Zendesk Relate 2026: The Shift From AI Assistants To Autonomous Systems

[00:00:00] - [Speaker 0]
I'm incredibly grateful to the team at Denodo for backing the Tech Talks Network and helping us produce over 60 interviews a month. And if you are looking for better ROI from your lake house, this message is going to be worth hearing because Denodo helps reduce complexity, control costs, and accelerate time to insight. And it does that by connecting all of your data sources in real time. So make your lake house work harder with Denodo. And you can do that by simply visiting denodo.com.

[00:00:37] - [Speaker 0]
What happens when AI agents stop being simple assistants and start behaving like a new form of digital labor? Well, today, I'm at Zendesk Relate in Denver, and I'm gonna be talking to their president of product engineering and AI. My guest is responsible for leading product engineering and AI at Zendesk, and also brings a fascinating background that spans Google Ads, Google Analytics, Performance Max, startup funding, physics, and enterprise AI. Yep. He knows what he's talking about.

[00:01:10] - [Speaker 0]
But today, we will try and go beyond the hype around general purpose bots and ask what businesses actually need from AI in the real world and explain why the future of AI service is likely to be built around coordinated networks for of specialized agents, each designed to handle specific tasks such as billing, collections, returns, refunds, employee service and industry specific workflows. I want to find out more about that and also how Zendesk itself is changing internally, From measuring seats to measuring outcomes and why automation rates, human acceptance of AI suggestions and customer satisfaction, all these are becoming metrics that matter. But it's also a conversation about trust, guardrails, and the human role in a world where agent to agent interactions will soon become normal. So today, my guest will provide a clear view of where enterprise AI is heading, why the pace of change feels unlike previous technology shifts, and why the next chapter of service might be less about software seats and more about managing intelligent systems, systems that are always learning. But enough for me.

[00:02:21] - [Speaker 0]
Let me introduce you to my guest right now. So thank you for joining me here at Zendesk Relate. Can you tell everyone listening a little about who you are and what you do?

[00:02:31] - [Speaker 1]
I'm Shashu Padhyay. I am the leader of product engineering and AI at Zendesk, and I've been here for eighteen months and driving the AI transformation of the company.

[00:02:41] - [Speaker 0]
Fantastic. And you've been incredibly busy today. I've seen your press briefings yesterday and on stage today, and you said something that really stood out to me. You said that the future future is not one general purpose bot trying to do everything, but coordinated networks of specialized agents. So do you think the industry previously has been oversimplifying AI by chasing the idea of a one all knowing assistant rather than designing systems that reflect how businesses actually operate.

[00:03:10] - [Speaker 1]
Yes. Absolutely. I mean, if, you know, the we found that it's best to think of AI agents as a form of labor. So it's it's a form of labor, Works twenty four seven. It knows everything, or it can everything that's know about the new company, works nice weekends, and in the digital realm, operates within very high level of intelligence.

[00:03:30] - [Speaker 1]
Right? But like any labor, any form of labor ultimately becomes specialized. Right? You can you can you can it's very hard to have a single individual good at everything. It's very hard to have a single process that solves all problems.

[00:03:45] - [Speaker 1]
So similarly with AI, what we are learning is that it's much better in a practical sense, not in a sort of theoretical AGI sense, but in a practical sense. It's much better to build systems that do specific tasks very well and then have a coordinating agent that brings it all together when something needs to be performed.

[00:04:06] - [Speaker 0]
And just to bring that to life, anyone that can't attend or didn't see the keynote today, you mentioned five or six different roles for multiple agents. Didn't you? Could you just mention those briefly as well?

[00:04:16] - [Speaker 1]
Yeah. So we we already had a number of agents that that we've been using. So first of all, we have what's called the autonomous AI agent that attempts to solve any customer problem or employee problem by itself. But then we also have a copilot that that helps individual agents human agents do their job. That also is actually an agent.

[00:04:38] - [Speaker 1]
It's but his his goal is not to solve the problem, it's to help the human. And then we introduced a number of other agents like the admin copilot and the knowledge copilot and the analyst copilot. Right? So we introduced a bunch of these things. But when we talk about specialized agents, we're talking about something slightly different.

