is to help take that that negativity out of the system, right? So that when a consumer comes in, they just have a great experience. They they're able to have a great experience and they don't have a great experience. Uh, I mean, they'll have a great experience with the AI agent because they're programmable, right? So we'll we'll make sure they do. But if they have to be handed over to a person, then that human being is fully equipped to solve the problem because they have a co-pilot. And is available more or less, you know, in in like much shorter period of time, like ideally live over time, right? That's like, that's the goal. So what, kind of coming back to your question, what I see happening is because services.

• 35:40Speaker 01

  Under invested in by most companies. The first thing we are actually seeing is customers are using AI agents. Not as a, hi, how do I shrink my workforce or like, actually, most of them are using this as. I had these unserved pockets of customers that I always felt very uncomfortable about. Like I have a free product, but I never provide a service for that. Let me go cover that, for example, right? Or my service has always been like, you know, nine AM to five PM. And, and of course, there's no service on holidays and like, let's, let's go cover that. Let's go get the weekends covered. Let's get like all of that covered. I think that is gonna be the first big, so all of that stuff I'm telling you about 90%, etc.

• 36:25Speaker 01

  All of that is going into that at the moment, right? It's like, let's go improve our service quality. Let's get our C set up. That's where the attention is. Over time, on the jobs question, like I'm a firm believer that there'll just be new, but different jobs. Right? So think, so one, one, one example of this is when you go from, uh, like, you know, doing everything by hand to like 90% of the work is now being done by AI. Then there, there are jobs to actually monitor the AI. There are jobs to train the AI. There are jobs to keep it, keep it from like going off the rails, right? Improving it, managing compliance with it, etc.

• 37:07Speaker 01

  And that requires the ability both to actually have done the service job. Cause the best people will know what it's actually like to be in a service role. But then also to be able to work with AI on helping the AI improve, right? So we just, I think the shape of these jobs will definitely change. But what's exciting here is like if it's possible to do both, which is automate more.

• 37:35Speaker 01

  And improve CSAT, the customer experience at the same time. Take away those negative experiences and then put a lot more investment in much more live experiences with much more knowledgeable people. That's I think the exciting future ahead here.

• 37:50Craig S. Smith?

  Yeah, and you keep using the acronym CSAT. What is CSAT?

• 37:54Speaker 01

  Oh, sorry, customer satisfaction. I'm sorry. It's just that it's a score. It's a score. So customer satisfaction, you know, we we tell like at the end of every interaction, you give them a little survey and say, how do we do, right? And they give you a rating. So it's it's a very commonly used metric to see if the customer had a good experience or not. So.

• 38:14Craig S. Smith?

  Right. And that's how you draw the line between deflection and genuine resolution because, you know, the ticket gets closed. You don't know until you get feedback from the customer. So you're looking for that signal. You don't get it all the time, but when you get it and it's positive, then you you look at the at the trace of what.

• 38:41Speaker 01

  Correct. I mean, i- i- if if a if the problem is solved, the customer lets you know whether or not they gave you a score that the problem was solved, right? Deflections are often like they include other things, right? Like, you know, you'll be surprised at how much spam these companies have to deal with. You know, they'll get something that looks like a support ticket, but comes through email. And like, you don't want a human like spending you in a microsecond on it. So like, machines are very good at saying, okay, this is not a real issue. Let's just throw it out. Like, that's more of an example of a deflection. Yeah.

• 39:16Craig S. Smith?

  Yeah.

• 39:18Craig S. Smith?

  The you you say that self learning systems are the future of service.

• 39:25Craig S. Smith?

  Who are you watching in this race? Other CRM, C M R platforms, hyperscalers, because

• 39:33Craig S. Smith?

  as you said, when

• 39:36Craig S. Smith?

  You know, when ChatGPT came in, everyone in the service industry saw this is going to change how things are done. And not only companies like Zendesk have hopped on it. I mean,

• 39:51Craig S. Smith?

  a lot of, you know, hyperscale in Microsoft, Google, Peer Play AI startups. There are a lot of people in the space. So,

• 40:01Craig S. Smith?

  yeah, how do you see that shaking out?

• 40:06Speaker 01

  Look, as with any new technology, right?

