Resolution as Architecture: Engineering Autonomous Service Systems That Actually Scale

Speed is the wrong metric for AI in service. Enterprises that optimized for faster response times found that faster wrong answers still produce the same escalation rates, the same customer dissatisfaction, and the same operational costs. The organizations redefining service architecture around AI are measuring something different: resolution. Whether the right outcome was reached, for the right customer, at the right point in the interaction, without unnecessary human intervention.

Melody Brue and Keith Kirkpatrick sat down with Zendesk’s Shashi Upadhyay, President of Product, Engineering, at Zendesk Relate 2026, with to examine what autonomous service systems actually require at the engineering and architecture level. The conversation covers why resolution is becoming the organizing principle for service platform design, what changes when AI moves from feature to operational system, and how fragmented tooling and data environments hold enterprises back from achieving consistent AI-driven outcomes at scale.

Shashi also addresses why specialized AI agents outperform generalist assistants in real service environments and what will ultimately define the next generation of customer and employee experience platforms as AI becomes embedded into core operational workflows.

‍Key Takeaways:

‍

🔹 Resolution is the organizing principle that changes how service platforms get designed. When enterprises shift from optimizing for speed to optimizing for resolution, the entire architecture changes, from how interactions are routed and escalated to how AI systems are trained, evaluated, and improved over time.

🔹 Treating AI as a feature and treating AI as an operational system are fundamentally different engineering challenges. AI as a feature augments an existing workflow. AI as an operational system owns the workflow, which requires different data architecture, different feedback loops, different governance, and a different relationship between the product and engineering teams that build it.

🔹 Fragmented architecture is the structural constraint that prevents consistent AI resolution at scale. Enterprises running patchwork tooling across service environments cannot feed AI systems the coherent, connected data they need to resolve issues reliably. Platform design must start from resolution and work backward to infrastructure, not the other way around.

🔹 Specialized AI agents are more operationally reliable than generalist assistants in production service environments. Purpose-built agents trained on domain-specific data, policies, and resolution patterns outperform broad-purpose models when consistency, accuracy, and compliance are non-negotiable requirements at scale.

🔹 Continuous learning without increasing complexity is the product design problem that defines the next generation of CX platforms. As AI systems become more autonomous, the organizations that win will be those whose platforms improve from every interaction without adding operational burden to the teams responsible for running them.

Watch the full video at sixfivemedia.com, and be sure to subscribe to our YouTube channel so you never miss an episode.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

‍

Shashi Upadhyay:
Look, I'll start by saying that Zendesk customers don't have that problem.

Keith Kirkpatrick: 

Okay.

Shashi Upadhyay: 

Let's have a clear definition of what resolution means. Okay. And then let's make sure that your customers, your employees, get to a resolution as fast and as accurately as possible.

Melody Brue: 

Hello and welcome to Six Five On The Road. I'm Melody Brew here with Keith Kirkpatrick and Shashi Upadhyay from Zendesk. Today we are going to talk about how AI is changing the architecture of customer and employee experience operations. Big day today. You were on the main stage. Big kickoff. How's it going?

Shashi Upadhyay: 

It's a great day. It's a very big day for us. Our annual customer event, several thousand customers are here, and it's in rainy Denver. So people had nothing else to do but to come listen to me talk. So that was great. It was really awesome. And we got to make a lot of announcements around new products, our AI direction, and so far we've been getting really good feedback.

Keith Kirkpatrick: 

You know, it's interesting when I listened to all of those announcements, they all seem to be focused around one thing, which is focusing on service, not just about doing it faster, but about really driving real outcomes and, and, you know, better customer service overall. Can you talk a little bit about that? Is that something that organizations really need to focus on as we enter this era of AI and agentic AI?

Shashi Upadhyay: 

Look, AI is just a technology, right? At the end of the day, you have got a human being with a problem that needs to be solved. It can be a customer, it can be an employee. And when they have a problem, the thing they want to have happen is have it solved as quickly as possible so they can get on with the rest of their life. So the business metrics that are built around that are things like how long are you making them wait, how accurate is your first solution, and did you solve it correctly so they didn't have to come back to you again, right? So those things are invariant. They don't change whether you have a manual call center, you're using a SaaS application, or doing it all in spreadsheets, or you're using AI. The way we see it is, what really matters is that outcome, that the problem got solved, that there was a resolution, and it happened as fast as possible. AI is very good at as fast as possible, because it works 24-7. It's available on weekends. It's available on holidays. It can speak in 100 languages. It can learn everything there is to know about all the knowledge that exists in your business, and it takes feedback pretty well. You can teach it to strike the right tone. detect emotion correctly, and then also ask for help when it fails. It reaches out to a human and says, hey, I couldn't solve this problem. You do it. So there is a lot of good that comes with it. And our take is, at the end of the day, what matters is the outcome. AI is good at doing a lot of things, but the conversation should be about outcomes. That's why we are, as Zendesk, making a huge commitment towards not just measuring outcomes, but also pricing for outcomes. That's called resolution-based pricing.