[00:04:56] - [Speaker 1]
So we're talking about agents that can take one specific function in the organization and do it extremely well. So, for example, a billing agent would be very good at the billing task. It would understand all the policies that are involved with billing a customer. It would understand all the edge cases. It would understand all the nuance behind it.

[00:05:16] - [Speaker 1]
Similarly, a collections agent would have a very different role. A collections agent's job would resemble that of a human collections person, which is reaching out to people who have not paid and trying to get them to pay within the kind of policies, the brand, the way the the company wants to approach the customer, and then carrying through that process, ensuring that it happens. And if it doesn't, it's not able to do the job, then to escalate to a human. Right? So there's a series of these types of agents that are function specific.

[00:05:43] - [Speaker 1]
Right? Agents for returns, agents for refunds that we can introduce. But there's also we also expect that over time, I mean and and this is the next set of releases coming out from us, there'd be vertical specific agents. Agents that specialize in e commerce. Agents that specialize in, for example, hospital operations.

[00:06:02] - [Speaker 1]
Agents that specialize in banking, for example, right? Or like the equivalent of the of of a bank teller, for example. So in order to do that, we built a very horizontal platform that allows you to build all these things on top of it and then get your own business' very specific requirements, specific policies, specific experiences, specific guidelines into it that operates as well and working in concert with humans doing the same role.

[00:06:35] - [Speaker 0]
And at Zendesk Relate this year, it feels like almost a moment where the industry is moving from just talking about AI assistance to building these truly agentic systems that you've just walked me through there. So from your perspective, leading products, engineering, and AI, what does that shift actually look like inside Zendesk? Can you give me an idea of what it's like behind the curtain?

[00:06:56] - [Speaker 1]
Well, there there are things that we do to build AI, and then there are things that we do with AI ourselves. So on the front of building AI, what we've been watching this one metric which we call automation rate. The automation rate basically says if a 100 customer requests come in or employee requests come in, out of those 100, how many what percentage, right, is the AI able to solve itself and what moves over to someone else? Like, what then has to be moved over to human? And basically, over the last year, we have seen that number go up from 30% -ish to 80% very consistently.

[00:07:34] - [Speaker 1]
The AI has become very powerful. We have learned to get it to do more of the work. It has become smarter. It has become, you know, hallucinates less, operates better within guidelines, all of that stuff. So building that system, which is very focused on this outcome, right, which is what's the automation rate, number of resolutions that we can actually solve, it's been a big transformation because remember, we used be a seat based company.

[00:07:56] - [Speaker 1]
We were a software as a service seat based company. We used to obsess over a number of seats. Right? Now we obsess over a number of outcomes. That's a cultural change.

[00:08:05] - [Speaker 1]
There's a second part, which is building with AI, which is our own engineers using AI to build these AI products. That is a very different type of transformation. They'll have to take very competent, very experienced engineers who are very good at their jobs, often have the identity tied to being very good at their jobs, and then convincing them that there's a better way. And that better way is to use AI itself as a helper, as an intern, as a, you know, as an army of very smart, but maybe raw talent that you can work with. And especially since November of last year, there's just been a massive, you know like, the entire industry is going through a massive shift, but we're also going through a massive shift in terms of how we do how we build products ourselves.

[00:08:57] - [Speaker 1]
So, for example, we are working with much smaller teams. We had about 200 teams, 2,000 people, roughly 10% per team, But the teams themselves are being broken up into smaller teams because we don't need that larger team. When you have a smaller team, it can actually move a lot faster because you don't have to carry You don't have to put every 10 people don't have to coordinate on it, right? So these teams can move faster. And we're consistently seeing that we can pull forward roadmap items that we thought we will do in 2027 is being done six months earlier or eight months earlier as a result of that.

[00:09:32] - [Speaker 0]
You raise so many great points there, and I think a big word that runs through everything is change, the pace of change as well. And what is needed to change both culturally and technically and operationally inside Zendesk to support this evolution? Because it is massive change, isn't it, we're talking about?