• 40:11Speaker 01

  New technology expands markets. I think we all know that, right? So with, with, you know, client server to cloud, there's like a ratio of 10 that's often used, which is every, every market that existed in the client server world became at least 10 times bigger in the cloud world. We believe with AI that number is probably even larger, just especially in service where it's so underserved, right? This is why everyone hates service, right? Coming back to that. So there will absolutely be players. I think there's room for many big players, not just like two or three, right? Like many big players.

• 40:44Speaker 01

  We of course have the, you know, the large, the large players that you mentioned, you know, Microsoft, Salesforce, et cetera. They're going to continue and they're absolutely, you know, everyone's working on becoming AI ready.

• 41:00Speaker 01

  Zendesk, I think among the incumbents has made two big bets. One, we moved very fast. We were private, so we could move very fast. We acquired seven companies, started integrating them right away. And bet all in on the, like, we really are in the process of disrupting ourselves. We recognize that C-based model is not here forever. So they've gone to resolution-based pricing, right? So we get paid if the customer's problem is solved. If it's not solved, we don't get paid. And we have moved in that direction faster than anybody else in the, in the kind of larger companies. And then on the On the startup side, you know, the last I counted, there were at least forty different

• 41:42Speaker 01

  companies that had been venture funded and that are going after the space. My expectation is maybe two or three will end up as large companies over time. And you know, some of them are founded by very impressive people doing very impressive things. We pay attention to all of them. We are trying to learn from all of them. I think it's good for the industry to have so many people thinking about how to solve this problem better.

• 42:04Craig S. Smith?

  Yeah, and the Zendesk is extending its Agentic AI to teams using other service platforms. Can you talk about that?

• 42:19Speaker 01

  That's a great, it's a great question, actually. So we had a number of our customers have come to us and said, we really like what we're seeing on your Agentic roadmap. We like your history of innovation. However, for a bunch of reasons, including, you know, enterprise agreements, et cetera, that we have with some of the larger companies. We want to stay with them. It's very impractical to move off of them. But we want to use your AI and not their AI because we like the pace at which you're moving. So one of the things that Forthought brings for us is because there were multi-platforms to begin with,

• 43:02Speaker 01

  they, they help us address the customer, our customers' needs in those situations. Right? And, you know, one of the great things about the SaaS world is it was very open. I mean, customers, like companies often work with each other, even when they competed with each other. We, I believe some version of that is going to continue in the AI world. It'll be slightly different, but it'll continue in that world. And customers are going to want to pick the best solution that works for them. You know, they may have someone else or the marketing or a sales solution. And maybe they want to keep all their data in one place. That's all great, but they may want, they may want their AI agent and analytics associated with service

• 43:47Speaker 01

  from somebody else because some other company is very good at providing that. And that's, we specialize in service, so we want to provide the, the AI agents for service to all customers, not just those that have chosen Zendesk as the ticketing platform.

• 44:01Craig S. Smith?

  Yeah. And the when, you know, these agentic resolution systems, they're not perfect.

• 44:15Craig S. Smith?

  Do you have examples where they made costly mistakes? The agentic layer, and when that happens,

• 44:23Craig S. Smith?

  who, who is responsible, the Zendesk, the Forthought's platform, or the underlying model? I'm just curious.

• 44:36Speaker 01

  Yeah, I mean, Forthought, we have not completed the, you know, the integration, right? So hard to speak, hard to speak on their behalf. In our case, I think one of the things that we got right because of our enterprise history was we put a lot of controls in place where, like, you cannot get our our agents to get to, you know, like, you cannot trick it into talking about politics, for example, which is a very common thing people try, right? They try, they're trying to get your agent to say something controversial, or try to get it to solve a math problem, right?

• 45:18Speaker 01

  And we, we, when we acquired Ultimate, right? They had already been building AI agents for, I think, at that point, four or five years. So, so they had already seen some of these modes at work. When I was at Google, you know, the first set of the

• 45:39Speaker 01

  talking agents that Google rolled out, my team had worked on that. So I've also seen like all the ways in which people try to break that. So we brought a lot of that knowledge into our system. So we have not had any like catastrophic failures of that kind where like, hey, the whole system has to be shut down or unrolled. I think the things that You know, we have had to improve upon along the way are much more like engineering, software engineering oriented. Getting the tone right is like a really, really important thing. We humans are incredibly sensitive to tone and especially when you have an upset customer and now they're talking to an AI agent, they get upset about some small thing

• 46:23Speaker 01

  or the other. Getting the local diction, the local, you know, the way people talk locally, right? Like, I mean, you know, as we all know, right? I mean, they're like, even within France, there are like four different ways people speak French, right? So you have to get it like that stuff, right? So especially with voice AI, that becomes an issue. I think those are, those have been the much bigger issues than the kind of classic stories that we hear in the press. The standard stories here on the press are much easier to control than the way the expectations are shifting for our customers.

