Adrian McDermott, Zendesk CTO: Transforming Customer Service in the AI Agent Era

Adrian McDermott, the Chief Technology Officer at Zendesk and the technical architect behind the company’s new AI strategy, recently discussed his views on agent-based customer service with BDM. As the pioneer of the Autonomous Service Workforce introduced this week in Denver, McDermott emphasizes that AI’s role is not to automate old processes but to create new capabilities that were previously impossible. He delved into topics such as Zendesk’s acquisition strategy, the concept of “service debt” he popularized, the significance of the Model Context Protocol, and the future reconfiguration of human roles in the industry.

From the Resolution Platform in 2025 to the Autonomous Service Workforce in 2023: What’s the Real Change?

AI is advancing so rapidly that our approach evolves alongside it. Looking back eighteen months, we began to grasp how agent-based AI could facilitate dialogues with customers, and aid our agents in enhancing those interactions through a human-in-the-loop system. These were our initial steps.

Yet, at its core, what we are enhancing is customer service. The challenge is leveraging AI across the entire team. The Autonomous Service Workforce is predicated on the notion that AI will impact every aspect of our work, either by assisting or automating a significant portion of tasks. With the introduction of Custom Agents Zendesk and our early adopters, we’re seeing the emergence of new use cases that were previously unattainable without human cognitive abilities.

Humans are invaluable, yet there are many things we could do but didn’t because it wasn’t economically feasible. Now, it’s possible. The Autonomous Service Workforce isn’t about automating old tasks; it’s about envisioning what we can achieve and build.

Zendesk’s Recent Spree of Acquisitions: Explaining the Strategy to Newcomers

When we acquire companies, we’re essentially buying three things.

1. Innovation. We acquire entities that have developed something we wish we had. Take HyperArc, which drives our newly unveiled Analytics Copilot, as an example. This represents a unique innovation in data memory and understanding that we couldn’t find elsewhere.
2. Teams. We acquire groups of people who deeply understand certain problem areas. Sure, I could start today by hiring ten people, but mastering these areas often requires learning from several failures—a journey I’m very familiar with. So, we’re also buying talent.
3. Time. This aspect should not be underestimated. Our acquisitions often involve a mix of these three elements. Some examples include entering new markets where we previously had no products, like customer service workforce management or customer service QA, with acquisitions like Tymeshift in Lisbon and Klaus in Tallinn.

In parallel, we’ve launched our Custom Agents platform, our IT Asset Management (ITAM), and our agent-based AI Actions Platform for integration. These represent what we call “zero to one” innovations, created from scratch. For instance, after evaluating existing players, we decided to build our ITAM from the ground up.

Whenever we evaluate a market, a company, or an innovation, we consider whether to build, buy, or partner, to determine the most effective approach.

Integration of Forethought Directly into Zendesk AI Agents: The Impact on Long-term Customers

To truly embrace continuous improvement, we leverage an intelligence system that analyzes every interaction and datum, recommending service enhancements. Consider a retailer whose AI agents initially failed to automatically handle return requests in Canada involving liquids due to shipping regulations. With our Resolution Learning Loop, when such an issue arises, the system can automatically generate and suggest a new procedure for the AI, asking the admin: “Would you like to approve or make a slight modification? We believe this could improve our service.”

Previously, tasks like gathering feedback from an agent, identifying patterns, and cross-verifying with another agent were managed by knowledge managers or service operations managers. However, each service agent typically sees only one ticket at a time. An experience in one location might be replicated elsewhere without any shared insights. Scaling this process for a large company would require a managerial organization that doesn’t exist. Hence, we aim to build systems that can perform these tasks instead of humans, enabling continuous improvement and capturing every signal.

Explaining the Concept of “Service Debt”

Here’s a counterintuitive observation: when a client automates 40% of their interactions, their overall volume of interactions with customers increases by 176%. I believe no company should ever say “we have enough service”. Contact centers are essentially human-powered factories, and the metrics used to manage them are based on human efficiency: tickets per day, time to first response, average handling time. Optimizing these metrics focuses on how many tickets an agent can handle in a day, not the quality of the service provided.

As a result, every company actually has what I call a “support debt”. Customers give up due to long waiting times, can’t find the help they need, or forego requests they wanted to make. This latent debt is exposed by automation. Once we properly automate customer experiences, we uncover all the issues customers have been trying to communicate that were previously unnoticed.

Interestingly, automation drives escalation: the median load on human agents has increased by 87%. They become more efficient, almost supercharged.

The concept of service debt revolves around the idea that we aren’t truly serving all our customers with everything they desire. And the more accessible we make our services, the more people use them. Economists refer to this as the Jevons Paradox, which is currently a trendy term in AI.