[00:09:51] - [Speaker 1]
This is a period of the most change that I've seen in my career. I started my I started out sort of in the workforce at the time of the first Internet revolution right before the crash. And then I also saw the the SaaS transition, the mobile transition. So I've been through some of these, but this is way bigger than that. And things that we thought would take several years to come.

[00:10:18] - [Speaker 1]
So for example, when AI agents when when LLN's first came out, we knew MCP, something like MCP or eight way would happen, but I would have put a 2026 timeline to it, but it actually happened in 2024. And then on, like, totally solving coding, we probably would have put a '27 or '28 timeline to what would happen in 2025. So everything is just so much faster this time. And managing this change is not easy, but I'll tell you how we are doing it. Right?

[00:10:46] - [Speaker 1]
So we are we spend a lot of time thinking about what are things that don't change, and then stay focused on that. Like, that's our North Star, right? In the middle of a storm, you gotta focus on something. So our North Star is certain things are not gonna change. Customers are still gonna have problems.

[00:11:06] - [Speaker 1]
Employees will still have problems. They will want to get those problems solved. They'll want it solved as quickly as possible, as accurately as possible, and the companies that serve these customers will want the highest CSAT possible. Right? So everything else is in service of that.

[00:11:23] - [Speaker 1]
If it if it helps with that quick time resolution, if it helps with increased accuracy of resolution, and if it helps with improving the customer satisfaction level, that's good. If it doesn't help with those things, then that's bad. And the the hardest thing, I think, is just as recognizing that stuff that we are working on today that we really believe is cutting edge, we may have to throw it all away in twelve months because new things become possible, and I think that's a very big shift.

[00:11:56] - [Speaker 0]
And over the last two, three years, we'll have many leaders listening from organizations that struggled with AI projects to get ROI measurable difference and generate better business outcomes. So how did you approach this at Zen Zendesk when approaching AgenTik AI in a way that delivers those measurable outcomes for customers in instead of simply laying AI features into existing workflows?

[00:12:21] - [Speaker 1]
Our customers don't have that problem. I think we're very consistently able to deliver value, and the reason is because we don't come in and say we're gonna do AI. We come in and say we are going to automate more and more of your the the ticket volume or your sub service volume using AI. There are specific metrics that have already been defined. Automation rate is one of them.

[00:12:46] - [Speaker 1]
There are two big metrics that come with AI. One is automation rate, which is what what portion of tickets or issues can AI solve out of the total. And the second is what we call auto assist rate, auto assist acceptance rate, which is a measure of when an AI makes a suggestion to a human, what percent of them the human accepts immediately versus, you know, edits or makes changes to it. Those two measure the effectiveness of AI. If you take those two things and then combine with what ultimately matters the most, which is did the customer leave happier, so the CSAT or equivalent of that, between those three things, we have a very clear set of metrics to work with.

[00:13:24] - [Speaker 1]
And so when we go into a customer, we start by saying, We think we can automate this much. We believe the auto access acceptance rate is gonna be this much, and we believe we can move your CSAT points by five or 10 or whatever. And then the rest of the discussion is about how to get there. And that aligns everyone because we one of the big transitions we made was also outcome based pricing. Meaning, if we are not solving the problem, we're not getting paid.

[00:13:51] - [Speaker 1]
Right? And that really aligns incentives. And I think that's where, you know, when people quote statistics, there was an MIT study saying, like, 95% of AI AI programs fail, is because they didn't have these. They didn't have clear metrics. They're often top down projects with no understanding of business context, and and they didn't have some, you know, the alignment of incentives.

[00:14:13] - [Speaker 1]
And we have all of those three things when we work with our customers.

[00:14:18] - [Speaker 0]
And you have also spoken about building autonomous AI systems that learn from every interaction. So tell me a little bit more about how that learning loop works in practice and why continuous improvement becomes such an important differentiator in modern customer service. That that feels like a real exciting moment for me.

[00:14:35] - [Speaker 1]
No. Thank you. Because it is a differentiator, and it's something only a company like us can do, because we have 80,000 customers. We sit on top of almost 20,000,000,000 tickets over our history, and there's a lot of data to learn from. As a result, we operate in a 100 plus countries, so we can even localize the learnings to individual countries and the individual geographies, individual verticals, individual functions, because we have this volume of data to learn from.

