Dec 5, 2025 · 40m · big-technology
Zendesk's Adrian McDermott: AI's Customer Service Potential, Adoption Cycle, Scale vs. Orchestration
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Zendesk CTO Adrian McDermott joins Alex Kantrowitz to examine the practical integration of generative AI in customer support, detailing the shift from knowledge search to agent orchestration, backend API challenges, and emerging frontiers in voice and persistent memory.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 28.2% of the talking time here. How this is scored →
Alex as informed peer 4.3 Guest teaching 4.6 Guest disagreement 1.5 Alex pushing back 1.4
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 19:20 Adrian rejects the premise that AI models cannot execute actions
Adrian counters Alex's assertion that AI support cannot process refunds or schedule changes, asserting that the capability exists today and is merely constrained by enterprise API readiness.
Hardest push from Alex ▶ 18:29 Alex presses on real-world refund and flight transfer bot handoffsAlex challenges Adrian by highlighting that automated bots consistently hand off to humans whenever actionable tasks like refunds or flight changes are requested.
Biggest teaching moment ▶ 19:48 Adrian educates on legacy system APIs and swivel-chair human tasksAdrian clarifies that the primary barrier to automated resolution is legacy infrastructure requiring human swivel-chair data entry across disconnected systems rather than LLM reasoning deficiencies.
Alex holds their own ▶ 28:06 Alex cites real-world voice cloning casework to challenge maturity estimatesAlex leverages specific case reporting on Evan Ratliff's multi-scenario voice cloning project to question Adrian's claim that voice AI is still far from enterprise readiness.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Impact of AI on Support Jobs and Evolving Metrics | 5 | 4 | 1 | 1 | Alex draws on his background covering marketing technology to contextualize customer service as a competitive brand differentiator. Adrian expands on this by explaining the economic tension between support debt and traditional throughput metrics like average handle time. | |
| The AI Adoption Continuum: From Generative Search to Copilots | 3 | 4 | 1 | 0 | Alex prompts Adrian to lay out the customer support AI adoption continuum. Adrian walks through the transition from generative search to copilot setups and autonomous agents, noting that procedural prompt authoring remains a nascent skill. | |
| Channel Shifts, Customer Intent Taxonomy, and Human Escalation | 4 | 5 | 2 | 2 | Alex asks whether customer service will migrate away from bespoke brand websites into general LLM hubs. Adrian uses Zendesk's 15-million-ticket taxonomy to illustrate that automation counterintuitively drives escalation to human representatives. | |
| Model Maturity, Determinism Guardrails, and AI Orchestration | 4 | 5 | 1 | 1 | Alex inquires about model sufficiency and orchestration. Adrian explains how software engineering teams have built deterministic guardrails around non-deterministic libraries and describes using frontier Claude models to verify issue resolution. | |
| System Action, Backend Integration, and the API Bottleneck | 5 | 6 | 3 | 3 | Alex brings up Mustafa Suleyman's views on orchestration and questions why bots still pass users off during actionable requests like refunds. Adrian pushes back, clarifying that backend API availability and legacy digital transformation are the actual bottlenecks rather than model intelligence. | |
| Coding Models and Automation Across Support Tiers | 4 | 5 | 2 | 2 | Alex asks if models need stronger capabilities like solving captchas to automate actions. Adrian explains that coding assistants like Cursor unlock internal integration building, dividing automation feasibility cleanly across support tiers. | |
| Consumer AI Agents and Agent-to-Agent Standards | 4 | 3 | 1 | 1 | Alex shares his vision for a consumer-side personal AI agent that bypasses support menus. Adrian outlines Anthropic's Model Context Protocol (MCP) and agent-to-agent interface standards under development. | |
| Latency and Emotional Reasoning Challenges in Voice AI | 5 | 6 | 2 | 2 | Alex references Evan Ratliff's voice cloning experiments to challenge why voice AI remains delayed in production. Adrian breaks down the latency stack involved in speech-to-text, LLM reasoning, and text-to-speech, emphasizing the compute barrier for handling emotional interruptions in real time. | |
| Contextual Summarization and Total Recall in Enterprise Memory | 4 | 4 | 1 | 1 | Alex asks how enterprise memory works and questions the architectural limitations of LLMs. Adrian explains that current implementations rely on prompt summarization injected into context windows rather than true total recall. | |
| Continual Learning Systems, Structured Governance, and Trust | 5 | 4 | 1 | 1 | Alex references Satya Nadella's commentary on continual learning and asks about organizational trust. Adrian contrasts black-box continual model training with human-governed macro creation and deterministic system updates. |
Statements from this episode (17)
McDermott: 54% of Zendesk support-contacting users drove 95% of 2024 revenue
“If we look at our own company data, 54% of our customers contacted us for support in 2024, but they represented 95% of revenue.”