Melody Brue: 

Yeah. So let's talk about getting to those outcomes from an engineering perspective. We say, you know, AI is just a technology. But when you involve the human in that, and AI is good at a lot of things, and one of them is not being tired and, you know, all the things you mentioned, but empathy may not be a strong suit. And so from an engineering perspective, because we are talking about a technology, how do you balance that human empathy, when to bring the human in the loop, and what AI can do on its own so that you get to those outcomes that you're talking about? Because at the end of the day, that's where you're trying to get to, human, AI, human plus AI, regardless of the combo there.

Shashi Upadhyay: 

That's a great point. So the empathy factor itself has like many pieces to it, right? So a big part of empathy is of course, asking more questions, listening intently.

Keith Kirkpatrick: 

Understanding.

Shashi Upadhyay: 

Understanding. Understanding what is it, like what's the real problem you're trying to solve, right? Understanding intent. And then using that information to figure out what the right solution is. Some of those things like, yeah, it doesn't get tired, doesn't get irritated, pretend to listen well, right? Those are all adjustable parameters in our models. So you can strike, you can, you know, we can adjust the tone to reflect the voice of the brand, the type of agent you would like to have. Yeah, it's very good at mimicking that. I think on the intent detection piece, one of the things we bring is we have a long history of being in the service business. We have over 20 billion tickets over years that we have collected. We have seen every problem in every industry, in every country, in every vertical, under every constraint. And that has allowed us, that has helped us build out a very comprehensive intent model, the type of problems people show up with. That, when you combine the AI sort of raw intelligence with that intent capability, you can get results almost as good as what you'll get with a human. Remember, the comparison is always, what will the human agent do in that situation? And what we train and what we tell our customers is, It's a great idea that AI should take the first shot, but if it's unsure, it should move to a human as quickly as possible.

Keith Kirkpatrick: You know, as we think about moving to an idea of really focusing in on resolutions, it seems that sometimes organizations have a real problem. They have, they're using a patchwork of tools, data is here, data is there, maybe it's not in the right format. How is having sort of a unified platform, how does that really help enable organizations to really drive these resolutions much more quickly, efficiently, and more cost effectively?

Shashi Upadhyay: 

Yeah, so the question you asked was like a really good question, right? In order for AI to be effective, it has to have all the information. So we call it context. It has to have all the context available. And that context is not just data. So data can be about like, hey, this customer has been our customer for this long. These are all the things they have purchased, these are all the things they returned, these are all the times they got a refund, these are all the support calls they have made. That is data. But context also includes other stuff, like what actually happened in that conversation. What are similar customers to? What was the actual state of the product, for example, on the thing that they had purchased? Maybe they had a recall later on. all that information, the structured, unstructured data together, bringing that into a single response and a single conversation, that is an art. I mean, that is where the engineering magic lies. So we come at it from two approaches, from two sides. One, we have this thing called a knowledge graph. The idea of a knowledge graph is take every piece of information that's available to you, not just in a Zendesk system, but also in other places. The information may live in Google Drive, or it may live in some folder of PDF documents somewhere, or it may live in Box or Dropbox. go connect it to all of those places, so it has all of that context, but also the data layer that we provide, so you can actually connect it to other systems where some of these information resides. So Stripe transactions or any customer database you may have built yourself, and that together with a timestamp on it, right? So like this is the data and this is when it happened together. That forms what we call the context graph. Once that context graph is available, AI becomes the most knowledgeable, intelligent being in the building, because there's no human being who can keep track of all of that stuff, right? So once you have that start, I think that's, in some sense, the journey a lot of our customers have to go to, because they have not really organized their systems that way. And we see an opportunity for them to start to think that way, which is, how do you organize a system So that just as you worry about how are my human agents going to be effective, you also have to worry about how will my AI agents be effective. And they're not exactly the same problem.

Melody Brue: 

In the same way that you have humans that specialize in certain things and human agents that specialize in certain things, you have AI agents that specialize in certain things or that are more kind of multi-purpose. And then today you talked about agent builder. In what scenarios does each one play the most important part? Where is specialized really needed versus where do you maybe want to look at building your own versus where is one that could maybe get most of it right most of the time?