[00:15:05] - [Speaker 1]
Now, the key idea behind resolution learning loop is, number one, there's not one. There are many loops. Right? There are many loops, and it's a matter of figuring out observing the AI do its job, figuring out where it failed, go to the root cause of why it failed, and then make sure that next time it runs into their job, it doesn't fail. That's the idea.

[00:15:30] - [Speaker 1]
So simple example of this the simplest example of this is a customer comes in, asks the AI question about, hey, I need to reconfigure my product. I'm struggling with it. The AI goes and looks for information from the knowledge base, comes back, and is not able to answer that question. Turns out so so you look at that interaction and say, it's not able to answer the question. It handed over to a human.

[00:15:56] - [Speaker 1]
So we look at those failures, and we say, why did it get confused? Well, it got confused because it actually found three different documents that specify the procedure, and they're conflicting with each other. And the AI was not able to decide which is the right one. Right? So that now becomes part of the resolution learning loop because it's feedback back to the knowledge base manager saying you have three conflicting things out there.

[00:16:18] - [Speaker 1]
You need to reconcile them. Which is the right answer? Right? And you can actually then use AI itself to figure out which one they should have picked because you see the human solved the problem. Right?

[00:16:31] - [Speaker 1]
So the way the human solved the problem now becomes training data. So the next time around, it won't stumble on that problem anymore. And as a result, the resolution rate will go up. So it's this constant looking for why it's failing, and then making sure that the thing that caused it to fail, that loop is closed. If it's seeking some piece of knowledge that is needed to solve the problem and it doesn't have the knowledge, then it goes and creates that knowledge in the future using similar tickets in the past that had solved that similar problem.

[00:16:58] - [Speaker 1]
So I think that's the basic idea, but then we carry it across the board, and we specialize it to verticals, geos, functions, etcetera.

[00:17:06] - [Speaker 0]
I think that'll resonate with so many people listening. I think you look in any knowledge base, there is Techno 1.1, 1.5, final. Correct. There are just so many different They're

[00:17:16] - [Speaker 1]
not very well maintained, as you know, right? Yeah. Usually, they're not very well maintained. That's correct.

[00:17:21] - [Speaker 0]
And there's so many big announcements this week. But, I mean, looking to the future, where do you see the next big opportunities for AI powered service experiences emerging outside of that classic context contact center model? I was speaking with IRC this morning. They blew me away with their unique angle that they were approaching it. But what excites you?

[00:17:40] - [Speaker 0]
Any other big opportunities where you see this evolving and heading?

[00:17:43] - [Speaker 1]
Well, we believe this is just the beginning of AI for service, first of all. I mean, for I mean, just in our space, the contact center the the CX space, customer experience space is the farthest along. The employee service space is just coming up. Right? It's it's sort of less mature at the moment because that space tends to rely very much on on on automation versus AIs.

[00:18:11] - [Speaker 1]
And then the contact center space is probably the most immature, although voice AI is now picking up speed. Right? So so first of all, there is simply this, like, this penetration problem. Like, many more customers are gonna start using this kind of tech. Second is what we talked about today, which was specialized agents.

[00:18:27] - [Speaker 1]
Right? Specialized agents are really important because they just allow you to expand the the set of tasks that you can reliably do across the system. And for service organizations, it it expands the impact. Right? So specialized agents specialize by functions, by verticals, specialized by, you know, the type of job that needs to be done.

[00:18:48] - [Speaker 1]
That's the second big opportunity. Third, we're gonna see a lot more, what I'd call, embedded AI. So not necessarily that have a and somebody buys the Zendesk AI and the Zendesk user interface, but they take the capabilities that our platform brings and then they build something on top of it that's very custom to them. And we actually wanna encourage that. Like, we want customers to be able to do that.

[00:19:17] - [Speaker 1]
And finally, I'm pretty confident that the world is going to go towards agent to agent, meaning every consumer will have their own agent just like you have your own computer. On the computer, you'll have an agent. That agent, you're gonna give you access to your email, you know, some information, maybe your bank account, maybe if you trust it enough, etcetera. And it will really be your chief of staff, your personal assistant, and it will monitor stuff. And it will, for example, see that you missed a flight.