McDermott: Traditional customer support metrics are obsolete in an AI era
“Customer service is this human powered factory, right? The metrics are ones of throughput and effort. It's like tickets today, per day, time to first response. You know, all of these things where you're measuring productivity, average handle time. In an AI wor…”
McDermott: Generative Search Handles 30% to 40% of Support Inquiries
“They're deploying generative search and they're seeing, You know, upwards of 30, 40% of inquiries being handled by generative search.”
McDermott: Direct-Answer Search Is Now Table Stakes in Customer Support
“Users who spend, you know, probably two human generations learning to type into a box and process 10 blue links have suddenly pivoted, and they just want the answer and the results, and I think that's table stakes for customer support at this point.”
Zendesk: 47% of Customer Service Inquiries Stem From Business Failures
“We kind of looked at, we took a sample of. Fifteen million or more customer service conversations across send us customers. And we actually use chat GPT to classify the contact reasons or the intents and just kind of group them into cohorts. And you get someth…”
McDermott: Customer Support Will Add Dedicated LLM Channels
“And so I think, yes, we will, you know, if you think about, you know, you have a chat channel and a voice channel, and you have, you know, a messaging channel and an email channel, you're going to have an LLM channel. Where the LLM will kind of be the initiato…”
McDermott: Automating Support Drives Increased Human Escalations
“One of the things that we say is automation drives escalation. You know, as I automate something, more and more people, the Alex's of the world are pressing zero and asking to talk to the operator. And that isn't coming away.”
McDermott: Zendesk has 20,000 customers using AI features
“Like Xendesk has 20,000 people using some kind, 20,000 customers using some kind of AI.”
Zendesk uses Claude to evaluate support conversations for resolution-based pricing
“We have an agent that listens to every conversation that is automated and said, you know, cause we run a resolution platform and we charge for actual resolutions, not just conversations. And so we need to know that the customer's problem was resolved. And so w…”
McDermott: Lack of backend APIs is the primary blocker for AI agents taking action
“I think if you don't have an API on that system, we can't swivel chair yet. With an AI agent effectively or, you know, reliably. And, you know, we can vibe code and build the integration for them, but the API has to be there. And so mostly the blocker, like wh…”
McDermott: Coding models will drive support automation more than computer-use agents
“I actually think what's really going to make a difference more likely than computer use models, computer using models is just the improvements in coding models. And that's a little bit basing the future on the immediate past. Coding models have gotten gotten s…”
McDermott: AI can automate 80-100% of text support, but voice is unready
“If it's an easy landscape and we can cover the API estate with what we have, you could really get to 80 or a hundred percent depending on where you want to go. So as of today, it's all there. I think voice isn't quite there, and that's it.”
McDermott: Anthropic's MCP and agent protocols are becoming industry standards
“There's a couple of standards out there. MCP, which was developed by Anthropic, and Agent to Agent, and a few other things, and they're becoming fairly standard”
McDermott: Production-ready mass-market voice AI is roughly a year away
“Well, I think it's almost, it's basically ready now, so we're only kind of a yearish away, but I think to get it mass market, mass scale, and always get the latency and reliability that we want, probably still, probably, you know, that's coming online in the n…”
McDermott: Enterprise AI Simulates Memory via Context Window Summaries
“I think we're experiencing memory through putting really good summaries into the context window, right, or into the prompt.”
McDermott: True AI Memory May Require Tech Beyond Current LLMs
“We might not get there with the exact technology that we're using for large language models at the moment, but it feels like some way that we want to get to with sort of frontier level at AI.”
McDermott: Autonomous continual learning creates an unverifiable black box for enterprise support
“When we're changing the machine and we're writing new articles and we're building new macros, it's sort of easy for the humans who own that machine to understand what's going on. Like the head of support technology can be like, yep, yep, that is the correct wo…”