Shashi Upadhyay: 

Yeah, so the approach so far for the most part has been to have a horizontal agent that does everything. And that's good for, I would call easy low value problems, right? So an easy low value problem is like the classic example of that is a password reset, okay? Like password reset was actually solved several years ago. You don't really need an LLM for that, but some people like to talk about that as a great example. For us, that's an easy problem. That's something a horizontal agent can do. But when you get into areas where the company has invested a lot in a specific business process, that makes that company good at it, special at it. then you're going to need a specialized agent. So let's say you have like this amazing refund process, you really optimize how refunds happen, which customers get a refund, which don't get a refund, and under what circumstances they get a refund, and you put all kinds of policies around it, and you have different rules for different regions, and All that specialization and distillation of that, we believe is better done by a specialized agent. So you train a specialized agent with your company knowledge, and then the more horizontal agent that fronts the interaction can just say, oh, you have a refund problem? Let me call the refund agent and let them do the job for you. And unlike with a human agent, whenever those handoffs happen, I'm sure you've experienced those, you are put on hold. That will not happen. It's seamless. From an end user perspective, it'll be a seamless interaction, but you're interacting with an agent in the back that has that specialized knowledge. So we kind of think of it as leave the simple stuff to the horizontal agent, but the stuff you value as a business where your knowledge, your specific processes, procedures are embedded, put those in specialized agents. And go crazy, build as many of them as you want because our platform can support that.

Keith Kirkpatrick: 

So as we start to see AI become really commonplace, almost ubiquitous in some ways, what are the things that are going to sort of separate the organizations that are able to actually see real value from AI as opposed to the ones who don't? Because I think that's a big question on, you know, really everyone's mind, you know, from CIOs on down.

Shashi Upadhyay: 

Look, I'll start by saying that Zendesk customers don't have that problem. Okay. And the reason is because of what we started the conversation with. Our goal is very simple. Our goal is let's have a clear definition of what resolution means.

Keith Kirkpatrick: 

Okay.

Shashi Upadhyay: 

And then let's make sure that your customers, your employees get to a resolution as fast and as accurately as possible. Once we define those as the metrics, which is given a certain number of cases that come through the door, how many of them are we able to resolve? Number one, metric number one. Metric number two is how quickly does the resolution happen? Does it take multiple turns, multiple days, or does it just happen in one shot? And metric number three is kind of a corollary to that, which we call auto-assist acceptance rate. So when it goes off of a human agent and goes to go off an AI agent goes to a human agent, the human agent has a co-pilot who's recommending like, hey, do this. And if it turns out that the AI agent got it right, then that is the auto-assist acceptance rate. Those are the three metrics that we really focus on. We establish those up front and say, hey, this is what we think will happen. We help our customers get that. That's why when we talk about our customers, we always start with what was the automation rate? Did you get 30 percent, 40 percent, 50 percent, 60 percent, 80 percent? If I go to a customer and they say, hey, I'm only getting 40 percent, I'm calling my team saying, hey, this is not acceptable. We need to be at the 80% level with this customer. Let's figure out how to get them there. That sharp focus on business outcomes is what determines success. I think a lot of times AI projects fail because there's no established metric up front. It's a top-down initiative as opposed to driven by experts and people in the business. And there is no clear alignment between what the outcome is and how the vendor or whoever is getting paid. Whereas with outcome-based pricing, resolution-based pricing, that problem is already solved.

Keith Kirkpatrick: 

Everyone's working towards the same.

Shashi Upadhyay: 

Everyone's working towards the same metrics and the same goals.

Melody Brue: 

Yeah. You touched a little bit about the, on the pricing, but I want to kind of take this from the approach of going from seat based to resolution outcome based. Where does that put you from a like pressure to deliver standpoint as a company?

Shashi Upadhyay: 

The pressure is very high, but we are up to the task.

Melody Brue: 

Good to hear.

Shashi Upadhyay: 

And we're actually happy to be in that state because, look, we have been in the service business for almost two decades. We understand this business extremely well. We think when we are put on the hook to ensure the customer is successful, we do our best work. So we are very happy and we are leaning into it. That's why you see us talk about it nonstop.

Keith Kirkpatrick: 

All right. Well, Shashi, thank you so much for joining us here. Thank you very much, Mel. And I want to thank all of you for turning into Six Five on the road from Denver at Zendesk Relate 2026. So don't forget to hit subscribe, follow us on the socials, and check out all of the videos at SixFiveMedia.com. Thanks, and we'll see you again very soon.

‍