[00:19:47] - [Speaker 1]
And instead of you worrying about, hey. How do I get a refund for this flight? It will take that job on itself. It will reach out to the United agent, AI agent. It's a nice two agents talking to each other trying to resolve your problem.

[00:20:00] - [Speaker 1]
And I think that's the big shift that's coming. I really believe the future is about agent to agent.

[00:20:05] - [Speaker 0]
Do you think there's a certain trust element there of, as you said, giving permission to plug everything in, whether it be your bank, your calendar, your email? Is that trust factor there? Have you I mean, I I I be honest, I've had a look at an AI browser, I thought this was really exciting, and as soon as he started asking for things, the the ex IT guy, we got a little bit cautious.

[00:20:24] - [Speaker 1]
Well, what

[00:20:24] - [Speaker 0]
am I doing here? But, yeah, I mean, it's just guardrails need to be in place.

[00:20:30] - [Speaker 1]
I I mean, look, but we we already have these guardrails, right? Like, I mean, just think of, like, one way One analogy I use is like, if you have teen kids or even older kids, right? Sometimes you get them a credit card and you give them a bank account, but you keep a cap on it, so they can only do so much damage, Right? But then but then you kinda let them run with it. Right?

[00:20:49] - [Speaker 1]
So it's like, you're gonna treat these things in a similar way until they build up enough trust. I mean, no one's recommending that you attach it to your main bank account. On the other hand, the mechanisms will, like, as these things become more powerful, especially personal agents, the mechanisms to manage this will also come up.

[00:21:06] - [Speaker 0]
And, of course, admin copilot is another big announcement. It feels like a a big move. It also feels part of a bigger industry trend where AI is increasingly becoming operational, proactively surfacing insights, recommended actions rather than waiting for commands. Do you think we are entering a phase here where enterprise software starts behaving more like that intelligent partner than than a static tool? We're looking in the future a moment ago where we think it's going, but it feels like a a real pivotal moment, doesn't it?

[00:21:34] - [Speaker 1]
Look, that's that is absolutely the vision. So if 80% of all the work is gonna be done by AI, then the job of humans is gonna become it's two jobs. One, do the other 20%. And two, monitor and manage the AI. Right?

[00:21:53] - [Speaker 1]
So monitoring, managing the AI, making sure it's behaving correctly, making sure that it's actually configured, for example, to get you the most value from your business, that is respecting your guardrails, your rules, understanding where it's failing, managing the resolution learning loops. All of that, for us, admin copilot is the portal into that whole world. Right? Because we're calling it admin copilot, it's sort of, broadly speaking, how do you administer the whole system? Right?

[00:22:21] - [Speaker 1]
Not the sort of old fashioned software, the service, and I give you this seat and I give you that seat and I Like, that's not what this is about anymore. Right? This is gonna become much more about managing a running humming system and ensuring this operating properly, creating the right kind of value for you. For example, if you have a certain automation rate and it starts to dip, like, we don't want it to be the human's job to go figure that out. The AI itself should try and figure it out first and come back with, hey.

[00:22:49] - [Speaker 1]
Here are five different reasons this is happening. Let's rule those out first, and then maybe it falls in the realm realm of the human. So you are absolutely right. I think admin copilot has a I'll call it like the name kinda hides how big a deal it is. Yeah.

[00:23:07] - [Speaker 1]
Yeah. Yeah.

[00:23:07] - [Speaker 0]
Finally, for people who seem to might be concerned that they see jobs going, etcetera, and training, upskilling, all the things around that to ensure that nobody gets left behind and that that real story of AI and changing how humans interact with systems, workflows, and each other. Any advice you'd give to people listening that have the concerns rather than the optimistic vision that you and I have?

