Eye On A.I.

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.

AI agents can now connect to every tool your employees use. The problem is that connecting them and trusting them are two completely different things, and most enterprises have figured out the first without solving the second. Oren Michaels, co-founder and CEO of Barndoor AI, joins Craig Smith to explain why that gap is the defining challenge of the agentic enterprise era. His framework is simple and sharp: agents are like enthusiastic interns. They will absolutely do something when you ask them to. Whether it's what you intended is another matter, and when an agent can act across Salesforce, Slack, email, and calendar simultaneously, the blast radius of a misunderstood instruction is far larger than anything a human intern could cause.
The conversation covers the 100,000 agent problem - the reality that each agent handling a discrete task needs its own set of rules about what it's allowed to do, and that number scales to a size no human team can govern manually - and why traditional identity management systems were never built for the failure modes AI agents create. The new threat isn't bad actors getting in; it's authorized people using allowed tools with agents that still do the wrong thing. Barn Door's governance layer sits between the agent and the tools it can access, specifying exactly what each agent is permitted to do in each context, and Venn brings that same capability to individuals who want to understand what's possible before their organizations catch up. This is one of the most practically useful conversations available about what enterprise AI governance actually looks like.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Every time you hit a phone tree or a chatbot with canned answers, you're experiencing the gap between what AI can already do and what most companies are still delivering. Craig Smith sits down with Tom Chen, Chief Product Officer at Aircall, to explore why that gap is closing fast, and what it means for any business that relies on voice as a customer communication channel. Tom makes a case that is both practical and counterintuitive: AI voice agents aren't better than your best human rep, but they are better than your average one. They never get frustrated. Their patience is infinite. Their tone never changes. And they can handle 100 concurrent calls at a fraction of the cost of a human operation, without lunch breaks, without bad days, and without going off script.
The conversation covers a finding that should change how any business thinks about AI adoption: when one of Aircall's customers gave callers the explicit choice between a human agent and a faster AI agent, far more people chose the AI than anyone expected, and satisfaction scores went up. Tom also identifies the real bottleneck that most businesses don't see coming: it's not the AI technology, which is increasingly commoditized. It's the tribal knowledge, the undocumented expertise that lives in the heads of long-tenured employees and never gets captured anywhere, that determines whether an AI agent performs well or not. Until that knowledge is surfaced, even the best voice agent will underperform.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

Most AI systems follow a gradient, a mathematical slope that tells them exactly how to improve, step by step, toward a known goal. Neuroevolution doesn't follow any gradient. Instead, it runs hundreds or thousands of competing solutions simultaneously, spreads them across the space of possibilities as broadly as possible, and lets the best ones recombine, the same logic that drives biological evolution. The result, as Risto Miikkulainen explains to Craig Smith, is creativity: solutions that no human designer would have anticipated, that emerge routinely from the evolutionary process.
Miikkulainen is a professor at UT Austin and VP of AI Research at Cognizant AI Labs, and he has been working on this field since the 1980s, which makes him both a historian of it and one of its most active frontiersmen.
The conversation covers a remarkable range: a mystery model that outperformed every competitor in a recent stock trading competition with forensic footprints pointing to neuroevolutionary AI; Sakana AI's system that autonomously designed experiments, wrote a paper, and had it accepted at a major machine learning conference; and a pandemic decision system that trained overnight and made country-specific recommendations by morning, with Iceland actually following some of them, all the way to the prime minister.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.

One in four people over 65 will experience a fall, and for most of them, the technology designed to help is a device that hasn't meaningfully changed since the 1980s. Chia-Lin Simmons, CEO of LogicMark, joined Craig Smith to make the case that this gap is both unnecessary and solvable, and that AI is finally making it possible to shift personal safety from reactive to predictive. Her company's Freedom Alert Max doesn't just detect falls after they happen, it builds a personalized digital twin of each user, tracking steps, sleep patterns, and medication adherence over time to identify the subtle signs of health decline that even daily caregivers often miss.
The conversation is one of the most grounded and human discussions of applied AI you'll hear, covering why Apple Watch fall detection was engineered for crash detection, not elderly falls; why AI can flag a problem but a human needs to hear the breathing on the other end of the line; and why the 700,000 caregiver shortage in America makes technology like this not a luxury but a scaling mechanism. For anyone navigating aging parents, their own future, or the sandwich generation pressures in between, this episode is both practically useful and genuinely moving.
Subscribe to Eye on A.I. for weekly conversations with the people building and deploying the future of AI.