• 47:00Craig S. Smith?

  Yeah. And in the standard problems you were talking about, the, was it Air Canada, where an AI system offered a refund and that was not policy or something like that? Those, you, you, you have a handle on that's.

• 47:23Speaker 01

  We have a handle. And by the way, it goes back to early part of our conversation, right? So like, take a refund example. So in a refund example, what you want is, once it's identified as a refund, right? It should, it should just go to a deterministic path that says, hey, the refund is less than $50, give it. If it's more than $50, it needs to go through a human in the loop, right? And then automation takes over and then you know you won't make a mistake, right? So like, this is what I mean. There's like real engineering here. You can't just take a model and slap an interface and say, you are my agent. That does not work, right? That does not work. So building like, building this at enterprise quality is like, it's, it's a real, it's a real challenge. And I think

• 48:07Speaker 01

  that's why knowing the domain, having a lot of experience, having a lot of customers makes a ton of difference.

• 48:13Craig S. Smith?

  Yeah. You know, this is moving very fast on your end.

• 48:21Craig S. Smith?

  The market penetration, I'm sure from, from Zendesk's point of view is, it must be a growing quickly, but from the public's point of view, this can't happen fast enough. And I, I still get, what are they called, ISR menus and, you know, press one for this, press two for that. It just drives me nuts. Or I'm still on hold for ridiculous amounts of time. How long do you think before this has touched everyone where the whole customer service thing has pretty much been solved?

• 49:05Speaker 01

  I think it's very early. I'd say we are maybe five, five percent through it, maybe. At best 10%, but I think it's closer to 5%. And, and you know, you and I are in a part of the world where this stuff is probably the most penetrated, but if you go to, even if you go to like Europe, it's like a long way to go. Of course there are pockets that are lead offers, et cetera, but yeah, I'd, I'd say, I mean, you kind of point to a really important thing here, which is the pace of technology is far, far exceeding the pace of adoption. And the pace of adoption is gated so much more by people working together,

• 49:49Speaker 01

  how companies run their processes, right? Like just the, how the purchasing cycle work, how the implementation cycle works. That is, those are like laws of physics, you know, they don't get, like, they don't improve by a factor of 10 ever. So that gates their adoption. So I'd say maybe we're 5% through. Yeah. Long way to go, a really long way to go. Yeah.

• 50:10Craig S. Smith?

  Yeah. And, and what do you think is the, the barrier? Is it just people have invested in legacy systems? And they're not going to make a change until really their system is hurting them because the market is, has moved beyond that or until enterprise, you know, SaaS contracts run out? Or, or, or they just have developed trust in, in the AI systems to switch over?

• 50:43Speaker 01

  Um, I, I, I think the real thing is that the chain management of associated with these is quite massive, right? So it just goes back to hundred percent of your work today is being done by uh human agents. And there's a new technology that can, let's say, do seventy percent of the work, okay? And the other thirty percent has to move somewhere else.

• 51:09Speaker 01

  You kind of have to go back to the drawing board and say, okay, like, this is what the new architecture looks like. This is what's gonna go here, this is what's gonna go there, this is what these different people are gonna do, right? So that chain management is always a slow-moving thing in an enterprise, right? I mean, I think we have gotten better at it with tools, etc., but chain management is the hard part. And very often, you know, organizations are not, like, when I talk to our customers, the thing they ask for is, show us the way, like, help us through that transition, right?

• 51:43Craig S. Smith?

  Yeah.

• 51:44Speaker 01

  Don't just sell us, don't just like, give us a technology, like, help us through that transition. I think that's a, uh, that's an underserved need, and especially in the uh, the mid-market, you know, SMB, like, I mean, large enterprises have all these big consultants who are running around after them, trying to help them with this. But, but the moment you go below like a billion-dollar company, right, like, that space is extremely underserved.

• 52:10Craig S. Smith?

  Yeah.