[00:23:30] - [Speaker 1]
Well, at the moment, first of all, we to run we have to go with data that we have, right? The data that we have shows that the number of jobs in contact center broadly, customer service, for example, broadly, is still going up and is growing faster than GDP. That's number one. Number two, when we look at our own customer and we filter for who is well AI adopted versus who is not AI adopted, the ones who are AI adopted are actually growing seats faster than the ones who are not AI adopted. And there's a little bit of this assumption, this belief that there's a finite amount of work to be done, and if AI takes on some of the work, then the human jobs will go away.

[00:24:13] - [Speaker 1]
But just think about the last time you had a great service experience. Think about the last time you're looking forward to calling someone, you know, and saying, hey, like, have this problem, and it's gonna be such a delightful experience, and you're gonna solve it for me. It never happens. Right? Every service experience is actually not that great.

[00:24:29] - [Speaker 1]
You're, you know, you're put on hold, you have a long hold time, then you talk to someone who doesn't know your problem because they're just triaging, then they give it to somebody else, and that person doesn't have your information. It's just a broken experience. So our belief is that there's a lot of unserved need in the service market, And the first thing that'll happen is companies are gonna go meet those unserved needs so they can keep the CSAT high or increasing, and they're gonna focus on that, the increasing CSAT as the metric, as opposed to the cost part of the metric. Right? And if you do that, then it's hard to make the case that there's gonna be a catastrophic job scenario here.

[00:25:10] - [Speaker 1]
That's what we're seeing at the moment. Who knows what'll happen in the future?

[00:25:13] - [Speaker 0]
I think that is a powerful moment to end on. I'd encourage everyone listening to feedback. Let me know your thoughts, your experiences, but thank you for sitting down with me today and starting this conversation. Really appreciate your time. Thank you very much.

[00:25:26] - [Speaker 0]
Very good questions, and really enjoyed the conversation. So many big takeaways from today's conversation around rather than chasing the idea of just one all knowing assistant, my guest argued that businesses need specialized agents to understand specific roles, policies, edge cases, workflows, and customer expectations. And this fills a much more realistic picture of how AI will actually operate inside modern organizations. And I was also impressed by that cultural shift that he described working inside Zendesk. Make no mistake.

[00:25:59] - [Speaker 0]
Moving from seats to outcomes is not just a pricing change. It is changing how teams build, measure, and ultimately think about value. And when success is measured by resolution, automation rate, and customer satisfaction, and the quality of human AI collaboration, at that point, the conversation becomes much more meaningful than simply asking whether a company has added AI. And maybe the most thought provoking idea there was his belief that the future will be agent to agent. Your personal AI assistant might be one day negotiating with a company's AI agent on your behalf.

[00:26:35] - [Speaker 0]
Whether that's resolving a missed flight, handling a refund, or just getting answers without you sitting on hold. And that future raises big questions around trust, permission, and governance, and the role that humans will need to play. And maybe the real opportunity here is not about simply reducing costs, but finally delivering the kinds of customer and and employee experiences that people have dreamed about for years. As you know, I am a tech optimist, but I'd love to hear your thoughts on this one. Are specialized AI agents the future of enterprise service, and how much autonomy would you personally be willing to give an AI agent that acts on your behalf?

[00:27:16] - [Speaker 0]
Love to have a quick straw poll on that one. So techtalksnetwork.com. You can send me a DM. You can record me a voice message. I'd love to hear from you.

[00:27:25] - [Speaker 0]
And if you are running a business right now, you may have noticed there's a quiet shift happening, one that most people are still underestimating, and that is your company doesn't live inside your network anymore. It lives inside the browser. That's where your SaaS apps sit. That's where your data moves. And increasingly, that's where attackers are focusing their attention.

[00:27:50] - [Speaker 0]
So NordLayer has just launched its new business browser, and it's designed specifically for small and medium sized companies that need visibility and control without the overhead of enterprise security tools. What I like here is the balance. You get advanced protection, better compliance, and full visibility into how your team is working online, but without slowing anyone down or forcing them to learn anything new. Feels like a practical step forward rather than another security layer that adds friction. So if you wanna see more about how it works, please head over to nordlayer.com/browser and check it out.

[00:28:31] - [Speaker 0]
But that's it for today. Time for me to hit the show floor again one more time now, but I will be back again tomorrow with a not a guest. Speak with you then. Bye for now.