Addressing Service Debt: Where to Start and How to Measure Progress

At Zendesk, we begin by tagging each customer contact with what we call the “intent”: what the customer was trying to accomplish, the reason for the contact, their sentiment, and the language they speak. We then map these intents by volume and average duration, assessing which are most automatable. We assist the client in automating these, connecting them to necessary systems, and setting up authentication.

Too often, corporate decisions are driven by the need for a comprehensive strategy before action is taken. There’s something to be said for iterating towards success, especially if you monitor your quality metrics along the way. That’s one reason why Zendesk launched the Quality Score this week: to provide an objective measure of the quality of service delivered.

A telling example is how Zendesk uses its own AI agent. Contrary to intuition, we aggressively automate 60 to 70% of our requests, and our NPS, TNPS (transactional Net Promoter Score), and CSAT (customer satisfaction index) are improving in tandem. The lesson here is to build good customer journeys, connect them to AI capabilities, and you will construct something superior.

Launching Both an MCP Client and an MCP Server: Why Zendesk Chose This Standard, and Which One is More Transformative for Clients

The Client and Server serve different purposes, representing two sides of the same protocol. For me, Zendesk as an MCP Client represents a new way to connect everything. I’ve been in tech for a long time—originally it was SOAP, then REST, XML, and now MCP. It’s useful because it provides a standard, easily implementable, flexible set of connection methods. You can build it with an AI model from a REST API. It’s become the lingua franca of AI agents, so it makes sense to adopt it. Is MCP a revolutionary technology? No. But has it reached escape velocity and is now in orbit? Absolutely. If you have an integration platform like our Zendesk AI Actions Platform, being able to consume it is critical because you then gain access to numerous different systems.

The other direction is the MCP Server: we make our data and workflows available so that third-party AI systems can access them. This week, we’re talking about LLMs as a new service channel. At the same time, traffic from Zendesk to our own help center has decreased by 20% because queries are being resolved via ChatGPT or Gemini. The MCP Server enables our clients to move in this direction: using their Zendesk platform as a component of a broader agent-based solution connected to external AI assistants.

The way to win for any software publisher at scale is to be both Client and Server, and to understand when you add value in each role. Some of our largest accounts are already building their own agents using their data and expertise, which opens up fascinating agent-to-agent use cases.

Tom Eggemeier Discusses the Rise of Specialized AI Agents: What Does This Mean for Human Customer Service Teams?

Undoubtedly, a reconfiguration of roles. Today’s large-scale contact centers operate like human-powered factories with clearly defined hierarchical roles: tier 0 (basic, very preliminary sorting), increasingly experienced tiers 2 and 3, team leads who manage human agents and organize schedules, and highly specialized and narrow roles like the knowledge management expert. To consider how these professions will evolve in an AI-driven world, it’s crucial not to confuse the tasks that occupy most of an employee’s time with the ultimate purpose of their job.

A good example is Sir Geoffrey Hinton, a Nobel Prize winner and a proponent of deep learning, who once advised students not to become radiologists because they might not be needed in five years. Five years later, there are more radiologists than before. Why? The FDA recognizes that AI-based medical scanners are now more accurate than humans. However, looking at scans was the task. The real job is interpreting the patient as a whole, developing treatment plans, connecting with them, and collaborating with other doctors. That’s what being a doctor is about.

We need to examine each role in customer service and ask: what is the task, and what is the purpose? Once we do that, we can redefine roles. Yes, tiers 0 and 1 will likely be automated in most cases. But you’ll need more tiers 2 and 3 because human-in-the-loop processes aren’t going away. And as I mentioned, automation drives escalation: it’s a human who steps in next. Team leads become a more sophisticated job because they have to manage both human and AI agents. As for the highly specialized and narrow roles, they expand into a new profession: the AI Service Architect, someone who considers knowledge, integrations, and procedures that drive AI.

In agent-based services, the main job becomes training AI agents for tasks at which they are exceptionally good. Not those that require humans.

Similar Posts

• OpenAI Unveils Frontier: New AI Management Platform Revolutionizes Business Operations!
• Slack Upgrades Its Slackbot: Meet Your New Enterprise AI Conversation Partner!
• Customer Focus & Team Excellence: Shashi Upadhyay’s Fast-Paced Method at Zendesk
• GitHub Unveils Agent HQ: Central Hub for AI Web Development Agents
• Microsoft Unveils Agent 365: Centralize and Manage AI Agents in Your Business!

Jordan Park

Jordan Park writes in-depth reviews and editorial opinion pieces for Touch Reviews. With a background in UI/UX design, Jordan offers a unique perspective on device usability and user experience across smartphones, tablets, and mobile software.