• 52:10Speaker 01

  In terms of, and, and that's why uh, we talk a lot about the intelligence and the capability of these models, but there's an art to making them easy to use, quick time to value, you know, like, all those, like, the basic stuff we learned during the SaaS era, I think they still apply. So if I could show you, right, like, hey, give me like a third of your traffic and I'll show you that I can automate seventy percent, it becomes much more likely that you will then say, okay, I get it, now I can slowly move the rest of the traffic to you, right?

• 52:43Craig S. Smith?

  Yeah.

• 52:43Speaker 01

  I think those are, uh, that's a little bit on us as providers to the space to make it easy to adopt. But the chain management, the chain man part is real, I think that's true for all AI, not just for support AI.

• 52:56Craig S. Smith?

  Yeah, and does Zendesk have like a consulting arm that?

• 53:01Craig S. Smith?

  That works with enterprises to implement the solutions or is it a pure platform play, SaaS play?

• 53:10Speaker 01

  No, we do. We actually have a team of consultants who help with the implementation. We also have a partner network. So if they want to work with a partner, that's okay too, which is and we've trained these partners on how to implement AI agents and bring customers to success.

• 53:33Speaker 01

  And then we also have like the roles themselves are changing, right? You've had customer success managers for a long time that used to watch over, watch over these accounts. For our larger accounts, we have technical account managers that help them ensure that they're successful in the journey, in the AI journey.

• 53:51Speaker 01

  This is a change. I mean, our customers want us to be much more involved than they were than we were before, and we are adapting to the change, but it's a real change. Yeah.

• 54:00Craig S. Smith?

  Yeah. Well, that's fascinating. Self-improving or self, what is it? Self-improving resolution AI loop.

• 54:12Craig S. Smith?

  Well, I'm eager to see how you guys develop that.

• 54:18Craig S. Smith?

  Is there anything I didn't cover that you want to say before I sign off?

• 54:23Speaker 01

  I know this has been a great, I mean, I loved your questions. This has been great. I think

• 54:30Speaker 01

  I really do think that there's a golden age of service ahead of us. I'm very hopeful that'll happen.

• 54:38Speaker 01

  Ninety percent of problems solved by AI and solved instantaneously, any time of the day. I actually think that a lot of traffic is going to just move to your AI agent is actually going to do a service for you. So that's even better, right? With with Open Claw and other other like personal agents coming online now. People are going to offload service requirements to agents, which are then going to talk to our agents and get problems solved. I think that's better. Everyone's happier. And for that last 10%, where you do need to talk to people,

• 55:11Speaker 01

  ideally a live experience, a great live experience. And so I'm super excited about the future and we'll let you know how it goes with the resolution learning loop.

• 55:22Craig S. Smith?

  Okay, great. If somebody wants to explore this, where do they go?

• 55:28Speaker 01

  They can start, I mean, our, our website is like a great place to start. We also have a, a series, we call it the internal note. So you can just check that out, Zendesk internal note, where we go into a lot of detail about how our products work and what they do and how to, you know, take advantage of them. So those would be two good places to start.

• 55:49Craig S. Smith?

  Okay, and it's Zendesk.com?

• 55:52Speaker 01

  Yeah, it's Z E N D E S K dot com, yeah.

• 55:56Craig S. Smith?

  Okay. Great. You said at one point that these coding agents can code something up on the spot.

• 56:04Craig S. Smith?

  And then they write these automations.

• 56:08Craig S. Smith?

  Will the day come where every

• 56:12Craig S. Smith?

  problem will, a coding agent will code up an automation on the spot and run it?

• 56:21Craig S. Smith?

  Rather than it learning and then creating an automation offline that then is triggered next time that problem comes up?

• 56:31Speaker 01

  Yeah, I mean, to do that, first of all, yes, that day will absolutely, it's probably already here, right? But again, right, like, can it do it with a hundred percent reliability is always the question. And that's why we pre-code things and then run them, because we know they're going to work. I, I think the, the, the real, the art of enterprise software at least is going to shift into providing context. Right? So coding agents can do anything if you give them the right context. So if you tell them, hey, these are the integrations available. These you have access to, these you don't have access to. And like, then based on that, yeah, absolutely, hundred percent. It's all there. Yeah, all the pieces are there.

• 57:14Craig S. Smith?

  Yeah. Okay, well, that's fascinating.